When the General Model Is Not Enough
Buy Cursor today and the invoice comes from X Subscription Corp, not from Cursor. Since October 12, 2026, Cursor has been part of xAI’s The X Subscription, one plan for Grok, Cursor and X. Plus costs $8 a month and carries no third-party models at all. Premium is $30, Super $100, Ultra $200, and Teams starts at $30 a seat. So the seller and the model list moved on the same day. In health, OpenEvidence’s privacy policy still says user questions aren’t covered by HIPAA, while its 2025 announcement says it’s fully compliant. That’s this category. Specialists are purpose-built tools for one domain. The domain layer is what you buy. The policy page is what you sign up to.
A different question from “which model is best”
General assistants are brilliant generalists. But in regulated, data-heavy or workflow-bound work, the less clever tool often wins. Strange, right? It wins because the domain is already inside it: the case law, the codebase, the EHR, the data room. The 2026 market has a credible specialist for most professional functions. The skill is knowing when to reach for one.
Domain layer over a frontier model
Here is the part the demo skips. Many vertical tools run on the same OpenAI or Anthropic models as your general assistant. The brain is often rented. Not always. Sierra now names in-house models, Curie and Fleming, and Cursor calls Grok, built by its sister company, its own. Either way, what you pay for is the layer on top: proprietary data, workflow, integrations, evaluation and accountability.
- The intelligence is often rented
- The domain layer is the product
- Judge the layer, not the demo
Depth a generalist cannot fake
A general chat starts every conversation cold. A specialist already holds your repository, your matter history, your permission model, your compliance posture and the workflow the work actually follows. You can’t fake that with a better prompt.
- Context that lives in your systems
- Guardrails for regulated work
- Integrations into systems of record
Price, lock-in and overlap
Specialists cost more. They embed your data deeply. And they often overlap with the general assistant you already pay for. So the call is economic and architectural. Capability is only one input.
- Your Copilot may already do 80%
- Migration is rarely easy
- Buy only for the part that clears the bar
A specialist for most professional functions
Pick a professional function. There’s probably a well-funded, credible specialist for it, doing work a general assistant only skims. That’s the shape of the market in late 2026. Names change fast, and so do the boundaries. So read this as a snapshot of categories, not a permanent leaderboard.
| Domain | What the specialist adds | Leading examples |
|---|---|---|
| Coding | Repository context, autonomous multi-file agents, CI and pull-request integration | Cursor (owned by SpaceX, sold through The X Subscription), Cognition (Devin) |
| Legal | Case law, contracts, firm knowledge, citation discipline and conflict-of-interest enforcement | Harvey |
| Finance and research | Deep analysis across private documents, public filings and licensed market data, now delivered into other tools through an API and MCP server | Hebbia |
| Customer service and long-horizon work | Outcome-priced agents across chat, voice, email and messaging, now extending to goals that run for weeks | Sierra |
| Enterprise knowledge | Permission-aware search and agents across all company apps, with an agent identity model Glean’s release notes call generally available and its docs still label beta | Glean |
| Healthcare | Ambient clinical notes inside the EHR, extending into workflow and revenue cycle; evidence at the point of care | Abridge, OpenEvidence |
The most mature vertical, and the most contested
Software is where vertical AI has gone furthest. It’s also where the general assistants fight hardest. Claude Code and OpenAI Codex are the generalist coding agents. Cursor and Cognition are the purpose-built ones. And this is the one domain on the map where you can see prices. Cursor publishes a tier list, though since October 12, 2026 it belongs to The X Subscription and X Subscription Corp sells it. Cognition publishes one too. Both stop at Enterprise. Cursor’s is custom. Cognition bills Enterprise in Agent Compute Units at an order-form rate it doesn’t publish. So the list price is where your negotiation starts, not where it ends.
Cursor IDE agent · SpaceX-owned
Cursor sells a coding agent for ambitious software: deep codebase understanding plus autonomous, parallel agents. Then the ground moved. On October 12, 2026 Cursor joined The X Subscription, xAI’s single plan for Grok, Cursor and X. New subscriptions are sold only through X Subscription Corp. And models are now gated by tier. Plus gets Cursor’s own models only: Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5, at fixed settings. Third-party models start at Premium and bill at API list price from prepaid credit. So the question isn’t which model your engineers prefer. It’s which tier unlocks that model, and who sells you the tier. Ownership sits underneath all of it. SpaceX acquired xAI on February 2, 2026 and Cursor on August 14, 2026, so Cursor and its Grok supplier are sister companies. Grok 4.7 arrived in Cursor on September 21, and xAI prices it from $2 per million input tokens and $6 output. On data, Privacy Mode is on by default for Teams and Enterprise. No Cursor page states the default for individual plans.
