The Senior Colleague Who Tells You When It Is Not Sure
Claude Haiku 5.5 resets the bottom of the price list, and on the Claude API the model it replaces has no guaranteed date left. Haiku 5.5 landed October 7, 2026 at $0.10 / $0.50 per million tokens for prompts up to 100,000 tokens, and $0.50 / $2.50 above that. Haiku 4.5 now sits under “Legacy models,” and its Claude API floor of October 15, 2026 fell without notice. Microsoft Foundry is firmer and retires it November 15, 2026. Opus 5.5 costs $4 / $20 with no retention gate. Sonnet 5.5 costs $2 / $10, with cache reads at $0.10. Fable 5.1 tops the lineup at $10 / $50. And Sonnet 4.5 stops serving on November 30, 2026.
The colleague who will not pretend the work is done
Claude’s best trait isn’t speed. It’s judgment. It will name the evidence that’s missing, argue with your brief, and keep the doubt in plain sight while it still gets the work done.
Careful without being passive
Claude reads closely and keeps the nuance. Offered a tidy conclusion it can’t support, or a caveat, it takes the caveat.
- Strong source fidelity
- Structured, professional writing
- Useful self-review and error detection
Work where judgment matters
Long documents. Complex coding. Regulated analysis, technical reviews, and anything that has to show its working months later.
- Board and client materials
- Repository-scale engineering
- Financial, legal and scientific workflows
The brief may itself be wrong
Claude earns its keep when the input is thin, argues with itself, or would tempt a model into a polished fiction.
- Ask it to inspect the premise
- Invite direct disagreement
- Require uncertainty to remain visible
Analyze this proposal for the executive team. Before recommending anything, identify what is missing, internally inconsistent or too weakly evidenced to support a decision.
Create the strongest defensible synthesis from these sources. Separate: verified facts, reasonable inferences, unresolved questions and claims the material cannot support.
Three new models in fifteen days, a cheaper cache read, and a floor that fell with no notice
The cheapest model is the one that moves a bill fastest. Claude Haiku 5.5 shipped October 7, 2026 at $0.10 / $0.50 per million tokens for prompts up to 100,000 tokens. That’s 90 percent under Haiku 4.5. It followed Opus 5.5 on September 22 at $4 / $20 and Sonnet 5.5 on September 28 at $2 / $10. Neither is a Covered Model, so neither carries the 30-day retention requirement. The same October 7 release cut Sonnet 5.5 cache reads to $0.10. Two dates matter more than any price. Haiku 4.5’s Claude API floor of October 15, 2026 arrived with no notice, while Microsoft Foundry retires it November 15. And Sonnet 4.5 stops serving on November 30, 2026.
Breaking change: tool_choice Four models return 400
This is the breaking change most likely to be sitting in your production code right now. Forcing a tool call has been the standard way to guarantee a parseable response. On Fable 5.1, Mythos 5.1, Opus 5.5 and Sonnet 5.5, the tool_choice types any and tool are not supported and return a 400. The types auto and none are unchanged. Anthropic’s stated alternatives are strict tool use or structured outputs. And here’s the trap. A codebase sweep scoped to “the 5.1 models” was correct in September and is wrong now. Scope it to the model ids you route to, then search for the two removed values.
anyandtoolreturn 400 on Fable 5.1, Mythos 5.1, Opus 5.5 and Sonnet 5.5autoandnonebehave as before- Use strict tool use or structured outputs instead
- A sweep scoped to “5.1 models” now misses two more
- This fails the request; it does not degrade quietly
Breaking change: thinking blocks 5.1 returns 400
A team can split down the middle on identical code, and the only difference is how old their accounts are. Here’s why. Thinking blocks are preserved only for the model that produced them, or a newer one. For accounts created on or after August 31, 2026, replaying a thinking block after a changed system prompt, changed tools or a changed message returns a 400. The beta header thinking-binding-controls-2026-08-01 controls that behavior. The exposure is any pattern that caches an assistant turn and replays it against an edited prompt. That’s common in agent frameworks and in evaluation harnesses that mutate the system prompt between runs. An older account may never see the error. So one engineer ships and another one can’t.
- Thinking blocks survive only into the producing model or newer
- Accounts created on or after August 31, 2026 get a 400 on a changed system prompt, tools or message
- Controlled by the beta header
thinking-binding-controls-2026-08-01 - Older and newer accounts can behave differently on the same code
Thinking blocks are now account-bound, and this one fails silently
The dangerous version of a breaking change is the one that returns a 200. On Sonnet 5.5, a thinking block works only in the account that produced it, or in an account linked to it. Send one from a different account and the API drops the block before the model ever sees it. The request succeeds. You get an answer. The reasoning you paid to generate and carefully replayed was thrown away, and nothing in the response says so. Read that beside the other thinking-block rule, because the failure modes are opposites. The account-creation-date rule returns a clean 400 you can catch in a test. This one returns a quiet, slightly worse answer. Anywhere you share cached assistant turns across accounts, an evaluation harness, a shared fixture library, a multi-tenant agent, you can’t detect this from status codes. Compare outputs instead. Haiku 5.5 thinking blocks carry the same account binding, but Anthropic’s Haiku 5.5 pages don’t say what happens to a block sent from another account. So don’t assume either a 400 or a silent drop there.
- Sonnet 5.5 and Haiku 5.5 thinking blocks work only in the producing account or a linked one
- On Sonnet 5.5, a block from another account is dropped and the request still returns 200
- For Haiku 5.5, Anthropic doesn’t say what happens to a block from another account
- The opposite failure mode from the 400 on a changed prompt, tools or message
- Shared fixtures and multi-tenant agents are the exposure
- Detect it by comparing output quality, not by watching for errors
| Model | Status on October 15, 2026 | Retirement as published |
|---|---|---|
Claude Sonnet 4.5, claude-sonnet-4-5-20250929 | Deprecated September 30, 2026 | November 30, 2026 on the Claude API; Microsoft Foundry also lists November 30, 2026 |
Claude Haiku 4.5, claude-haiku-4-5-20251001 | Active; under “Legacy models” on claude.com/pricing | Claude API: not sooner than October 15, 2026, floor reached with no deprecation notice. Microsoft Foundry: November 15, 2026, no replacement named |
Claude Opus 4.5, claude-opus-4-5-20251101 | Active | Claude API: not sooner than November 24, 2026. Microsoft Foundry: November 24, 2026 |
| Claude Opus 4.6 | Active | Not sooner than February 5, 2027 |
| Claude Sonnet 4.6 | Active | Not sooner than February 17, 2027 |
| Claude Opus 4.7 | Active | Not sooner than April 16, 2027 |
| Claude Opus 4.8 | Active | Not sooner than May 28, 2027 |
| Claude Fable 5 and Claude Mythos 5 | Active | Not sooner than June 9, 2027 |
| Claude Sonnet 5 | Active | Not sooner than June 30, 2027 |
| Claude Opus 5 | Active | Not sooner than July 24, 2027 |
| Claude Fable 5.1 and Claude Mythos 5.1 | Active; Mythos 5.1 requires verification | Not sooner than September 1, 2027 |
| Claude Opus 5.5 | Active | Not sooner than September 22, 2027 |
| Claude Sonnet 5.5 | Active | Not sooner than September 28, 2027 |
| Claude Haiku 5.5 | Active | Not sooner than October 7, 2027 |
| Claude Mythos Preview | Deprecated June 9, 2026 | To be announced |
Where a blocked request goes Two categories, not one
If a runbook in your product names the fallback model, somebody guessed. The pages that describe the routing don’t name one. Anthropic’s readable wording is plural and unnamed: “Our safeguards do, however, still redirect several kinds of dual-use cybersecurity tasks...to our Opus models. This includes penetration testing, exploit generation, and binary-based vulnerability scanning.” So write “an Opus model” in runbooks and user-facing notes, not a version number. There’s a second routing category on the same page. It’s easy to read past: “Queries related to research and development in the life sciences will still be directed to our Opus models.” If your safety documentation describes routing as a cybersecurity mechanism, it’s describing half of it. The same page also restates the safeguard figures for the 5.1 generation and scopes them more narrowly than the August 7 post did. Anthropic says its latest biology safeguards for Fable 5.1 and Fable 5 fire 85% less often for benign requests related to elementary biology and medical questions. It also says Claude Code users can expect an average of around 60% fewer interventions per session from the cyber safeguards, measured against the previous safeguards on Fable 5. Both are vendor figures. One capability change belongs on a security team’s desk rather than a safety team’s. Anthropic says Fable 5.1 can now be used to discover software vulnerabilities, though not to develop exploits for them. On the API, the fallbacks parameter accepts a "default" mode applying Anthropic’s recommended fallback per refusal category, and server-side fallback is in beta. One change here costs money rather than behavior. Pre-output refusals in three categories are billed again from September 24, 2026, and the refusal callout above covers it.
