The Fast, Real-Time Operator With an Edge
Grok 4.7 is the flagship, out September 21, 2026, and it costs exactly what 4.6 did. $2.00 per 1M input, $0.50 cached and $6.00 output below 200,000 tokens. Still a 500k context window. What is new is a reasoning effort dial with four settings, high by default. And the flagship can’t use the Batch API at all, so batch pricing is out of reach on the model you’ll actually call. A fast variant is announced at twice the price and still has no ID. Underneath all of it, xAI rewrites undated pages and tells nobody. Three of its own pages give three different knowledge cutoffs for 4.7. So the risk here isn’t the model. It’s your own notes.
The colleague closest to the live conversation
Some questions change by the hour. That’s Grok’s window. And the edge isn’t the cheeky tone everyone talks about. It’s live X context, frontier reasoning, coding agents and media production sitting in one place.
Direct, fast and current
Grok will take a position. That helps when you’re trying to make sense of a live event. It hurts when the confidence runs ahead of the evidence.
- Real-time social context
- Strong engineering orientation
- Less polished corporate default tone
Live signal and high-velocity execution
Breaking narratives. Market reaction. Technical work, rapid prototyping, coding agents, and image and video experiments in the same window.
- What’s happening now
- What developers are saying now
- What can be built or rendered quickly
Recency or candor changes the answer
Reach for Grok when yesterday’s summary isn’t enough. Or when the social reaction is part of the evidence. Or when a blunt stress test makes the work better.
- Live pulse
- Direct critique
- Fast action through Build or Imagine
Map the three dominant narratives about this announcement on X. Show who is driving each one, what evidence they cite and where the narratives disagree. Don’t decide which is true yet.
Now verify the factual core against the company release, filing and any other primary source. Separate confirmed facts, disputed interpretation and unsupported social claims.
A new flagship at the old price, and a discount you can no longer reach
Grok 4.7 shipped on September 21, 2026 at the same price and the same 500k window as Grok 4.6. The new control is reasoning effort: low, medium, high and xhigh, with high as the default. The new limit is a subtraction: it can’t use the Batch API at all. Rate limits are 150 requests per second and 50,000,000 tokens per minute. Distribution kept moving too. Grok 4.6 reached GitHub Copilot on August 14, Amazon Bedrock on August 19, the Gemini Enterprise Agent Platform on August 21 and Microsoft Foundry on August 26, and Grok was already on Bedrock and Databricks in June. A sixth route belongs to Microsoft, not to xAI. Since September 12, 2026 Grok has sat inside Word, Excel and PowerPoint through the Microsoft Frontier Program, off by default and closed to the EU, EFTA and the UK during the preview. Grok Bot gained enterprise access, network and audit controls on September 3, and Team Bots arrived in public beta on September 28. And underneath all of it, three xAI pages give three different knowledge cutoffs for one model. No xAI page mentions the Frontier Program route. It rests on Microsoft’s own Copilot Blog post. And it has nothing to do with the xAI Frontier AI Framework, a safety policy dated June 30, 2026.
grok-4.7: none, so batch pricing never applies to the flagshipgrok-imagine-image-quality; requests then move to grok-imagine-image-2.0 at low quality, $0.01 per image cheaper at every resolutionGrok 4.7 Flagship · September 21
Out on September 21, 2026, and priced exactly where 4.6 was. $2.00 per 1M input, $0.50 cached input and $6.00 output below 200,000 tokens, with a 500k context window. Cross 200,000 tokens and every token in the request bills at the long-context rate. The real change is a dial. Reasoning effort now takes low, medium, high or xhigh, and it defaults to high, so a request you don’t configure is a request on the expensive setting. The other change is a subtraction: the Batch API doesn’t support this model. Three xAI pages describe it three different ways, so name the page whenever you quote one.