· Cursor’s help page: “Cursor is part of The X Subscription, one subscription for Grok, Cursor, and X.”
· On Teams and Enterprise, “third-party model requests include a Cursor Token Rate of $0.25 per million tokens.”
· cursor.com/security, last updated August 25, 2026: “When enabled, we will not train on your data.”
- Sold through X Subscription Corp since October 12, 2026
- Plus runs Cursor’s own models only; third-party from Premium up
- Privacy Mode on by default for Teams and Enterprise
- Cursor Router is Teams and Enterprise only
Cognition (Devin) Autonomous SWE
Cognition runs Devin, marketed as the first autonomous software engineer. It plans, writes, tests and ships code inside your codebase and your tools. September was busy. A Series E on September 8 says Cognition raised over $2B at a $48B valuation, led by Andreessen Horowitz and Accel. SWE-2, which Cognition calls its most advanced coding model yet, landed September 10. Fusion shipped into Devin Desktop and the Devin CLI on September 11. A multi-year Strategic Collaboration Agreement with AWS followed on September 15. And on September 25 Cognition said it crossed $1B in annualized revenue run rate. Vendor statements, all of them, and none audited. The price list binds self-serve buyers only. Enterprise is billed in Agent Compute Units, and no ACU rate is published anywhere. Training is on by default. Paid plans can opt out. Somebody has to go and click it. And on FedRAMP, the surfaces split: Devin Desktop and the CLI are authorized, the full platform isn’t yet.
· The Series E wording: “Cognition has raised over $2B at a $48B valuation.”
· The training wording, from the undated Devin documentation: “By default, we may use your data for model training purposes.”
- $1B annualized revenue run rate claimed September 25, 2026 (vendor statement)
- Self-serve list published; Enterprise ACU rate unpublished
- Training on by default; paid plans can opt out, Enterprise needs written consent
- Devin Desktop and CLI FedRAMP High Authorized; full platform In-Process
| Coding vendor detail | Cursor | Cognition (Devin) |
|---|---|---|
| Owner and seller | Owned by SpaceX since August 14, 2026; subscriptions sold by X Subscription Corp since October 12, 2026 | Independent; Series E on September 8, 2026 at a stated $48B valuation |
| Latest model news | Grok 4.7, September 21, 2026; priced by xAI from $2 input and $6 output per 1M tokens | SWE-2, September 10, 2026; Ultra card names Fable 5.1 and GPT-6 Astra, September 23, 2026 |
| Self-serve price | Free $0; Plus $8; Premium $30; Super $100; Ultra $200 a month; Teams $30, $100 or $200 per seat | Free; Pro $20 a month; Max $200 a month; Teams $80 a month |
| Models by tier | Free and Plus: Cursor’s own models only. Premium and up: third-party models at API list price from prepaid credit | GPT usage can bill to a linked ChatGPT Go, Plus or Pro plan |
| Enterprise price | Custom; third-party requests on Teams and Enterprise add $0.25 per 1M tokens | Agent Compute Units; no ACU rate published |
| Training default | No training with Privacy Mode on; on by default for Teams and Enterprise; individual default not stated | May train by default; paid plans can opt out; Enterprise only with express prior written consent |
| Retention published | No Cursor period; most model providers under zero data retention, Fable 5 and 5.1 excepted | Duration of the customer relationship |
| Certifications | SOC 2 Type II, ISO 27001, ISO 42001, AIUC-1 | SOC 2 Type II since September 2024, per undated docs; FedRAMP High Authorized for Devin Desktop and CLI |
| Status flags | Projects in beta since September 10, 2026; switching for existing subscribers rolling out since October 12, 2026 | Full platform and Devin Cloud FedRAMP High In-Process since July 13, 2026 |
| Open questions | Privacy Mode default on individual plans; data terms across a combined Grok and Cursor account; Cursor Start | Enterprise ACU rate; Teams member cap, unlimited on devin.ai/pricing and 200 in the docs |
| Price source | cursor.com/pricing and Cursor help, undated, read October 15, 2026 | devin.ai/pricing and docs.devin.ai, undated, read October 15, 2026 |
| Deals since July 2026 | Firetiger, August 13, 2026 | TierZero; The Interaction Company; the Dioxus team, September 10, 2026 |
Where domain data is the moat
Raw intelligence is the cheap part here. What matters in legal, finance and enterprise knowledge is grounded access to the right documents, under the right permissions, with the right citation and audit discipline. These tools compete on the corpus and the controls. The model underneath is almost beside the point, right? And all three are shipping. Harvey raised at $15.5B on September 9 and shipped a connector library in October. Hebbia went headless on October 1, with an API and MCP server. Glean’s October 1 release notes say agent identity is generally available, while its docs still call it beta.