- The destination is “our Opus models,” plural and unnamed
- Two routed categories: dual-use cybersecurity work and life sciences research and development
- Named cyber examples: penetration testing, exploit generation, binary-based vulnerability scanning
- Vendor figures: 85% fewer fires on benign elementary biology and medical questions for the 5.1 generation, and around 60% fewer Claude Code interventions per session
- Anthropic says Fable 5.1 can discover software vulnerabilities but not develop exploits for them
- From September 24, 2026, pre-output refusals in three categories are billed again, on all platforms
- Repoint any runbook that names a specific fallback model
Compliance and Admin API mature, and one change reaches backward August 26 to September 24
A records process that was tested in September can be broken today, retroactively, with no error anywhere. On September 24, 2026 the Compliance API Activity Feed stopped returning file names, project document names and artifact titles. The filename and title fields are now always empty or omitted, including on activities recorded before the change. Past records changed shape. Names must now be resolved by ID, using a Compliance Access Key with read:compliance_user_data. If your e-discovery or records workflow reads those fields, re-test it this week. The rest of the arc. On August 26 the Compliance API session endpoints left beta for Claude Cowork and Claude Code. The Admin API arrived in the ant CLI and across all SDKs. The September 1 release notes added the anthropic-version header to the Admin API Claude Enterprise endpoints, the Analytics API and the Compliance API guides. They instruct you to send it on every request. So anything built against the August 26 item may not be sending it. On September 24 local sessions for Microsoft 365 left beta, covering product_surface values beginning office_agents across Excel, PowerPoint, Word and Outlook. Claude in Chrome, at claude_in_chrome, is still beta. Every item here lives on rolling release notes, so confirm each directly before a records process depends on it.
- Activity Feed file names, document names and artifact titles are gone, retroactively
- Resolve names by ID with a Compliance Access Key carrying
read:compliance_user_data - Microsoft 365 local sessions out of beta: Excel, PowerPoint, Word and Outlook
- Claude in Chrome transcripts remain beta
- Send
anthropic-versionon every Admin, Analytics and Compliance API request
Cache diagnostics are out of beta September 23
You no longer need a beta header to see what your cache is doing. As of September 23, 2026, cache diagnostics don’t require cache-diagnosis-2026-04-07, and POST /v1/messages responses always include a diagnostics field. Cache reads now run from $1 on Fable 5 down to $0.01 on Haiku 5.5, and that Haiku rate jumps five times once a prompt passes 100,000 tokens. So measuring your hit ratio isn’t an engineering nicety any more. It’s the difference between two forecasts. This sits on a rolling release note, so confirm it against your own responses.
- No beta header needed;
diagnosticsships on every Messages response - Use it to measure the cache-hit ratio before you bank a cache saving
- Release-note only: confirm directly
Claude Code enterprise controls Dated changelog on code.claude.com
Three models shipped in fifteen days. If your model allowlist is loose, all three are already reachable. That’s the setting to look at first. The key availableModelsMatch takes the value "exact", which makes an availableModels entry allow only the version it names. New releases stay blocked until an admin lists them. Alongside it sit deniedModels, allowedProviders and allowClaudeInChromeWithManagedMcp. Here’s the trap. Both availableModelsMatch: "exact" and deniedModels need version 2.1.283 or later, and earlier versions ignore both keys without complaint. So set requiredMinimumVersion too. Version 2.1.299, dated October 14, 2026, fixed the Claude apps gateway serving models those keys block. Now the fallback. When an admin removes access to the default model on Bedrock or Google Cloud, the session falls back to an earlier version of the default model. If the default is an Opus model and no Opus version is available, it drops to the default Sonnet. Microsoft Foundry users see errors instead. So the quiet downgrade can cross a tier. The earlier cluster still stands. Managed settings decline an unauthorized model switch. The maxEffortLevel key caps reasoning effort across Bedrock, Vertex and Foundry. Unparseable managed settings fail startup by name. The managedMcpServers key distributes MCP servers organization-wide. And permissions.blockReadsOutsideWorkingDirectories prompts before the first read outside the working directories. Defaults are published per plan and per provider. Opus 5.5 is the default on Pro, Max, Team, Enterprise and the Anthropic API. Haiku 5.5 became the default Haiku model on the Anthropic API in version 2.1.293, dated October 7, 2026. And the sonnet alias resolves to Sonnet 5.5 on the Anthropic API, but to Sonnet 4.5 on Bedrock, the Gemini Enterprise Agent Platform and Microsoft Foundry. Sonnet 4.5 is the model Anthropic retires on the Claude API on November 30, 2026, and Microsoft retires on Foundry the same day. Pin cloud-routed fleets explicitly.
· Sources: the Claude Code changelog and model configuration pages on code.claude.com, both living pages read October 15, 2026. The changelog dates each version.
availableModelsMatch: "exact"anddeniedModelsneed v2.1.283 or later; older clients ignore them- Set
requiredMinimumVersionso the allowlist binds - Bedrock and Google Cloud fall back to an earlier default, and Opus can fall to Sonnet; Foundry errors
- The
sonnetalias means Sonnet 4.5 on Bedrock, Google Cloud and Foundry - Opus 5.5 is the Claude Code default on Pro, Max, Team, Enterprise and the API
Agent governance Release notes and the ant CLI
An operator can now talk to a running agent instead of watching it finish being wrong. Three additions belong on an agent architect’s desk, and all three sit on rolling pages. Managed Agents permission policies gained an auto mode in which, in Anthropic’s words, “the server evaluates each agent or MCP tool call and runs it, denies it, or pauses for your approval.” A new evaluation field on the agent.tool_use and agent.mcp_tool_use events makes that decision visible in the event stream rather than inferred after the fact. The ant CLI 1.30.0 added ant apply, which declares agents, environments, skills, memory stores and deployments as files with a committed claude-lock.json lockfile. That’s infrastructure as code for an agent estate. It pairs naturally with Skills loading from GitHub: reusable behavior becomes reviewable in source control, and now so does the estate it runs in. The ant CLI 1.32.0 added ant beta:sessions connect, which attaches to a running agent session so an operator can message it and approve or deny in flight. None of these is anchored to a dated article, so confirm all three directly.
autopolicies evaluate each tool call and can pause for approval- An
evaluationfield on agent tool-use events makes the decision auditable ant applywith a committedclaude-lock.jsonlockfileant beta:sessions connectfor live attach, message and approve or deny
On-demand compaction September 14 beta
A 1M-token window is capacity, not recall. Compaction is Anthropic’s own admission that a long window is a thing to manage rather than a thing to fill. The Messages API can now return a signed compaction block summarizing the messages you sent, behind the beta compact-2026-09-04. You replay that block in place of the messages it covers. Recent turns are kept verbatim, and on models that preserve thinking, that thinking stays valid. It cuts both ways, though. The cache-hit column is where the money is, and compaction shortens the prompt while changing the prefix. So a workload tuned around a stable cached prefix should measure the net effect before assuming a saving. This is a beta on a rolling release note. Pin it, regression-test it, and confirm it directly.