- Pin
grok-4.7rather than a moving alias - Reasoning effort: low, medium, high, xhigh. Default is high
- No Batch API support, so batch pricing doesn’t reach the flagship
- Rate limits of 150 requests per second and 50,000,000 tokens per minute, region us-east-1
- Text and image in, 500k context, long-context tier at $4 / $1 / $12
- xAI positions it as “Twice as fast, at half the price of comparable models,” which is a vendor claim
- Model card dated September 21, 2026, revision 2026-09-21
Grok 4.6 Demoted, not retired
Grok 4.6 is no longer the flagship, and nobody has told you what it is instead. The dedicated model page still renders with no deprecation marking and no legacy label. The price table still carries a grok-4.6 row. But x.ai/api has dropped it from the pricing tables and the docs models index no longer describes it at all. No retirement date appears on any xAI page. So don’t read the marketing page as a shutdown notice, and don’t read the surviving row as a commitment. Ask xAI in writing how long that row stays.
- Still priced at $2 / $0.50 / $6 below 200,000 tokens, 500k context
- No retirement date published on any xAI surface
- Dropped from the
x.ai/apipricing tables and from the docs models index - Grok 4.5 is a generation further back, still carries a price-table row, and still has its own model card dated July 14, 2026
- If a workload depends on 4.6 output behavior, start the 4.7 comparison now rather than on the day the row goes
Four surfaces in twelve days Distribution
Grok 4.6 landed on GitHub Copilot on August 14, 2026. Then Amazon Bedrock on August 19, alongside Grok Build on web and mobile. Then the Gemini Enterprise Agent Platform on August 21 and Microsoft Foundry on August 26. Four third-party surfaces in twelve days, and the wave starts at Copilot, not at Bedrock. Count from Bedrock and you lose a surface and flip the order. Bedrock wasn’t new in August either. xAI published a Grok on Amazon Bedrock listing on June 17, 2026 and Grok on Databricks on June 18. So Databricks is a fifth catalog, it predates the August wave, and only Grok 4.6 on Bedrock was new. Either way the buying question changes. Plenty of organizations can now get Grok through a contract they already signed, without ever signing one with xAI.
- GitHub Copilot August 14, Bedrock August 19, Gemini Enterprise Agent Platform August 21, Microsoft Foundry August 26
- Grok reached Bedrock on June 17, 2026 and Databricks on June 18, 2026
- Copilot access may need enabling in Copilot settings on business and enterprise tenants, as xAI noted at launch
- Check which entity holds the data terms on each platform
Grok in Word, Excel and PowerPoint through the Microsoft Frontier Program September 12 · EU, EFTA and UK excluded
The sixth surface is the one xAI didn’t announce. Microsoft put Grok into Word, Excel and PowerPoint through its Frontier Program. Not Cowork. Not Copilot Studio. Access sits behind a dedicated admin setting that’s off by default. And Grok isn’t available to Frontier customers in the European Union, the European Free Trade Association and the United Kingdom during the preview. No end date is given. The gate, the regions and the data terms on this route belong to Microsoft, not to xAI. One name trap before you search for it. The only “Frontier” document on x.ai is the xAI Frontier AI Framework, a safety policy dated June 30, 2026. Different thing, different company, no connection to Word.
- Word, Excel and PowerPoint only: not Cowork and not Copilot Studio
- A dedicated administrative setting, disabled by default, controls access
- EU, EFTA and UK customers are excluded during the preview, with no published end date
- Not the xAI Frontier AI Framework, which is xAI’s own safety policy dated June 30, 2026 and has nothing to do with Microsoft
- A different mechanism from xAI’s own Office add-ins: see the Workplace Add-Ins section
- Source: Microsoft’s own post “Expanding model choice in Copilot with Grok” on the Microsoft Copilot Blog, published September 12, 2026 and updated September 14. No xAI page mentions this route
Three first-party answers on one cutoff Read the source, not the number
Grok 4.7’s knowledge cutoff depends on which xAI page you happen to open. The docs model reference says May 2026. The model card says a pretraining cutoff of June 2026, plus data generated as late as August 2026 in supplemental training. The dedicated model page says nothing at all. May 2026 is the one written as a plain date on the page everybody lands on first, which makes it the easiest to lift and the easiest to treat as settled. It isn’t settled. Quote a cutoff with the page it came from, and don’t hang a recency assumption on it.