Harvey Legal
Agents for law firms and enterprise legal teams, grounded in case law, contracts and firm knowledge. Harvey is the most prolific publisher here. Its homepage, read October 15, 2026, claims more than 200,000 professionals, 3,500 or more firms and in-house teams, more than 85 countries and 80 or more AmLaw 100 firms (vendor claims). Those figures move often, so read the homepage again on the day you quote it. On September 9 Harvey raised $550M at a $15.5B valuation. It has held AIUC-1, an agent-security certification, since July 30. What shipped lately: a Connector Library, Agentic Search for SharePoint and Google Drive, a Motion to Dismiss Workflow with LexisNexis, and Fable 5.1 and GPT-6 Astra, all in its October 9 Brief. What hasn’t: Harvey II, announced August 18, still carries no availability language. On data, Harvey says it doesn’t train on your work, and now says so contractually. Customers can set retention policies, but no duration is published. And there’s no price anywhere.
· Harvey’s training wording opens: “Harvey contractually guarantees through our Platform Agreement that your data stays yours.”
- 3,500 or more firms and teams, 85 or more countries (vendor claims)
- AIUC-1 certified since July 30, 2026
- No training on inputs, outputs or uploaded documents
- Retention set by policy, with no duration published
Hebbia Finance
Hebbia calls itself the leading AI platform for finance. Its product page leads with two products, Matrix and Max. Matrix runs analysis across private documents, public filings and licensed market data. Connectors include SEC filings, CapIQ, FactSet, PitchBook, Preqin, expert networks, Snowflake and SharePoint. October brought three posts. Headless Hebbia on October 1, an API and MCP server. The October Disclosure on October 5, with Max Memory, Version History, Agent Creation, Search and Scheduling and an MS Word plugin. And an Octus partnership on October 13. Max is the puzzle. Hebbia keeps shipping features for it, yet its July 30 post still says it’s “rolling out to a small set of firms first,” as of October 15, 2026. On price, hebbia.com/pricing is broken rather than empty. On data, Hebbia gives the cleanest single sentence here, and its DPA adds what happens at the end.
· Hebbia’s training wording: “Your documents, prompts, and outputs are never used to train Hebbia’s models or any model provider we work with.”
· The DPA, last updated April 27, 2026, says Hebbia will destroy or return data on termination, at your election, “without undue delay.” No day count.
- Headless Hebbia API and MCP server, October 1, 2026
- Documents, prompts and outputs never used for training
- Pricing page returns a Contentful error, not a price
- Max ships features with no general-availability statement
Glean Enterprise knowledge
Work AI that unifies permission-aware search, an assistant and agents across your company apps. Start with what’s now readable. glean.com/security lists SOC 2 Type II, ISO 27001, ISO 42001, HIPAA, TX-RAMP Level 2 and GDPR. Its legal index lists a Business Associate Agreement too. Glean publishes no price. Plan names and a currency, FlexCredits, show up only in release notes. Glean’s May 28, 2026 press release puts the business at $300M ARR (vendor claim). Transform, launched August 26, is still “Coming soon” on Glean’s October 2026 Product Drop. September brought interactive artifacts, shipped the same day. On September 29 Glean joined OpenAI’s marketplace. And its release notes are good news for you. Each release has its own dated page, back to April 2025.
· Transform, in Glean’s words: “an early preview of what we’re building.”
· Glean gives Transform no ship date.