- Signed
compactionblock, replayable in place of the messages it summarizes - Recent turns retained verbatim; preserved thinking stays valid
- Beta header
compact-2026-09-04 - Compaction changes the prefix, so re-measure your cache-hit ratio
New betas worth knowing API
Three beta surfaces are worth knowing. Per-message effort, behind the header mid-conversation-output-config-2026-07-01, is available on Fable 5.1, Mythos 5.1 and Opus 5. It lets one conversation move up and down the effort ladder instead of committing at the first turn. Turn-scoped system messages narrow a system instruction to a single turn. And thinking.display: "updates" changes how reasoning is surfaced. All three are beta. Regression-test before any of them sits in a critical path.
- Per-message effort on Fable 5.1, Mythos 5.1 and Opus 5
- Turn-scoped system messages
thinking.display: "updates"- Beta surfaces move; pin and re-test
Claude in Chrome Pro feature and Compliance API surface
A browser agent isn’t a feature of chat. It’s its own environment, and it has a different risk shape from a terminal agent or a governed desktop workspace. The untrusted content is the page itself. The credentials are already in the session. The blast radius is whatever the signed-in browser can reach. Claude in Chrome also sits on both sides of the buying decision. And it’s the surface most likely to be missing from an inventory built before it existed. On claude.com/pricing it’s a Pro plan feature, listed as “Claude in Chrome and Microsoft 365.” And it’s a Compliance API surface. Local session transcript retrieval for Claude in Chrome, with a product_surface value of claude_in_chrome, is still in beta for Claude Enterprise organizations. The Microsoft 365 surfaces left beta on September 24, 2026. That half comes from a rolling page, so confirm it directly before a retention process depends on it.
- A Pro plan feature on claude.com/pricing
- Compliance API transcript retrieval still in beta for Claude Enterprise, while Microsoft 365 local sessions went GA on September 24, 2026
- Page content is untrusted input, and the browser already holds the session
- Give it its own row in your environment and boundary inventory
Read the price column, because the names point the wrong way
The newer model costs less than the older one, twice over. Hold onto that while you read the tables. Haiku 5.5 is $0.10 / $0.50 per million tokens up to 100,000 tokens, against Haiku 4.5 at $1 / $5. Opus 5.5 is $4 / $20 and Opus 5 is $5 / $25. Sonnet 5.5 and Sonnet 5 both cost $2 / $10, but Sonnet 5.5 reads cache at half the price. Fable 5.1 still tops the generally available lineup at $10 / $50, and it’s still retention-required. And the “Legacy models” heading on claude.com/pricing holds ten models, most of them still Active. So don’t read a label. Read the price, the cache column and the lifecycle row.
| Model | Standard base input | 5m cache write | 1h cache write | Cache hit | Standard output |
|---|---|---|---|---|---|
| Claude Fable 5.1 | $10 | $12.50 | $20 | $0.25 | $50 |
| Claude Mythos 5.1, verification required | $10 | $12.50 | $20 | $0.25 | $50 |
| Claude Opus 5.5 | $4 | $5 | $8 | $0.20 | $20 |
| Claude Sonnet 5.5 | $2 | $2.50 | $4 | $0.10 | $10 |
| Claude Haiku 5.5, prompts up to 100,000 tokens | $0.10 | $0.125 | $0.20 | $0.01 | $0.50 |
| Claude Haiku 5.5, prompts over 100,000 tokens | $0.50 | $0.625 | $1 | $0.05 | $2.50 |
| Claude Fable 5 | $10 | $12.50 | $20 | $1 | $50 |
| Claude Mythos 5, invite only | $10 | $12.50 | $20 | $1 | $50 |
| Claude Opus 5 | $5 | $6.25 | $10 | $0.50 | $25 |
| Claude Opus 4.8 | $5 | $6.25 | $10 | $0.50 | $25 |
| Claude Sonnet 5 | $2 | $2.50 | $4 | $0.20 | $10 |
| Claude Haiku 4.5, Legacy | $1 | $1.25 | $2 | $0.10 | $5 |
| Model | Where it sits | Status on October 15, 2026 |
|---|---|---|
| Claude Fable 5.1 | Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry | Active, floor of “Not sooner than September 1, 2027”; the only Fable model in the models overview comparison table |
| Claude Opus 5.5 | Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry; Pro, Max, Team and seat-based Enterprise | Released September 22, 2026; Active, floor of “Not sooner than September 22, 2027”; not a Covered Model, so no retention requirement |
| Claude Sonnet 5.5 | Current Sonnet; named as the migration target for Sonnet 4.5 | Released September 28, 2026; Active, floor of “Not sooner than September 28, 2027”; not a Covered Model |
| Claude Haiku 5.5 | Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS; the default Haiku in Claude Code on the Anthropic API | Released October 7, 2026; Active (latest), floor of “Not sooner than October 7, 2027”; zero data retention position unpublished |
| Claude Mythos 5.1 | Not generally available; organizations verified through an Anthropic verification program, such as the Cyber Verification Program | Active (verification required), floor of “Not sooner than September 1, 2027.” Same specifications and pricing as Claude Fable 5.1 |
| Claude Mythos 5 | Not generally available | Active, floor of “Not sooner than June 9, 2027” |
| Claude Fable 5 | Still selectable and still priced at $10 / $50 | Under “Legacy models” on claude.com/pricing, but Active on the deprecations page, not sooner than June 9, 2027 |
| Claude Opus 5 | Still selectable and still priced at $5 / $25, above the newer Opus 5.5 | Released July 24, 2026; under “Legacy models” on claude.com/pricing, Active on the deprecations page, not sooner than July 24, 2027 |
| Claude Opus 4.8 | Still selectable and still priced, at exactly the Opus 5 rate | Legacy on claude.com/pricing, Active on the deprecations page, not sooner than May 28, 2027 |
| Claude Sonnet 5 | Still selectable at $2 / $10, the same base price as Sonnet 5.5, with a $0.20 cache read | Under “Legacy models” on claude.com/pricing, Active on the deprecations page, not sooner than June 30, 2027 |
| Claude Sonnet 4.5 | Serving until the retirement date, then not at all | Deprecated September 30, 2026, with retirement on the Claude API scheduled for November 30, 2026, the same date Microsoft Foundry lists. Anthropic recommends migrating to Claude Sonnet 5.5 |
| Claude Haiku 4.5 | First under “Legacy models” on claude.com/pricing; no longer in the models overview comparison table | Active on the Claude API, where claude-haiku-4-5-20251001 reached its floor of “Not sooner than October 15, 2026” with no deprecation notice; Microsoft Foundry retires it November 15, 2026 |
| Claude Opus 4.1 | Nowhere; the identifier stops serving requests | Retired August 5, 2026; Anthropic’s current migration pointer sends Claude Opus 5 and earlier to Claude Opus 5.5 |
Claude Fable 5.1 Still the top tier
The defaults are the story here. 1M token context is the default now, not a beta flag. Maximum output is 128k. And adaptive thinking is always on, so there’s no configuration in which this model isn’t reasoning before it answers. It’s still the top of the generally available lineup, at model id claude-fable-5-1. Anthropic’s models overview now reserves it for demanding reasoning and long-horizon agentic work, or for when your evals on Opus 5.5 at higher effort still fall short. Anthropic positions Fable 5.1 and Mythos 5.1 together as “the world’s most advanced models for coding and knowledge work.” That’s a vendor claim. Treat it as one until your own evaluation says otherwise.