- Docs model reference: May 2026
- Model card, revision 2026-09-21: pretraining to June 2026, supplemental data to August 2026
- Dedicated 4.7 model page at
docs.x.ai/developers/models/grok-4.7: no cutoff stated - The same three-way split showed up on the previous flagship, so expect it on the next one
- The docs pages are rewritten in place, so confirm directly
An announced price with no model to buy Contradiction, twice running
xAI has now announced a fast variant at twice the price twice, and shipped an ID for it neither time. The August 12, 2026 Grok 4.6 post said it. The September 21, 2026 Grok 4.7 post says it again: a fast variant with twice the output speed at twice the price. No model ID with the word fast in it appears anywhere on the docs price table, and neither dedicated model page names a variant. You can’t put a budget line against a model you can’t name. And a pattern that repeats across two releases is a publishing habit, not an oversight.
- Announced for 4.6 on August 12, 2026 and again for 4.7 on September 21, 2026, both at twice the standard rate
- The eight price-table rows are
grok-4.7,grok-4.6,grok-build-0.1,grok-4.5,grok-4.3,grok-4.20-multi-agent-0309,grok-4.20-0309-reasoningandgrok-4.20-0309-non-reasoning - No model ID, context window or pricing row on any surface
- Don’t commit a fast-variant budget line until an ID exists
- Price table read on a docs page rewritten in place: confirm directly
The Cursor connection Strategic · same owner
Cursor and xAI are now sister companies, and that changes how you read everything they say about each other. The sequence matters more than any one date. They announced a partnership in April 2026. On July 8, 2026 they shipped together. On August 14, 2026 Cursor said SpaceX had acquired it, finishing the process that started in April. No terms, no valuation. SpaceX had already acquired xAI on February 2, 2026. So the training relationship here’s inside one group, not arm’s length. No further terms have been published first-party. The closest thing to a follow-up is a September 22, 2026 xAI post about its own support desk, which mentions a ticket surge after the Cursor merger and nothing else about the deal. Record the rest as not found, not as nothing happened.
- Cursor has been SpaceX-owned since August 14, 2026, with no terms disclosed
- Partnership announced April 2026, joint Grok 4.5 release July 8, acquisition completed August 14
- July 8 wording: xAI said Grok 4.5 “was trained alongside Cursor,” on NVIDIA GB300s, and Cursor said “Today we are releasing Grok 4.5 together with SpaceXAI, our most intelligent model”
- August 14 wording: Cursor posted that it “has officially been acquired by SpaceX,” completing “the acquisition process that started in April, when we announced our partnership with SpaceXAI to accelerate our model training efforts”
- No further first-party detail found, which isn’t the same as none existing
- A buyer evaluating Grok or Cursor should know both products now share an owner
Grok Bot Eight plans · enterprise controls
Grok Bot looks generally available. xAI has never said it is. It launched August 11, 2026 in early beta for three tiers. By August 26 it covered eight plans. On August 29 it gained an X integration that xAI calls the first version. On September 3 it gained access, network and audit controls so enterprises can govern Bots at scale, and xAI claimed thousands of organizations had adopted it. Then on September 28, forty-eight days after launch, xAI shipped Team Bots and called that public beta in writing. So the beta language is still live. Widen your availability claim to the named plans and stop short of calling it GA.
- Eight named plans, up from three: SuperGrok, SuperGrok Plus, SuperGrok Heavy, Cursor Pro, Cursor Pro+, Cursor Ultra, and Cursor Teams Standard and Premium
- Access, network and audit controls for enterprises since September 3, 2026
- Team Bots, September 28, 2026, described by xAI as available in public beta
- Adoption and isolation statements are vendor claims
- No first-party statement that Grok Bot has left beta: don’t claim GA
Team Bots Public beta · September 28
A Bot the whole team shares, holding one set of context, plugins, credentials and memories. xAI says each person’s conversations with it stay private, and the Bot keeps separate context and memories per user while drawing on skills shared across the team. That is a design claim, not an audit. No price is published. And there’s a wrinkle a buyer hits in about ten seconds. xAI says Team Bots is available on Teams and Enterprise plans. There is no Teams tier on the pricing page. The seven columns are Free, SuperGrok Lite, SuperGrok, SuperGrok Plus, SuperGrok Heavy, Business and Enterprise. Ask which published tier the word Teams means before anyone budgets for it.