- Lists SOC 2 Type II, ISO 27001, ISO 42001 and HIPAA
- Dated per-release notes you can cite, back to April 2025
- No published prices or seat model
- Glean Transform is “Coming soon,” not available
| Vendor detail | Harvey | Hebbia | Glean |
|---|---|---|---|
| Latest shipped feature | Connector Library and Agentic Search for SharePoint and Google Drive, October 2026 Brief | Headless Hebbia API and MCP server, October 1, 2026 | Platform Skills API and three connectors generally available, October 1, 2026 |
| Announced, not available | Harvey II, August 18, 2026; Everlaw MCP integration, September 3, 2026, targeted for fall 2026 | Max, gated since July 30, 2026 | Transform, August 26, 2026 |
| Status in dispute | None found | MCP launch dated August 31 in one post, October 1 in another | Agent identity: generally available per October 1 release note; beta per docs updated October 12 |
| Published price | None; harvey.ai/pricing is a 404 | None; the pricing page returns a Contentful error | None; plan names and FlexCredits appear only in release notes |
| Training position | No training on inputs, outputs or uploaded documents; backed by the Platform Agreement | Documents, prompts and outputs never used for training | Not stated on Glean’s security page |
| Retention published | No duration; customers set retention policies; audit logs kept one to ten years | Service data kept for the business relationship; destroy or return on termination, no day count | No |
| Certifications listed | AIUC-1, SOC 2 Type 2, ISO 27001, 27701 and 42001, IRAP, TX-RAMP; trust center lists a HIPAA Report and BAA | SOC 2 Type 2, ISO 27001:2022, ISO 42001:2023, GDPR, CCPA; no HIPAA | SOC 2 Type II, ISO 27001, ISO 42001, HIPAA, TX-RAMP Level 2, GDPR |
| Subprocessor list | Updated September 21, 2026, with a processing location per provider | DPA points to trust.hebbia.ai, which renders no names; 10 days’ notice of new ones | Updated July 28, 2026, 21 named entities |
Agents with an identity of their own, and two answers on status
An agent can “use its own scoped service credentials instead of borrowing the identity of the person who invoked it,” with admin controls for credential scoping and audience management. So access rights split from invocation rights, and audit trails name the agent rather than whoever triggered it. It’s the most interesting governance move in this market. Now the status. Glean’s dated October 1 release note says “Agent Identity is now generally available” and “Available by default.” Glean’s admin overview, updated October 12, still opens with “Beta: This feature is in beta and may change.” Two Glean pages, opposite answers, and neither can be called the sole current truth until Glean reconciles them. So don’t pick one. Get the status, and the support terms that come with it, in writing before you put it in a dated rollout plan.
· The service-credentials builder page, last updated September 25, 2026, also carries the beta banner.
· The admin page says the agent runtime never receives the raw secret.
- Public beta from July 15, 2026
- October 1, 2026 release note: generally available
- Admin docs updated October 12, 2026: still beta
- Audit trails name the agent
A vendor benchmark against a named competitor
One of the six August 26 posts names a rival in its title: “Right-sizing intelligence: Glean saves 81% on token costs and is preferred 78% of the time versus Claude Cowork.” Both figures are Glean’s own claims about a named competitor. Produced by Glean. Published by Glean. In Glean’s comparative marketing. Record that the post exists and what it asserts. Don’t endorse either number. If a token-cost or preference figure is going to move a purchase, run the comparison on your own corpus and your own queries. A vendor’s idea of “right-sizing” is a claim about which jobs need which model. That judgment is exactly what you’re buying.
- 81% token-cost saving: Glean’s claim
- 78% preference rate: Glean’s claim
- Vendor comparative marketing, not an independent test
- Reproduce it on your own workload before citing it
A specialist arriving inside a suite
Harvey showed up as a new connector in Microsoft 365 Copilot in Microsoft’s August 2026 roundup. The specialist and the generalist have stopped being alternatives on a shortlist. They’re layers now. So ask three questions. Who owns the connector? Whose permission model applies inside it? Where does the audit trail land? For a law firm, Ethical Wall enforcement with Intapp, which Harvey announced as generally available on July 23, is why that matters. Conflict policy has to travel with the tool into every surface it appears on. The September 3 Everlaw announcement is the same pattern again. Ask those questions twice there, because Harvey and Everlaw say they “are building” the MCP integration and expect it in fall 2026. That’s not live.
- Harvey is now a Microsoft 365 Copilot connector
- The Everlaw MCP integration is targeted, not shipped
- Two permission models meet in one workflow
- Confirm walls and audit coverage in the connected surface
Beyond your own documents
The corpus is no longer just the data room. Hebbia’s July 10 integration post is the clearest statement of that scope: private files sit alongside public filings and licensed market data in a single view. So the diligence question changes. It used to be “can it read our documents.” Now it’s “which third-party licenses do we already hold, and which does this tool require of us.” The October 13 Octus partnership makes it concrete, because it needs an Octus subscription. And Headless Hebbia adds a second version of the question, because output now leaves through an API and MCP server into Claude and ChatGPT. That’s an egress question.
· Headless Hebbia, October 1, 2026, names no availability tier and no price.
- One view over three kinds of data
- License terms become part of the buy
- API and MCP distribution add an egress question
- Source every number back to a filing
For this vertical tool, tell me: which underlying model it uses, exactly what proprietary data or workflow it adds on top, how it handles citations and source traceability, and where our data is stored and processed. Separate vendor marketing from verifiable fact.
Compare what this specialist does against what our existing Microsoft 365 Copilot and ChatGPT already cover. Identify the specific 20% of the workflow the specialist really adds, and whether that 20% justifies the cost and integration.
This vendor published a comparison claiming a cost saving and a preference rate against a named competitor. Design a test I can run on our own documents and queries that would confirm or refute both numbers, and tell me what the vendor’s methodology would have to look like for the claim to hold.