- 1M token context by default; 128k maximum output
- Always-on adaptive thinking, not a toggle
- Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry
- $10 / $50 per million tokens, with cache hits at $0.25
- A Covered Model: 30-day retention required, with a per-workspace override for ZDR organizations
Claude Opus 5.5 New, September 22
The capable tier for work that can’t sit under a retention window. Opus 5.5 is claude-opus-5-5, at $4 / $20 per million tokens with a $0.20 cache hit, 1M token context and 128k maximum output. That rises to 300k on the Batch API in beta. Thinking is always on and adaptive, and the default effort is medium. Anthropic says it performs at Fable 5.1 level on most work, costs 40 percent less to run than Opus 5 and produces output more than 30 percent faster. Vendor claims, every one. What isn’t a vendor claim is the list of Covered Models, and Opus 5.5 isn’t on it. Anthropic says zero data retention is available. Three breaking changes against Opus 5. Thinking can’t be turned off at all now: sending disabled fails, and so does setting your own budget_tokens. The tool_choice types any and tool return a 400. And computer_20251124 returns a 400 on the Claude API and Google Cloud, where you need computer_toolset_20260801 instead, though Anthropic says the old tool keeps working on Amazon Bedrock.
- $4 / $20 per million tokens; cache hit $0.20, below Fable 5.1’s $0.25
- Not a Covered Model: no 30-day retention requirement, and ZDR available
- 1M token context, 128k maximum output, 300k on Batch API beta
- Thinking cannot be disabled and a manual budget_tokens also returns a 400
- On Pro, Max, Team and seat-based Enterprise with increased five-hour usage limits
Claude Sonnet 5.5 New, September 28
Same price as Sonnet 5, different behavior, and the named landing place for everyone leaving Sonnet 4.5. Sonnet 5.5 is claude-sonnet-5-5, at $2 / $10 per million tokens with a $0.10 cache hit since October 7, a $4 one-hour cache write, 1M token context and 128k maximum output. Anthropic says it’s more than 30 percent faster than Sonnet 5 and costs up to 30 percent less for most work. Vendor claims. The default effort differs by surface: medium in the Claude apps, high on the Claude Platform. So the same prompt is a different bill depending on where it runs, and nothing in your code shows that. Five changes break Sonnet 5 code. Forced tool use returns a 400. Disabling thinking at high effort or below is now thinking: {"type": "between_tools"} rather than "disabled". And the advisor tool rejects Claude Opus 4.8, Claude Opus 4.7 and Claude Sonnet 5 as advisors, which is an agent-architecture change dressed as a model release. Sonnet 5.5 isn’t a Covered Model, and Anthropic says zero data retention is available on all platforms.
- $2 / $10 per million tokens; cache hit $0.10, half the Sonnet 5 rate
- Default effort is medium in the Claude apps and high on the Claude Platform
- The advisor tool rejects Opus 4.8, Opus 4.7 and Sonnet 5 as advisors
- Forced tool use returns a 400, same as on the 5.1 pair and Opus 5.5
- The migration target Anthropic names for Claude Sonnet 4.5
Claude Haiku 5.5 New, October 7
The high-volume tier just dropped to a tenth of the price per token, with two strings attached. Haiku 5.5 is claude-haiku-5-5, or anthropic.claude-haiku-5-5 on Bedrock, at $0.10 / $0.50 per million tokens for prompts up to 100,000 tokens and $0.50 / $2.50 above that. Cache reads cost $0.01, then $0.05 past the line. String one: the threshold counts cached tokens, so long cached prompts land on the higher rate. String two: the tokenizer. The docs say the same text counts as roughly 30 percent more tokens than on Haiku 4.5, which is part of why Anthropic’s own average saving is around 75 percent rather than 90. Vendor claim. It has 1M token context, 128k maximum output and a June 2026 knowledge cutoff, with medium effort by default. It replaces Haiku 4.5, which is now Legacy, past its Claude API floor and set to retire on Microsoft Foundry on November 15, 2026. Code written for Haiku 4.5 can break on it, so treat the move as a migration, not a model-id swap. And don’t assume zero data retention. No Anthropic page states a position on it.
- $0.10 / $0.50 up to 100,000 tokens; $0.50 / $2.50 above
- Cache reads $0.01, or $0.05 past the threshold
- The same text runs about 30 percent more tokens than on Haiku 4.5
- Floor of not sooner than October 7, 2027
- Zero data retention position unpublished
Claude Mythos 5.1 Verification required
The access question has one answer now, and it runs through verification. Mythos 5.1 is the gated tier, at model id claude-mythos-5-1, with the same specifications and pricing as Fable 5.1. Its model page lists the status as Active (verification required). It says Mythos 5.1 goes only to organizations verified through Anthropic’s verification programs, such as the Cyber Verification Program, and tells you to apply to the program that covers your use case. The October 6, 2026 expansion of that program is the bigger news. It has three tiers: Defense Access, Red Team Access for organizations only, and Specialized Access, which Anthropic reviews in depth with the US government. Each tier reaches Opus 5.5, Sonnet 5.5, Mythos 5.1 and new models as they arrive. Existing Project Glasswing members move to Specialized Access without reapproval for current models. Two conditions bite. Data retention is required for enrolled organizations, so Anthropic can monitor for cyber misuse. And on Amazon Bedrock, the program is open only to customers eligible for Enterprise Frontier Safeguards. The life sciences route is separate. The Life Sciences Verification Program, given its own article on September 17, 2026, runs Standard Use grants renewed annually and High-risk Use grants renewed every six months, after a review of research credentials, security standards and ethical oversight. Its Standard Use grants name Mythos 5.1, Opus 5 and Sonnet 5, and the page hasn’t added the 5.5 models. A September 30 changelog entry on that page says its form is an expression of interest, not an application. The deprecations table gives Mythos 5.1 a floor of not sooner than September 1, 2027.
· Anthropic’s wording on the model page: “To request access, apply to the program that covers your use case.”
- Active (verification required), with the specification and price of Fable 5.1
- Cyber Verification Program tiers: Defense Access, Red Team Access and Specialized Access
- Each CVP tier reaches Opus 5.5, Sonnet 5.5 and Mythos 5.1
- CVP requires data retention, and on Bedrock it needs EFS eligibility
- LSVP Standard Use grants name Mythos 5.1, Opus 5 and Sonnet 5, not the 5.5 models
- Not generally available, so not a procurement option for most buyers
Claude Opus 5 Outclassed by its own successor
If a routing rule in your estate sends regulated traffic to Opus 5, it’s now sending it to the slower, dearer option. Opus 5 was released July 24, 2026 at $5 / $25 per million tokens with a $0.50 cache hit, 1M context and 128k maximum output. For two months it was the most capable Claude model with no mandatory retention requirement. Opus 5.5 took that position on September 22 at $4 / $20 with a $0.20 cache hit, with Anthropic claiming Fable 5.1 level performance and zero data retention availability. Cheaper on input, cheaper on output, far cheaper on cache reads. Opus 5 has also moved under the “Legacy models” heading on claude.com/pricing while staying Active on the deprecations page with a floor of not sooner than July 24, 2027. So it’s still selectable, still priced, still supported. It’s just no longer the answer to the question it used to answer. One comparison a buyer should still see: on the standard table, Opus 5 and Opus 4.8 are priced identically, line for line, and both now cost more than Opus 5.5.
- $5 / $25 per million tokens on the standard tier; cache hit $0.50
- No data-retention requirement, but Opus 5.5 offers the same at a lower price
- Under “Legacy models” on claude.com/pricing, Active to not sooner than July 24, 2027
- Priced identically to Opus 4.8 across every column
- Re-point regulated routing rules at Opus 5.5 and measure the difference
Claude Fable 5 Legacy in marketing, Active in lifecycle
Legacy in marketing. Active in lifecycle. Still priced and still selectable. A procurement policy that keys off the word “legacy” will read the marketing page and stop, so give a buyer all three parts. Fable 5 hasn’t been retired and hasn’t been repriced: $10 / $50 per million tokens on the standard tier, with a $1 cache hit. What changed is who is editorializing. The pricing page used to read as the unopinionated inventory beside an opinionated models overview, and it’s neither of those now. It carries a heading reading “Legacy models,” and that heading has been getting crowded. It opens with Haiku 4.5, then Sonnet 5, Opus 5, Fable 5, Opus 4.8, Sonnet 4.6, Opus 4.7, Opus 4.6, Sonnet 4.5 and Opus 4.5. Ten models. The platform pricing documentation doesn’t use that heading at all; it calls the same group “Additional models.” Most of those are still Active on the deprecations page, Fable 5 among them, with a tentative retirement of not sooner than June 9, 2027. One isn’t: Sonnet 4.5 is deprecated and retires November 30, 2026. So the heading tells you nothing about whether a model is about to stop serving. Neither page is wrong. They answer different questions. Fable 5 is also one of the four Covered Models, designated on June 9, 2026, so it carries the same 30-day retention requirement as the 5.1 pair and the same per-workspace override. Both pages are living documents, so check them on the day.