- Shared context, plugins, credentials and memories across a team
- Per-user privacy and per-user memory are vendor design claims, not tested findings
- No published price
- xAI names a Teams plan that doesn’t exist on
x.ai/pricing: get the tier named in writing - Still beta, so don’t write it into a dependency
Grok Build Coding agent
A terminal CLI agent, and on web and mobile since August 19, 2026. Plan mode lets you approve, annotate or rewrite the plan before anything runs. Bigger tasks get handed to parallel subagents in their own worktrees. And since September 16, 2026 it keeps memory between sessions as markdown files on disk. The open question is who is entitled to it. No x.ai page says which plans grant access. The pricing page never mentions X Premium or X Premium+, and X Premium pricing lives on x.com rather than on any xAI surface. So the widely repeated claim that X Premium Plus includes Grok Build has no first-party support. Confirm the plan before you promise it to anyone.
- Approve the plan before it runs
- Parallel subagents and worktree integration
- Persistent memory across sessions since September 16, 2026
- X Premium Plus entitlement is unconfirmed on any x.ai page
EU availability isn’t one answer Per surface
One date settled one surface, and people keep quoting it for all of them. Grok 4.5 wasn’t available in EU products or the API console at launch. That gap closed on July 17, 2026, when the developer release notes recorded Grok 4.5 as available to EU users in the API console. Read it for exactly what it covers. One model. xAI’s own API console. July. It isn’t the EU position everywhere, and at least one surface is shut today. Grok reached Word, Excel and PowerPoint through the Microsoft Frontier Program on September 12, 2026, and Frontier customers in the EU, EFTA and the UK are excluded during the preview, with no end date. Anyone who treats July 17 as settled and then switches Grok on in Word will be wrong about the region they’re standing in.
- July 17, 2026 covers Grok 4.5 on xAI’s API console, not every surface
- The Microsoft Frontier Program route excludes the EU, EFTA and the UK during the preview, with no end date
- Each third-party catalog carries its own regions and terms
- Establish EU access per surface and per model before promising it
- Source for the Frontier exclusion: Microsoft’s Copilot Blog post of September 12, 2026, not any xAI page
Skills and connectors May 18
Skills reached web, iOS and Android on May 18, 2026. Grok can build documents, decks and spreadsheets, and reach external systems through reusable skills and integrations.
- Persistent expertise
- Business workflow automation
- Permission-aware enterprise use
Office and Workspace add-ins
xAI ships its own free add-ins, and they put Grok inside PowerPoint, Word, Excel, Outlook, Sheets, Slides and Docs. That takes the work story well past X and grok.com. The program started in June 2026, not July. Keep this route apart from the Microsoft Frontier Program route dated September 12, 2026. They reach three of the same Office apps by different mechanisms, with different gates and different regional terms.
- Draft and revise documents
- Turn outlines into slides
- Not the same thing as Grok in Copilot through the Frontier Program
- Keep tone and source review explicit
- See the Workplace Add-Ins section for the dated releases and the comparison
A coding agent designed for speed and sustained work
Grok Build moves the product from advice into doing. The /goal command, the session dashboard and the plugin system make the unit of work an outcome, not a run of terminal prompts. Two dates change what a security review looks like. Since July 15, 2026 the agent itself is open source, so a reviewer can read it instead of trusting a description. And since September 16, 2026 it writes memory to disk between sessions, so that review has more ground to cover.
Open source July 15
xAI open-sourced Grok Build: the agent loop, the developer tools, the terminal interface and the skills, plugins, hooks, MCP and subagent systems. That changes what a security review can actually see. One catch. The announcement page doesn’t state a license, so check the terms in the repository before you build on it.
- Enables local-first use with your own inference
- Read the agent loop instead of trusting a description
- License not stated on the announcement, so verify it
- Repository: github.com/xai-org/grok-build
Web and mobile August 19
Grok Build reached web and mobile on August 19, 2026, the same day Grok 4.6 reached Amazon Bedrock. A coding agent that’s no longer terminal-only is a governance change, not just a convenience, right? The group who can start a long-running agent just got a lot bigger than the group with a configured CLI.
- Not terminal-only any more
- Decide who may launch agents from a phone
- Keep repository permissions as the real control
/goal Long running
A command for handing over a sustained outcome instead of driving every agent turn by hand. This is the CLI’s long-running execution mode, and it isn’t the same thing as consumer Automations.
- Good for bounded projects
- Requires completion criteria
- Use stop conditions
Agent dashboard
Run many coding sessions, watch progress, answer questions and send out more work.