Specialists that act, and specialists that must be right
Two frontiers where being wrong costs something real. And on one of them, a vendor’s privacy policy and its own HIPAA announcement say opposite things about the same clinician asking the same question. Customer-facing agents take actions for actual users. Clinical tools carry a different weight of error entirely. Both show the same thing: the guardrails and the paperwork matter more than any capability claim.
Sierra Agents that act
Sierra began in customer experience, across chat, voice, email and messaging, priced on outcomes. It has pushed well past that. Horizon, announced July 16, targets goals that run for days, weeks or months, such as loan origination and prior authorization. September added multimodal agents, AIUC-1 certification on the 17th and Ghostwriter. Then Sierra Summit on October 6. Sierra named its own fine-tuned models: Curie runs the core orchestration loop, and Fleming detects when the caller is another agent. It also announced the Personal Agent Protocol, an open standard it says Meta and Sierra are developing, with a v0.1 spec planned for later in October 2026. On price, say it plainly: there’s no rate card. No per-resolution rate, no platform fee, no minimum, no billing term. What’s new is the number of ways to pay: the OpenAI Marketplace, unnamed cloud and frontier lab marketplaces, and the Stripe App Marketplace. On data, Sierra publishes one good sentence and nothing more.
· Sierra’s data wording: “Your data is only used as you instruct, and is never shared with other customers.”
· Sierra’s slogan, from its Summit recap: “Rent the intelligence and own the context.”
- In-house models Curie and Fleming, October 6, 2026
- No rate card on any Sierra page
- Sold via OpenAI, unnamed cloud marketplaces and Stripe
- Data used only as instructed, never shared between customers
Abridge Clinical documentation
Abridge turns doctor and patient conversations into structured clinical notes in real time, inside Epic. Its product page now reads “One platform, three tailored experiences”: Clinicians, Revenue Cycle and Nursing, with Assistant and Care Signals as capabilities. It adds Real-time Prior Authorization, with Highmark Health and AHN, and Abridge Inside Epic. Abridge credits its own Contextual Reasoning Engine, with Linked Evidence for auditability. On October 8 the University of Texas System signed a systemwide agreement across eight named institutions. The VA item from September 22 still needs careful reading. Abridge was selected under the VA enterprise contract “through our distribution partner,” who isn’t named. The $775.72 million ceiling is shared across all eligible vendors. It isn’t Abridge’s award, and no task-order value is disclosed.
· Abridge states no training or retention position on any page a buyer can read without asking.
· Abridge publishes no price anywhere.
- Three experiences: Clinicians, Revenue Cycle, Nursing
- UT System agreement, October 8, 2026
- VA selection came through an unnamed distribution partner
- Clinician reviews every note
OpenEvidence Clinical evidence
A medical answer engine grounded in licensed clinical literature, free to clinicians and deployed across health systems. Start with the contradiction, because it decides whether you can use this at all. The privacy policy says user questions and prompts may be sold or monetized as a commercial data feed, and that user inputs aren’t covered by HIPAA. The security page repeats the HIPAA line. The vendor’s April 25, 2025 announcement says it’s fully HIPAA compliant, and its Terms say covered-entity PHI is handled under a business associate agreement “incorporated into the Terms by reference.” The Terms don’t link that agreement. All four are first-party. All four are live. So ask for the BAA itself, and treat the privacy policy as governing until OpenEvidence confirms otherwise in writing. The content deals are dated: NEJM Group, February 19, 2025, and JAMA Network, June 5, 2025, though OpenEvidence’s announcements index lists an August 21, 2025 item on both agreements. But don’t call the July 15, 2025 round its latest news, because OpenEvidence serves announcement pages that sit on no index.
· The policy wording: “Your user data, including your questions/prompts, may be sold to third parties or otherwise monetized as a commercial data feed.”