- Still active and still priced at $10 / $50 on the standard tier
- Under the “Legacy models” heading on claude.com/pricing, which holds ten models
- Active on the deprecations page, not sooner than June 9, 2027
- Cache hits cost $1, four times the 5.1 rate and five times the Opus 5.5 rate
- A Covered Model since June 9, 2026, so retention-required
Claude Sonnet 5 Same price, newer sibling
Sonnet 5 costs what Sonnet 5.5 costs on the sticker, and more on the cache. Both are $2 / $10 per million tokens. But since October 7, Sonnet 5.5 reads cached context at $0.10 while Sonnet 5 still charges $0.20. So there’s no price reason to stay, and now there’s one to leave. The pricing question that hung over Sonnet 5 was settled by the calendar rather than by a promise: the scheduled September 1 increase didn’t happen, and $2 / $10 is the standard price. Since then Sonnet 5 has moved under “Legacy models” on claude.com/pricing while staying Active on the deprecations page to not sooner than June 30, 2027. There’s one reason to look before you jump. Moving to Sonnet 5.5 isn’t a drop-in: five changes break Sonnet 5 code, including forced tool use and the advisor tool refusing Sonnet 5 itself as an advisor. Budget the migration, not just the tokens.
- $2 / $10 standard, the same as Sonnet 5.5, but a $0.20 cache hit against $0.10; $1 / $5 on the batch tier
- Under “Legacy models” on claude.com/pricing, Active to not sooner than June 30, 2027
- Five breaking changes stand between Sonnet 5 code and Sonnet 5.5
- The advisor tool on Sonnet 5.5 rejects Sonnet 5 as an advisor
- Claude Code publishes per-plan defaults at version level; the Claude apps don’t
Claude Opus 4.8 Two pages, two labels
The legacy label isn’t on the page a buyer would check for it. The models overview no longer has a Legacy heading at all. Its four-row comparison table is now Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. And the migration pointer on it has moved twice in ten weeks. It now tells anyone on Claude Opus 5 or earlier to read Anthropic’s migration notes for Claude Opus 5.5. That sentence quietly includes Opus 4.8. The same page tells an undecided buyer to start with Claude Opus 5.5 for most workloads. Fable 5.1 is for demanding reasoning and long-horizon agentic work, or for when evals on Opus 5.5 at higher effort still fall short. A different first-party page carries the label. Under “Legacy models” on claude.com/pricing, Opus 4.8 sits at $5 / $25. Beside it are Sonnet 4.6 and Sonnet 4.5 at $3 / $15, and Opus 4.7, 4.6 and 4.5 at $5 / $25. And the deprecations page still shows Opus 4.8 as Active, not sooner than May 28, 2027. So the accurate three-way position is legacy in marketing, active in lifecycle, still priced. And priced identically to Opus 5 on the standard table, which means both now cost more than the newer Opus 5.5. Opus 4.8 isn’t described as a safety fallback anywhere readable. One more thing an agent architect should catch: on Sonnet 5.5 the advisor tool rejects Opus 4.8 as an advisor outright.
- Legacy on claude.com/pricing, Active on the deprecations page, still priced
- Identical standard pricing to Opus 5, and both cost more than Opus 5.5
- The migration pointer now sends Claude Opus 5 and earlier to Claude Opus 5.5
- The models overview recommends starting with Opus 5.5 for most workloads
- On Sonnet 5.5, the advisor tool rejects Opus 4.8 as an advisor
| Plan | Price as published | Model availability | Notable inclusions |
|---|---|---|---|
| Free | No cost | Sonnet and Haiku. Fable is shown as “No” | Claude Design, Slides, Docs. Family names only; the page never names a version number |
| Pro | The page gives two annual figures that do not reconcile, $17 per month billed annually and $200 up front, against $20 monthly | Opus, Sonnet and Haiku, with Fable on “Usage credits.” Opus 5.5 is included, with increased five-hour usage limits | Everything in Free, plus Claude Code, Claude Science, Projects, Claude Dashboards (beta), task hand-off and scheduling, and Claude in Chrome and Microsoft 365. Claude Code defaults to Opus 5.5 |
| Max | From $100 per month | Choose 5x or 20x more usage than Pro; Opus 5.5 included with increased five-hour usage limits; Fable use capped at 50% of weekly limits | Monthly API credits from October 7, 2026: $100 on Max 5x, $200 on Max 20x. Higher output limits and priority access at high traffic times |
| Team | $20 per seat per month billed annually, $25 monthly; premium seats $100 billed annually, $125 monthly | Premium seats carry 5x more usage than standard seats, with Fable use capped at 50% of weekly limits; Opus 5.5 included | For teams of 2 to 150; includes Claude Code and Claude Cowork. Monthly API credits: $20 per Standard seat and $100 per Premium seat, pooled and capped at $500 |
| Enterprise | Seat price plus usage that scales with model and task. US$20 per seat per month billed annually, 20 seat minimum, as read on claude.com/pricing on October 15, 2026; identical to a standard Team seat and unexplained | Fable, Opus, Sonnet and Haiku all marked “Yes,” in both self-serve and sales-assisted columns | The Compliance API surfaces; Claude Design, Slides and Docs in beta and off by default; not eligible for the monthly API credits |
Claude for Teachers Dated, with a changelog
Verified US K-12 educators get free premium access, and the sign-up window runs through June 30, 2027. This one is easier to trust than most. The article is dated July 14, 2026 and carries a published changelog, with updates stamped July 21 and August 28, 2026. So a change to the window would leave a trace on the page rather than happening silently.
- Announced July 14, 2026, with dated updates July 21 and August 28, 2026
- Sign up by June 30, 2027 for a full year of access
- Verification required
- A recruiting channel as much as a product
Self-serve HIPAA Process, not scope
Self-serve is a process change, not a scope change. HIPAA configuration became self-serve for Enterprise and Platform customers on July 14, 2026, which takes a manual step out of healthcare deployments. Read it beside the Covered Model retention requirement, the per-workspace override and the platform split on where retained data sits, before you assume the two compose.
- Configure without a custom process
- Still requires your own compliance review
- Read it beside the Covered Model retention requirement
- Confirm scope before moving regulated data
Claude is now a family of work surfaces
Choosing Claude isn’t choosing a model any more, right? The real question is where the work should run. Conversation, terminal, governed desktop workspace, team channel, design and document surface, the browser, or a business application. The browser is the newest of those, and the one most likely to be missing from an inventory you already built.
| Surface | Best for | Context advantage | Primary caution |
|---|---|---|---|
| Claude chat | Analysis, writing, research and files | Conversation, uploads, connectors and memory | Large context still needs coverage checks |
| Claude Code | Repository-scale engineering | Codebase, terminal, tools and project instructions | Command and prompt-injection risk |
| Claude Cowork | Desktop knowledge work | Selected folders, local apps, skills and plugins | Mount permissions and action review |
| Claude Tag | Shared team delegation | Channel history, scoped tools and shared memory | Spend, permissions and public mistakes |
| Claude in Chrome | Work that lives in a web application | The open page, the signed-in session and the browser’s own reach | Page content is untrusted input; transcript retrieval is beta |
| Claude Design | Rapid visual exploration | Brand inputs, conversation and prototypes | Still a Labs product; professional finishing required |
| Claude Slides and Claude Docs | Decks and written deliverables | The analysis already in the conversation | Listed from Free upward; beta and off by default on Enterprise |
| Claude Science | Auditable scientific work | Packages, tools, compute and artifacts | Expert validation remains essential |
From source material to finished work without changing context
People rebuild the same brief four times a day in four different tools. That’s the waste Anthropic is aiming at. Claude is meant to follow the work across files, applications and teams instead of making anyone start over on each surface.