- Parallel engineering
- Central supervision
- Avoid duplicate changes
Plugin marketplace
Extend Grok Build with packaged tools and workflows inside the coding environment.
- Review plugin trust
- Pin required versions
- Keep permissions narrow
Composer 2.5
A specialized coding model you can pick inside Grok Build for fast software tasks.
- Use the model menu
- Benchmark on your repository
- Don’t assume the flagship is always the best route
External coding integrations
Grok subscriptions or models can work through environments such as Warp, Kilo Code and OpenCode.
- Know which tool owns execution
- Check data terms
- Keep configuration documented
Memory September 16
Grok Build now writes notes to disk and keeps them. xAI describes them as conventions, decisions and project facts, stored as markdown, scoped per project workspace and globally. The /memory command opens a read-only browser of the files, and /dream sorts notes into topic files. xAI says it deliberately leaves out task state, temporary conclusions and secrets. Take that as design intent, not as something you’ve tested. Memory applies to new sessions, so run /new or start a fresh grok to pick it up.
- Per-project and global scope, stored as markdown on disk
- Read the stored files with
/memorybefore trusting what the agent remembers - A new artifact that follows a repository around, and a new thing to review when someone leaves
- Applies to new sessions only
Build Mode July 28
xAI shipped Build Mode on July 28, 2026 in a dated article of its own, not in the release notes. Easy to miss if the release notes are how you track Build. Check how it behaves with plan mode and with the permissions you’ve already set before you write a workflow around it.
- A dated first-party release, not a release note
- Confirm its current behavior in the product
- Re-test any Build workflow written before July 28
Workflows July 23
Workflows in Grok Build shipped on July 23, 2026, also in its own dated article. It belongs in the same inventory as /goal, plugins and, since September, memory. That inventory is the list of things a Build installation can start, carry or remember without anyone typing the next prompt.
- Dated July 23, 2026
- Inventory it alongside /goal, plugins and memory
- Decide who may create and share one
Grok Build API grok-build-0.1
A public-beta API path lets developers wire the coding model into their own workflows. Both the model and the price are published this time, which isn’t something you can take for granted here.
- Model ID
grok-build-0.1, 256k context window, listed as a priced text model on the docs price table - $1.00 input, $0.20 cached input and $2.00 output per 1M tokens at short context; $2.00, $0.40 and $4.00 at long context
- Beta stability: evaluate and instrument before you depend on it
- Prices read from a docs page rewritten in place, so confirm directly
Implement organization-wide audit logging for authentication changes. Follow existing architecture, add migration-safe storage, update API and admin UI, write abuse-case tests, and finish only when all checks pass.
Inspect this repository for the highest-risk authentication and authorization flaws. Don’t change code yet. Rank findings by exploitability and show the exact evidence behind each one.
Split this migration into independent workstreams that can run safely in parallel. Identify shared files, sequencing constraints and the final integration test before dispatching sessions.
Report only: files changed, user-visible behavior, tests run, security implications, unresolved issues and the exact review I should perform before merge.
Use X as a sensor, not a source of universal truth
X will tell you what people are saying faster than anything else. It will also tell you what bots are saying, what campaigns are pushing and what a confident stranger got wrong. Same feed, right? So treat it as a sensor and not as a source. Since August 29 Grok Bot works with X directly, which puts an agent on that same surface.
Narrative emergence
See which explanations and phrases are spreading before they reach formal analysis.
Real-time sentiment
Track reaction to products, earnings, policy and events while people are still arguing about them.
Practitioner commentary
Find the engineers, researchers, operators and witnesses closest to an event.
Attention dynamics
See who is amplifying a claim and how fast it moves through communities.
Not representative
X users and visible posts aren’t a census of your customers, voters or professionals.
Coordinated activity
Bots, incentives and campaigns can manufacture the look of consensus or urgency.
Primary records still win
Use releases, filings, transcripts, source data and direct evidence for factual claims.
Recency decays quickly
Capture the date and time, because a real-time answer can go stale within hours.