- Privacy policy: questions and prompts may be sold as a data feed
- Policy and security page: inputs not covered by HIPAA
- Announcement and Terms: covered-entity PHI under a BAA
- The BAA is referenced in the Terms but not linked
| Vendor detail | Sierra | Abridge | OpenEvidence |
|---|---|---|---|
| Latest dated item | Summit 2026 and Personal Agent Protocol, October 6, 2026; draft shared October 9 | UT System agreement, October 8, 2026 | Latest on the index: JAMA and NEJM content agreements, August 2025; pages off the index can’t be dated |
| Published price | None; outcome pricing described in prose only | None | Free; no pricing page, seat model or enterprise plan |
| Training position | Data used only as instructed, never shared with other customers | Not stated on any readable page | Questions and prompts may be sold or monetized as a data feed; Terms allow selling usage data to licensed third parties |
| Retention published | No | Not stated on any readable page | Up to one year after account deactivation |
| HIPAA | Listed on the trust page | Listed in the trust center | Policy and security page: inputs not covered. Announcement and Terms: BAA for covered entities |
| Certifications listed | SOC 2, HIPAA, GDPR, PCI DSS Level 1, FedRAMP High, CCPA, CSA STAR, ISO 27001, ISO 42001; AIUC-1 announced September 17, 2026 but not listed | CCPA, HIPAA, SOC 2 Type 1 and Type 2, TX-RAMP | None named; the security page states AES-256 storage |
| Gated or missing documents | Trust center shows no document detail; AIUC-1 scope and issue date not stated | Trust center offers access requests without saying which documents are gated | Business associate agreement referenced, not linked |
A decision framework, not a vibe
These calls are expensive and they keep coming back. Specialist or generalist. Buy or build. Run them on purpose, then run them again, because the thing you evaluated in July may arrive through a different surface in September, and from a different seller in October.
Six habits for buying a specialist without regret
None of these habits is about the model. They’re about what you read, what you write down and who owns the result. Each one comes from a vendor above that already caught somebody out.
Read the live page before the launch post
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. And when a vendor’s own dated note and its own live docs disagree, as Glean’s do on agent identity, don’t pick a side. Get it in writing.
Treat announced as unavailable
Harvey II, Hebbia Max, Glean Transform and the Everlaw integration are announcements, not products. Diarize the date, re-read the source page on the same cadence, and keep a dated copy of the row you depend on. A date written down once and trusted is how you miss a deadline that moved.
Get the price in writing
Here a written quote is usually the only price that exists. Six of the eight vendors publish nothing usable. Cursor and Cognition publish lists, and both stop at Enterprise. Cursor’s list now comes from X Subscription Corp. And tell a broken pricing page from a deliberate one before you draw a conclusion.
Give every training switch an owner
Cursor turns Privacy Mode on by default for Teams and Enterprise and doesn’t say what individual plans get. Cognition may train by default until a paid plan opts out. So a default like that is a configuration task, not a feature, right? Name the person who sets it, and have them check it again after every plan change, renewal or switch to a new subscription.
Ask which surface it arrives on
Hebbia arrives inside Claude and ChatGPT over MCP. Harvey arrives inside Microsoft 365 Copilot. Cursor now arrives inside a bundle with Grok and X. Each time, ask whose permission model, data terms and audit trail apply there. And ask whether the embedding vendor’s model terms match the ones the model maker publishes.
Keep a human of record
Every filing, diff, clinical note and customer action needs a named professional who signs it. Every long-running goal needs an owner and a checkpoint schedule. “The agent handled it” isn’t a sign-off. A name is.
A $48B valuation is not a fit for your workflow
Three risks here that a general assistant doesn’t carry. You often pay a premium for someone else’s model. You embed sensitive data deeply. And you bet on a market that’s consolidating while you read. At least eight deals have landed across these eight vendors since July 10, 2026. Cursor acquired Firetiger on August 13. SpaceX acquired Cursor the next day, and on October 12 Cursor’s plans moved into xAI’s subscription. Deals come in bursts, so don’t mistake one busy stretch for a yearly pace, or a quiet one for a settled market.
Rented intelligence
Many verticals wrap OpenAI or Anthropic. Some now build their own, like Sierra’s Curie and Fleming. Either way, confirm the domain layer is real before you pay specialist prices for a model you could reach directly.
Overlap you already own
Your Copilot, ChatGPT or Claude may already cover most of the job. Buy the specialist for the measured gap only. And this is getting harder to judge, not easier. Hebbia runs inside Claude and ChatGPT over MCP. Harvey is a Microsoft 365 Copilot connector. So the same capability can reach one user through two contracts at once. Nobody wants to pay twice, right?
Vendor comparisons name competitors now
Glean’s August 26 post claims it “saves 81% on token costs and is preferred 78% of the time versus Claude Cowork.” Those are Glean’s figures about a named rival, published in Glean’s own marketing. Treat any head-to-head number from a vendor as a hypothesis to test on your corpus. Stay just as skeptical when a competitor answers with numbers of its own.
Data and compliance
Vertical tools touch your most sensitive data: legal, clinical, financial. Three of the eight vendors here say plainly that they don’t train on your work. Cognition defaults the other way, with an opt-out you have to go and find, and OpenEvidence’s policy says prompts may be sold. Cursor protects Teams and Enterprise by default and doesn’t say what individual plans get. Two publish no answer a buyer can read. And three publish anything like a retention period. The table below sets out each position in the vendor’s own terms. Add one more question for MCP distribution and bundles. When the specialist runs inside another vendor’s product, whose data-processing terms apply?