Claude Code
Plans, edits, tests and debugs inside the real repository.
- Agent teams and subagents
- Long-running workflows
- Code review and verification
Claude Cowork
Runs multi-step tasks against selected desktop files and applications.
- Read-only or controlled write modes
- Reusable skills and plugins
- Dispatch from text or voice
Excel, PowerPoint and Word
Claude works directly in business documents and carries context between applications. Write tools shipped on July 7, 2026: Claude can draft and send email, act on the calendar and create or update files.
- Model, audit and explain spreadsheets
- Build decks from live analysis
- Review write permissions before granting them
Claude Tag
A shared team agent in Slack with scoped identity and memory.
- Channel-specific access
- Shared continuation
- Logs and budget controls
Claude in Chrome
Works inside the browser on web applications that never exposed an API. A Pro plan feature, and now a Compliance API surface in beta for Claude Enterprise.
- Read and act on the open page
- Page content is untrusted input
- Scope the signed-in session deliberately
Claude Design, Slides and Docs
Explores prototypes and polished visual directions through conversation, and produces decks and documents in the same place. The claude.com/pricing page lists the three together on the Free plan.
- Software experience ideation
- Slides, docs and one-pagers
- Canva handoff
MCP and apps
Pulls live context from enterprise and creative systems.
- Google Drive and Gmail
- Market and portfolio data
- Blender and specialist tools
Skills and plugins
Packages instructions, tools and workflows into repeatable organizational capability. The Skills API left beta across the SDKs on August 27, 2026.
- Reduce prompt folklore
- Standardize recurring work
- Deploy domain templates
Memory and context
Carries relevant context across sessions and, increasingly, across applications. On July 10, 2026 memory was reworked into individual categorized entries, replacing the daily memory summary.
- Preserve audience and standards
- Inspect and remove entries one at a time
- Review what should persist
Claude is becoming useful before, during and after the creative tool
Claude doesn’t replace taste. What it does is widen the number of directions a creative team can try, take the repetitive production off their hands, and keep the strategy attached to the thing being made.
Claude Design Labs
Explore software experiences, prototypes, slides and one-pagers with visual iteration inside the conversation.
- Start with audience and desired behavior
- Include brand rules and references
- Export into professional finishing tools
Creative connectors MCP
Claude can work with tools such as Blender through connectors, scripts and the application’s own APIs.
- Analyze and debug scenes
- Batch repetitive changes
- Create custom tools inside the creative application
Cross-format production
Claude can turn research into a document, document into deck, data into charts and creative brief into a buildable prototype.
- Keep one narrative thread
- Reduce manual reformatting
- Use humans for taste and final selection
Build three distinct directions for this onboarding flow. Use our brand system, explain the behavioral idea behind each route, and make every prototype testable rather than decorative.
Inspect this Blender scene and identify the repetitive production steps. Propose a reversible automation plan, then implement only the steps I approve.
Turn this research and workbook into a 12-slide executive deck. Use the approved master, keep one claim per slide, and link every chart to its source.
Do not redesign this. Review it as a creative director: identify where the execution departs from the stated concept, audience or brand behavior.
Eight habits that make Claude substantially better
A better model doesn’t make a team better, right? Behavior does that. The upgrade matters, but what people do with it decides whether all that capability turns into useful work, a bigger bill or a real risk.
Front-load the context
Give Claude the complete source pack and explain how the files relate before asking for conclusions.
Let it inspect the brief
Ask what’s missing or contradictory before it starts producing anything.
Choose effort deliberately
Use lower effort for reversible work and escalate only when consequence justifies it. On 5.1, adaptive thinking is always on, so effort is the only lever left.
Define completion
Name the artifact, the checks it has to pass and the evidence report. Not just the topic.
Keep context across stages
Move from analysis to file to deck without rebuilding the narrative in a new chat.
Use an adversarial second pass
Ask Claude to name the weakest claims, the broken assumptions and the missing tests in its own work.
Name the audience and stakes
Claude changes depth, tone and caution when it knows who will use the result and what failure costs.
Trust the loop, inspect the result
Don’t micromanage every step. Do review the commands, diffs, formulas and finished artifacts.
Produce the completed board memo, not an outline. Validate every recommendation against the attached evidence and include a one-page appendix of unresolved assumptions.
Add pagination to the users API, update the contract and documentation, add edge-case tests, run the full suite, and stop only when CI passes or you can name the blocker precisely.
Re-read your work as a skeptical domain expert. Identify the three conclusions most likely to be wrong and show what evidence would change each one.
Before synthesizing, create a coverage map of every supplied document. After the answer, list any sources or sections you did not materially use.
The more Claude can act, the more its environment matters
A weak chat answer is annoying. A tool-using agent with broad file and network access can do real operational damage. Different failure mode, right? Two more categories belong here. One, a model generation that fails requests your current code makes legally today. Two, a set of first-party contradictions a control document can inherit without anyone noticing.
Scope your breaking-change sweep to model ids, not to “the 5.1 models”
A sweep that was correct in September is wrong today, and that’s the point. The tool_choice types any and tool now return a 400 on four models, not two: Fable 5.1, Mythos 5.1, Opus 5.5 and Sonnet 5.5. Opus 5.5 adds two more. Thinking can’t be turned off at all, so both disabled and a manual budget_tokens fail. And computer_20251124 returns a 400 on the Claude API and Google Cloud, where computer_toolset_20260801 is required, while Anthropic says the old tool keeps working on Amazon Bedrock. Sonnet 5.5 adds five of its own, including thinking: {"type": "between_tools"} in place of "disabled" at high effort or below, and an advisor tool that refuses Opus 4.8, Opus 4.7 and Sonnet 5 as advisors. Haiku 5.5 brings its own list against Haiku 4.5: budget_tokens, the sampling parameters and assistant prefill can each return a 400. Separately, the four Covered Models require 30-day retention, so a request from an organization or workspace without retention enabled returns a 400 invalid_request_error. That last one is a configuration problem with a documented fix, not a wall: see the workspace override above. Test every path before you switch a model id in production.
Zero data retention is not a hard gate on 5.1
The 30-day requirement on Fable 5.1 and Mythos 5.1 reads like a wall. It isn’t one. A zero data retention organization can turn on 30-day retention for a single Console workspace while every other workspace stays at zero. The path is the Console’s Data retention page, where you pick the workspace and select Standard data retention. Use any workspace except the Default Workspace, which always follows the organization’s setting. If a procurement file anywhere says no configuration flag fixes this, that’s the sentence to go and fix first. The requirement is real and the failure mode is a clean 400. So the work is a configuration decision and a scoping decision, not an exception request.
Nobody publishes who is eligible for interim zero data retention
Anthropic’s September 1 news pages say eligible customers will be able to use Fable 5.1, and Fable 5, with zero data retention until Enterprise Frontier Safeguards is ready. The September 1 API release note and the retention documentation both say the models aren’t available under zero data retention unless expressly authorized by Anthropic. Anthropic’s October 6 Cyber Verification Program article uses the eligible wording too. No page defines eligibility. No page says whether eligible and expressly authorized are the same status. This is an unresolved conflict between first-party sources rather than a stale claim, so treat it as one. Don’t let either sentence into a control document on its own.