Image, video, speech and audio are becoming one creative stack
One pass now gives you motion, physics, ambience, sound effects and speech. Imagine Video 1.5 does all of it together, and it gained a References release on July 31, 2026. The image side moved to 2.0 on August 7, 2026, then took an API update on August 28. Faster experiments, yes. It also bundles several quality risks into one output, so a single clip can fail in four ways at once. Four lifecycle facts belong in a calendar. The voice alias repriced itself on August 5, 2026. grok-voice-think-fast-1.0 then vanished from the price table with no dated notice. grok-voice-transcribe-1.0 is now headed for deprecation on no date at all. And grok-imagine-image-quality retires on November 2, 2026.
Imagine image API, August 28 Parameter change
A changed default is a silent change in what comes out. The August 28, 2026 release notes add an auto quality parameter on grok-imagine-image-2.0 and move the default from medium to auto. Editing now takes up to 5 source images, up from 3, and two aspect ratios were added: “21:9 (cinematic widescreen) and 5:2 (wide banners).” Re-run your reference prompts. Don’t assume parity.
- Default moved from medium to auto
- Up to 5 source images for editing, up from 3
- New 21:9 and 5:2 aspect ratios
Grok Imagine Image Quality Retires November 2
The dedicated quality endpoint stops being its own model on November 2, 2026. Requests keep working. They just get served by grok-imagine-image-2.0 with quality set to low, a cent cheaper per image at every resolution. Cheaper output at lower quality with no error is the kind of change that reaches a brand review before it reaches an engineer.
- Quality today: $0.05 at 1K, $0.06 at 1.5K, $0.07 at 2K. Low quality on 2.0: $0.04, $0.05, $0.06
- Migrate deliberately rather than being migrated
- Set the quality parameter explicitly on 2.0
- Add a visual regression check before November
Grok Voice Think Fast 2.0 Live since August 5
This is what grok-voice-latest has served since August 5, 2026. The docs price table now names the speech-to-speech row inline as grok-voice-think-fast-2.0 at $0.08 per minute, which works out at $4.80 per hour. That’s stronger evidence than guessing the current model from what the alias seems to serve.
- Announced July 29, 2026 at $0.08 per minute of audio plus $0.004 per text input
- Priced per minute of audio, not per token
- Named inline on the price table, so the current model is no longer an inference
- Budget the hourly rate for anything conversational
Grok Voice Think Fast 1.0 Removed, not deprecated
The previous voice model, and a good lesson in reading a page twice. On September 3, 2026 the docs marked it [Deprecated] at $0.05 per minute of audio, with the same $0.004 per text input. That read like a rate you could still get by naming the version. The page no longer supports that. On October 1, 2026 the Voice API section has three rows, speech to speech, speech to text and text to speech, and grok-voice-think-fast-1.0 is in none of them and carries no deprecation marking. Its own model page does render now: modalities text and audio, region us-east-1, and a Pricing heading with nothing beneath it. So the model is documented and unpriced at the same time. Whether it still answers, and at what rate, remains unestablished. Either way, stop planning on $0.05.
- Gone from the price table with no dated retirement notice
- No longer marked [Deprecated] anywhere
- Model page renders with an empty Pricing section, which is not a rate
- Still callable or not: unresolved
- Budget at the 2.0 rate and migrate deliberately
Grok Voice Transcribe 2.0 Becoming the default · no date
This one flipped. Grok Voice Transcribe 2.0 was the opt-in model you had to name. On September 18, 2026 xAI announced that 2.0 will become the default in the Speech-to-Text API and that 1.0 will be deprecated in the coming weeks. No date. That’s the worst shape a change can take: a committed deprecation with nothing to put in a calendar and nothing to raise a change request against. The prices are published at least. Batch transcription stays $0.10 per hour of audio and streaming $0.20 per hour, with diarization, timestamps and key terms included. xAI says you can pin grok-voice-transcribe-1.0 to stay on it during the transition. That buys weeks, not a plan. If your pipeline needs 1.0 behavior, budget the migration now and benchmark 2.0 on your own audio this month.
- $0.10 per hour batch, $0.20 per hour streaming, diarization, timestamps and key terms included
- 2.0 will become the default and 1.0 will be deprecated, both “in the coming weeks” with no published date
- No switchover date exists, so there’s nothing to diarize and the first sign may be in your output
- The 2.0 model page carries a Pricing heading with no rates under it: take the rates from the article
- Benchmark 2.0 against 1.0 on your own audio before the default moves
Imagine Video 1.5 June 16
Better motion, better physics, synchronized audio and clearer speech. Released June 16, 2026, then extended by Imagine Video 1.5 with References on July 31, 2026. xAI describes making videos from text or image references with preset voices, up to 1080p.