The terms someone has to read in person
Glean’s data terms sit on pages a person has to open and read. Its legal index links a DPA, AI addendum, security standard, Limited Retention Addendum, terms, BAA and two dated PDFs, and trust.glean.com shows little beyond its headings. Those pages are where the answer sits on whether Glean trains on your data, how long it keeps it and when it deletes it. Sierra’s trust center and Hebbia’s subprocessor page are the same. Harvey gates its HIPAA Report and BAA. Abridge offers documents on request. And cursor.com/unify, which may explain the combined Grok and Cursor account, needs the same reading. So put this on a person’s calendar, not in a backlog. It’s an hour of someone’s time and it decides whether the contract is signable.
Lock-in
Proprietary data and workflow make these tools hard to leave. Check export and migration before you embed one in a critical path. And check whether a switch is reversible. Cursor says moving to The X Subscription is permanent.
Announced is not available
Four things on this map have been announced without becoming buyable, and the wait is getting hard to defend. Harvey II was introduced August 18, 2026 with no availability language of any kind and no pricing, and the October Brief doesn’t mention it. Hebbia Max has been “rolling out to a small set of firms first” since July 30, and the October Disclosure ships five Max features with no general-availability statement. Glean Transform launched August 26 and is still “Coming soon” on Glean’s October Product Drop. And Harvey and Everlaw said on September 3 that they “are building” an MCP integration. As of October 15, 2026 the wording is the same, with a fall 2026 target. Two more need care. Glean agent identity is generally available in Glean’s October 1 release note and beta in its October 12 docs. Cognition’s full platform has been FedRAMP High In-Process since July 13, 2026. None of these belongs in a dated rollout plan. If one of them is load-bearing in yours, ask for a date in writing and read “soon” as a no.
A capability matrix expires in days
Fable 5.1 reached Harvey and Microsoft 365 Copilot on its September 1 launch day, Glean on September 3 and Cognition’s Fusion harness by September 11. Then Glean added Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna on September 22, Claude Sonnet 5.5 on September 29 and GPT-6.1 Sol on October 6. OpenAI calls GPT-6.1 Sol an upgrade to GPT-6 Sol, and GPT-6 Sol still powers ChatGPT’s paid tiers since October 7. The vendors don’t agree on the roster either. Harvey and Cognition name GPT-6 Astra. Cursor’s model list shows GPT-5.6 Sol, Terra and Luna, with no GPT-6 row visible. So any capability matrix you circulate should carry the date it was compiled, the page each row came from and the tier each model needs. Two weeks is long enough to make it wrong in either direction.
Consolidation, at speed
At least eight deals across these eight vendors since July 10, 2026. In the seventeen days to July 27: Cognition bought TierZero and welcomed The Interaction Company, Harvey bought Benchmark, Sierra bought Takeoff, Abridge took on the Altrina team. On August 13, Cursor acquired Firetiger, which builds agents that work on software once it reaches production. No terms were disclosed. On August 14, Cursor announced that it “has officially been acquired by SpaceX,” completing a process begun in April. Again no valuation and no terms. On September 10, the Dioxus team joined Cognition, no terms. Money moved too. Harvey took a strategic investment from Goldman Sachs and J.P. Morgan on July 28, then raised $550M at a $15.5B valuation on September 9. Cognition announced a Series E on September 8 it says raised over $2B at a $48B valuation. One of these eight vendors is now owned by a launch company and sold through that group’s subscription. The specialist you sign with may not merely be a different company in a year. It may sit inside a group whose priorities have nothing to do with your codebase.
Do not infer ownership, but do not ignore the signal
A partnership, a joint model launch or a shared brand is evidence of a relationship. It isn’t evidence of ownership. Cursor ran a public partnership with xAI from April 2026, and when xAI released Grok 4.6 on August 12, Cursor published its own launch post the same day. Cursor announced the acquisition itself only on August 14, 2026, in a dated post titled “Cursor is now a part of SpaceX.” Wait for the dated post rather than inferring from the pattern. Then look at what followed. Cursor now calls the Grok models its own. Its pricing page names X Subscription Corp as the only seller. Its enterprise page still says nothing about the change of ownership. SpaceX acquired xAI on February 2, 2026 and Cursor on August 14, 2026, so ownership, model supply and now billing sit inside one group. There’s one parent instead of two counterparties. The terms between them are intra-group rather than negotiated at arm’s length. That puts affiliate terms, transfer pricing and a single change-of-control event on the agenda in place of ordinary third-party supply risk. Get ownership, control, model supply and the contracting entity written into the contract rather than inferred from a launch post.