The retention answer changes with the route you buy through
Read this beside the two cards above, because it’s a different kind of conflict. Those are Anthropic pages disagreeing with each other. This is Anthropic’s direct terms and Microsoft’s subprocessor terms describing the same model and not reaching the same outcome. On Anthropic’s own pages, Fable 5.1 is a Covered Model with a 30-day retention requirement. On Microsoft’s side the off-by-default retention gate attaches to Fable 5.0. And for certain organizations Fable 5.1 runs with Anthropic as a subprocessor under Microsoft’s Product Terms and Data Protection Addendum. There, Microsoft states that Anthropic doesn’t retain customer content. Neither source resolves the other. Don’t move Microsoft’s sentence into an Anthropic control description, or the reverse. And don’t let a procurement file inherit a retention posture from whichever page it read first. Establish the contracting route, then take the retention answer.
· Source: “Anthropic models in Microsoft Online Services,” Microsoft Learn, a living page stamped September 18, 2026.
EFS is a promise with no dated milestone, and the two pages disagree on tense
A plan with no delivery date is a different animal from a plan with a date that slips. Enterprise Frontier Safeguards is described as giving privacy equivalent to zero data retention, by keeping customer data on the customer’s own cloud infrastructure. The scope named is Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, the Claude Platform on AWS, the Gemini Enterprise Agent Platform and Microsoft Foundry. The September 1 article says the rollout happens in phases starting later this fall. The request form says EFS is rolling out to eligible customers in phases. Future tense on one page, present tense on the other, and no phase, customer or date on either. Anthropic’s October 6 Cyber Verification Program article, the newest dated page to mention EFS, still uses future tense. Whether any phase has shipped is an open question, and it’s open because the only two pages that would carry the answer contradict each other.
Prompt injection
Instructions inside email, documents, websites or pasted prompts can redirect an agent. Treat external content as untrusted input.
Mount and write access
Cowork permissions should match the task. Read-only, read-write and no-delete boundaries are meaningful controls.
Higher effort cost
More reasoning helps on hard work. It also adds latency, token use and the odd pointless detour. On 5.1 and on Opus 5.5 you can’t turn thinking off at all. And on Sonnet 5.5 the default effort is medium in the Claude apps but high on the Claude Platform. So the same prompt costs more through the API than it does in the app.
Preview instability
Dynamic workflows, Labs products and beta surfaces require regression testing before production use. Per-message effort, turn-scoped system messages and the updates thinking display are all beta.
Long context omissions
A 1M-token window is capacity, not proof of complete recall or equal attention. That’s now the default window on 5.1 rather than an opt-in, so the coverage discipline applies to every call.
No native media generation
Claude can plan, analyze and prototype visuals, but final image, video and audio generation still depends on other tools.
Model-version drift, and one failure that returns a 200
A model upgrade can change formatting, tool choices and reasoning style. Re-test critical prompts and workflows. Watch for the quiet one. A thinking block that Sonnet 5.5 produced works only in the account that produced it or a linked account. A block sent from any other account is dropped before the model sees it, and the request still succeeds. No error, no warning, just a slightly worse answer. Haiku 5.5 blocks are bound to the producing or linked account too, though Anthropic doesn’t say how a block from another account fails. Anywhere you share cached assistant turns across accounts, compare outputs rather than status codes.
Diplomatic language
Claude may soften criticism unless directness is explicitly requested. Ask for the harshest defensible review.
Opus 4.1 is gone, and the replacement chain has moved twice
If you haven’t hit this error yet, your pinned job isn’t running, and that’s its own finding. On Anthropic’s deprecations page, read October 15, 2026, claude-opus-4-1-20250805 still shows status Retired with the line “This model was retired August 5, 2026.” Anything still pinned to the 4.1 identifier, an integration, a saved evaluation or a scheduled job, is failing now as a request error rather than a warning. Watch the destination, because it keeps moving. The deprecations row still names claude-opus-4-8 as the replacement. The models overview now says anyone on Claude Opus 5 or earlier should migrate to Claude Opus 5.5. That’s a different target from the one it named in July. And Opus 4.8 and Opus 5 are both filed under “Legacy models” on claude.com/pricing. Migrating 4.1 to 4.8 today buys you one hop into a bucket you’ll have to leave twice.
A floor is not an announcement, in either direction
Floors cut both ways. On September 30, the day after its floor expired, claude-sonnet-4-5-20250929 became a deprecation with a firm retirement of November 30, 2026, 46 days from October 15. Haiku 4.5, claude-haiku-4-5-20251001, reached its floor, “Not sooner than October 15, 2026,” and on the Claude API nothing happened. No deprecation row, no history entry, no API release note, no status notice. The rest of the evidence points one way. Haiku 5.5 shipped October 7. Haiku 4.5 moved under “Legacy models” on claude.com/pricing and left the models overview comparison table. But that’s inference, not an announcement. So don’t read the silence as a reprieve, and don’t read the pressure as a date. Start the Haiku 5.5 migration and watch the deprecations page. On the Claude API, Opus 4.5, claude-opus-4-5-20251101, carries “Not sooner than November 24, 2026,” 40 days from October 15, with no notice either. Microsoft Foundry is the exception: its retirement schedule ends claude-haiku-4-5 on November 15, 2026 and claude-opus-4-5 on November 24, 2026.
Python SDK v1.0 will stop a build
The August 20, 2026 release moves httpx to httpx2, requires Python 3.10 or later, removes Text Completions and removes temperature, top_p and top_k from Messages. An unpinned dependency will pick this up on its own schedule, which is rarely yours.
The Workbench URL is dead
The legacy Workbench at platform.claude.com/workbench sunset on August 17, 2026 and the replacement Playground at platform.claude.com/playground was documented in the August 18, 2026 release note. Three experimental /v1/experimental/*_prompt endpoints retired on the same August 17 date. Internal links, onboarding docs and screenshots will still point at the old surface.
July 24 breaking changes still apply
Setting thinking to disabled is only valid at effort high or below. Combine it with extra high or max and you get a 400. Requesting fast speed on Opus 4.7 errors. Both break a working integration quietly, right up until a call fails.
A default moved underneath Pro and Team budgets, and the Sonnet alias points at a retiring model
Claude Code’s model configuration page sets Opus 5.5 as the default on Pro, Max, Team, Enterprise and the Anthropic API. If you sized Pro or Team Claude Code usage against a Sonnet default, that assumption expired without anyone telling the budget holder. The cloud routes are the sharper problem. The sonnet alias resolves to Sonnet 4.5 on Amazon Bedrock, the Gemini Enterprise Agent Platform and Microsoft Foundry, and Foundry’s default model is Sonnet 4.5. That’s the model Anthropic retires on the Claude API on November 30, 2026, and Microsoft retires on Foundry the same day. The Claude apps are vaguer. The July 24 announcement put Opus 5 as the Claude Max default and left Sonnet 5 on Free and Pro, and no current page restates either at version level. Pin a model id in anything you evaluate, or your baseline changes underneath the test.
The safety fallback has no published name, and two categories rather than one
Anthropic says dual-use cybersecurity work is redirected “to our Opus models,” plural and unnamed, naming penetration testing, exploit generation and binary-based vulnerability scanning as examples. It also says queries related to research and development in the life sciences are directed to the Opus models. No first-party page names a specific fallback model. Two things follow. If a runbook, incident playbook or user-facing message in your product names one, it’s repeating a guess rather than a documented fact. And if your safety documentation describes routing as a cybersecurity mechanism only, it will surprise a research team the first time a life sciences query behaves differently from the rest.
Stolen API keys and clients spoofing Claude Code
Treat an API key as a production credential, not a convenience. A stolen key lets an attacker run operations at your expense while the activity is attributed to you. That’s a cost exposure and an attribution problem in the same incident review. Anthropic’s September 10, 2026 threat intelligence report also describes credential harvesting through fake Claude resellers, and malicious clients that spoof “Claude Code.” That second one is a supply-chain warning about the tool developers install most casually. So name it in an internal advisory: install Claude Code from Anthropic’s own distribution and nowhere else.
Records built on the Compliance API Activity Feed changed shape retroactively
This is the rare vendor change that reaches backward into data you already collected. From September 24, 2026 the Compliance API Activity Feed no longer returns file names, project document names or artifact titles. The filename and title fields are always empty or omitted, including on activities recorded before the change. So a retention or e-discovery process that passed its test in September can produce different output today against the same period, with no error to tell you. Names must now be resolved by ID with a Compliance Access Key carrying read:compliance_user_data. Re-test any records workflow that reads those fields, and re-test it against historical activities, not just new ones.
The anthropic-version header on Enterprise endpoints
Check what your client sends, not what your notes say it sends. The September 1 release notes say the Admin API Claude Enterprise endpoints, the Analytics API and the Compliance API guides now show the anthropic-version header. They instruct you to send it on every request to these endpoints. An integration built against the August 26 Compliance API release may not be sending it. This is a rolling release-note item.
Date Claude Code claims from code.claude.com, not the GitHub redirect
Two addresses, two different records. The address platform.claude.com/docs/en/release-notes/claude-code still redirects to a changelog on GitHub, numbered by version with no dates. But code.claude.com publishes its own changelog with a date on every version: 2.1.293 on October 7, 2026, for example, and 2.1.300 on October 14. It also runs a weekly “What’s new” digest labeled by date range. So cite the code.claude.com page. One gap remains. That changelog lists versions back to 2.1.290, dated October 5, so the 2.1.280 to 2.1.286 items can’t yet be dated from a first-party page. The Opus 5.5 default change, availableModelsMatch and the fallback behavior are among them. Check those against your installed version. Three other items rest on rolling pages rather than dated articles: the self-serve HIPAA item, the August 17 Workbench sunset with the August 18 Playground rename, and the Python SDK v1.0 release.
· Source: “Changelog,” Claude Code Docs, code.claude.com/docs/en/changelog, a living page read October 15, 2026.
Vendor benchmark claims, and there are more of them now
Anthropic made a lot of claims in fifteen days and none of them has been independently checked. Opus 5.5 performs at the level of Fable 5.1 on most work, costs 40 percent less to run than Opus 5 and produces output more than 30 percent faster. Sonnet 5.5 is more than 30 percent faster than Sonnet 5 and costs up to 30 percent less for most work. Haiku 5.5 costs around 75 percent less to run than Haiku 4.5 on average, on a tokenizer that counts the same text as about 30 percent more tokens. Fable 5.1 and Mythos 5.1 are the world’s most advanced models for coding and knowledge work. Opus 5 sits within 0.5% of Fable 5 on the benchmark Anthropic cites. Every one of those is the vendor marking its own homework. The Opus 5.5 claim is the expensive one to get wrong, because a whole tiering decision now hangs off it. Run a task-level evaluation before you move production work up or down a price tier on the strength of an adjective.
The Usage Policy changes on November 12, 2026
A policy update is a contract change you didn’t negotiate. Anthropic published its 2026 Usage Policy update on October 8, 2026, and it takes effect November 12, 28 days from October 15. Three changes reach a buyer. The high-risk use case section now lists which recommendations are covered. Where Claude controls hardware that takes autonomous physical actions, new requirements apply, including a safe state if Claude is disconnected. And the Supported Regions rules now reach entities “majority-owned or controlled by persons or entities in those regions.” If your ownership chain crosses a border, check it. The surveillance and law-enforcement rules were rewritten too. Read the new text against your use cases before November 12, not after.
· Anthropic’s wording: “the equipment must also be able to hold a safe state if Claude is disconnected.”
| Environment | Useful boundary | What still requires review |
|---|---|---|
| Claude chat | Ephemeral server-side container | Files produced, connector scope and claims |
| Claude Code | Local sandbox plus human command approvals | Shell commands, diffs, secrets and outbound access |
| Claude Cowork | Sealed VM and selected folder mounts | Mounted workspace, MCP servers and completed actions |
| Claude in Chrome | The browser tab and the signed-in session it already holds | Page content as untrusted input, form submissions, and what the session can reach |
| Claude Tag | Channel-scoped tools, memory and spend | Public responses, tool actions and cross-team access |
Where these facts come from
Product status, availability and pricing move fast. Below are the dated first-party articles behind the facts above, plus the deprecations table, which is the only authoritative record of retirement dates. Several load-bearing facts sit on living pages instead: the retention documentation, the Covered Models article, both pricing pages, the models overview, the Claude Code changelog and the rolling API release notes. Those pages carry no reliable publication date, so they’re attributed inline rather than listed. The cards after the list say which facts rest on which page, and where Anthropic’s own pages disagree.
Retention rests on living pages and two dated ones
The per-workspace 30-day setting, the two-year flagged-content figure and the three six-year tails come from Anthropic’s API and data retention documentation, a living page with no date. The two-year and seven-year figures also appear on two Anthropic Privacy Center articles dated July 1, 2026. The Covered Models list comes from a support article that shows “September 1, 2026” under its title without saying what that date marks. Read the living pages before a control description quotes them.
Prices, floors and the lineup are readings
The plan and seat structure, the API credits and the “Legacy models” heading come from claude.com/pricing. Per-token rates, including the two Haiku 5.5 price bands and the $0.10 Sonnet 5.5 cache read, come from the platform pricing documentation, which groups older models under “Additional models.” Every status and floor comes from the deprecations table listed above. The four-model comparison, the Opus 5.5 recommendation and the migration pointer come from the models overview. All of these are rewritten in place.
Operational changes sit on rolling release notes
The September 22, September 28 and October 7 launch entries and their breaking changes. The October 7 Sonnet 5.5 cache cut. The September 30 Sonnet 4.5 deprecation. The September 24 refusal billing and both Compliance API changes. Cache diagnostics on September 23. The anthropic-version header, the compaction beta, the Managed Agents auto policies, the Workbench sunset, Python SDK v1.0, the three betas and the Sonnet 5 price confirmation. Anthropic rewrites those notes in place, so none of them can anchor a date.
Claude Code items are dated on code.claude.com
The Claude Code changelog on code.claude.com dates every version it shows, and the model configuration page there publishes per-plan and per-provider defaults and aliases. Both are living pages. The changelog lists versions back to 2.1.290, dated October 5, 2026. Items from 2.1.280 to 2.1.286 sit outside that range, so they can be ordered but not dated. The ant CLI items rest on rolling notes too. Check each against your installed version.
Where Anthropic’s pages disagree
Two disagreements are live. The September 1 news pages and the October 6 Cyber Verification Program article speak of eligible customers using zero data retention, while the September 1 API release note and the retention documentation say not unless expressly authorized by Anthropic. No page publishes the eligibility criteria or reconciles the two. And the EFS article speaks in future tense while the request form says EFS is rolling out. On Opus 5.5, the Opus 5.5 and Sonnet 5.5 announcements both say zero data retention is available, and the Covered Models article and the retention documentation agree by leaving Opus 5.5 off their lists. One dated article and the live pricing page differ on a price: the September 28 Sonnet 5.5 announcement prints $0.20 cache reads, and the pricing page prints $0.10.
Microsoft’s subprocessor terms rest on a Microsoft page
Microsoft’s subprocessor terms for Fable 5.1, its up-to-30-day wording and its data-retention category naming Fable 5 and Mythos 5 come from “Anthropic models in Microsoft Online Services” on Microsoft Learn, stamped September 18, 2026. Microsoft rewrites the page in place, so read it again before it goes near a control document.
When a dated article and a live page disagree
A dated article is authoritative for what was announced and when. A live page is authoritative for what is true today. When they disagree, the live page governs the present and the article governs the record. Keep your own dated copy of the live page, because it carries no history.
· Example: the September 28, 2026 Sonnet 5.5 article records cache reads at $0.20. The live pricing page governs today, and it shows $0.10 from October 7, 2026.
AI Mindset