A missing price is not a discontinued product
Cursor Start, the ₹649 a month India tier announced July 28, still isn’t on Cursor’s pricing page as of October 15, 2026. The post is still live with no update note. No page says whether Start was withdrawn, geo-gated or folded into The X Subscription. So leave it unconfirmed rather than guessing. Hebbia is a different case and shouldn’t be filed with Cursor. hebbia.com/pricing returns “Failed to load page data from Contentful.” and no price at all. That page is broken, not minimal, and you can’t tell whether a price was ever published there. Now look at the field. Harvey, Sierra, Glean, Hebbia, Abridge and OpenEvidence publish no usable price between them. Cursor and Cognition are the only vendors here with published lists. Three habits follow. Don’t diagnose a vendor’s page from what you expected to see on it. Tell a broken page apart from a deliberate one. And get any tier you intend to buy confirmed in a written quote, because on this map a written quote is usually the only price that exists.
Valuation is not fit
A famous, well-funded tool can still be wrong for your one workflow. Evaluate on your job, not on the funding round.
| Vendor | Published position on training, in its own terms | Retention period published |
|---|---|---|
| Harvey | No use of inputs, outputs or uploaded documents to train underlying models; contractually guaranteed through the Platform Agreement | No duration. Customers set retention policies; logs 1 to 10 years |
| Hebbia | Documents, prompts and outputs never used to train Hebbia or any model provider it works with | Yes. Service data kept for the business relationship |
| Sierra | Data used only as instructed, never shared with other customers | No |
| Cursor | No training when Privacy Mode is on; on by default for Teams and Enterprise; individual default not stated | No |
| Cognition | May train by default; paid plans can opt out at any time; Enterprise requires express prior written consent | Yes. Duration of the customer relationship |
| OpenEvidence | Questions and prompts may be sold to third parties or otherwise monetized as a commercial data feed | Yes. Up to one year after account deactivation |
| Glean | Not stated on Glean’s security page | Not published |
| Abridge | Not stated on any readable page | Not stated on any readable page |
| Pre-purchase check | What to check | Who owns it |
|---|---|---|
| Frontier model underneath | Which model, which tier unlocks it, and how fast the vendor adopts the next one | Technical evaluation |
| Sensitive data | Where data is stored and processed; compliance certifications | Security and legal |
| Overlap with existing tools | Gap versus the current general assistant | Budget owner |
| Delivery through another vendor | Whose permission model, terms and audit trail apply in that surface | Security and procurement |
| Bundle sale | Invoicing entity, and whose data terms cover the combined account | Procurement and legal |
| Deep embedding | Exit and data-export path | Procurement |
| Work of record | Human review of every filing, diff, note or customer action | The professional of record |
| Acquisition risk | Change-of-control terms, and what happens to price, roadmap and support | Procurement and the budget owner |
Where these facts come from
These vendor pages change fast, and the category consolidates fast, so verify before you act on any specific tool. Capability claims are anchored to each vendor’s own dated posts. Valuations, funding, customer counts, benchmark scores and comparative figures are as reported by the vendor and labeled that way. Now the limits, which matter more than the list. Several facts here live only on rolling pages that carry no date and can change without notice. Cursor’s tier prices, its model gating and its Privacy Mode defaults come from its pricing, help and docs pages, read October 15, 2026. Remote control for local agents and the Origin App Marketplace appear only in Cursor’s changelog. The note that existing subscribers keep included third-party usage until they switch comes from Cursor’s October 12, 2026 forum post, which isn’t cited below. Glean’s GPT-6.1 Sol, Haiku 5.5 and FlexCredits entries sit only on its release-notes index. Devin’s ZDR and memory notes come from its rolling release notes. Harvey’s customer counts come from its homepage, read October 15, 2026. Trust-center listings at trust.harvey.ai, trust.abridge.com and trust.sierra.ai are rolling too, and aren’t cited below. Confirm any of these directly with the vendor before you rely on them. Pages only a person with a browser can open: Glean’s DPA, AI addendum, security standard, Limited Retention Addendum, terms and BAA, plus cursor.com/unify, all robots-blocked; trust.glean.com, trust.sierra.ai and Hebbia’s subprocessor page, which render metadata only. The 404 at harvey.ai/pricing and the Contentful error at hebbia.com/pricing were read October 15, 2026. sierra.ai/pricing, cognition.com/pricing, abridge.com/security and glean.com/products/transform failed to load that day, which isn’t the same as a confirmed 404. Other items referenced in the body were read on the vendors’ own channels without being separately cited: Harvey as a Microsoft 365 Copilot connector in Microsoft’s August roundup, the JAMA Network agreement of June 5, 2025, and the Ethical Wall launch of July 23, 2026.
AI Mindset