- Quality and Fast modes
- References release dated July 31, 2026
- Product and API access
- Library search for generated media
Documented video resolutions and rates
xAI lists 480p, 720p and 1080p output. The rates are $0.08 per second at 480p, $0.14 at 720p and $0.25 at 1080p.
- At the listed rate, six seconds of 1080p is $1.50 of generation
- Choose the resolution the channel actually needs
- A living docs page, so re-check the rates before you budget
Video 1.5 Fast
Six-second 720p clips in about 25 seconds, according to xAI.
- Rapid iteration
- Useful for social and concepts
- Inspect every clip for sync and warping
Audio and speech
Sound effects, ambience and dialogue come out with the video instead of arriving as a separate first pass.
- Faster concepting
- Voice and brand risk
- Still needs final sound review
Imagine Image 2.0 August 7
The image side moved to version 2.0 on August 7, 2026 and gained new parameters on August 28. New version, new output. Re-test any prompt library or brand guardrail you built against the old one before you assume the pictures still match.
- New version, new output characteristics
- Re-run your reference prompts
- Re-check text, likeness and product truth
Voice APIs
xAI offers speech-to-text, text-to-speech, voice-agent tools and custom voice work, and shipped 21 new flagship Grok voices on July 6, 2026. The Voice Agent Builder is still marked Beta.
- Multilingual workflows
- 21 flagship voices since July 6, 2026
- Consent and impersonation controls
- Secure voice assets
Grok moved into Microsoft 365 and Google Workspace
The add-in program started in June, not July. Grok for PowerPoint shipped on June 16, 2026 and Grok for Word on June 18. Then Excel on July 20, Outlook on July 21 and Google Workspace on July 24. The July releases are the ones people noticed. They aren’t the beginning, and a policy written around them was written around the wrong date. Grok now sits beside the documents your teams already work in, on surfaces Microsoft and Google own. Since August it also sits inside those companies’ model catalogs and inside GitHub Copilot. And since September 12, 2026 a second, completely separate mechanism reaches Word, Excel and PowerPoint, this one run by Microsoft through its Frontier Program. Keep the two apart, because they share nothing but the app names. Treat all of it as a distribution decision, not just a feature release.
Grok for PowerPoint June 16
The first add-in in the program, out on June 16, 2026, about a month before the July wave people remember. If your add-in policy was written around the July dates, it was written around the wrong beginning.
- The add-in program starts here, not in July
- Confirm the tenant add-in policy before any rollout
- Review what the add-in may read from a deck
Grok for Word June 18
Two days later, on June 18, 2026, drafting and revision arrived beside the open document.
- Second in the program, ahead of Excel
- Drafting and revision beside the open document
- Review connector scope with your Microsoft 365 admin
Grok for Excel July 20
A free Microsoft 365 add-in. Ask a spreadsheet questions in plain English, get formulas back, and run scenarios and forecasts against the open workbook.
- Plain-English queries over your data
- Formula generation and scenario runs
- Connectors for SharePoint, Drive and email
Grok for Outlook July 21
Thread summaries, reply drafts matched to how you write, and inbox triage inside Outlook. Who is entitled to it’s unconfirmed. No x.ai page says which paid plans include it. The pricing page never mentions X Premium or X Premium+, and X plan entitlements live on x.com rather than on any xAI surface. So the common claim that every paid X and SuperGrok user gets it has no first-party support.
- Entitlement unconfirmed: no x.ai page states which paid plans include it
- Style-matched draft replies
- Review drafts before sending, because the tone is imitated rather than verified
Grok in Google Workspace July 24
One free add-on from the Google Workspace Marketplace that puts Grok beside the open file in Sheets, Slides and Docs.
- A single add-on covers all three apps
- Installed through the Marketplace
- Confirm your admin allows Marketplace installs
Two separate routes reach Word, Excel and PowerPoint Don’t merge them
Same three Office apps. Two mechanisms that share nothing but the app names. The first is xAI’s own: free add-ins dated June 16 to July 24, 2026, installed from the Microsoft 365 and Google Workspace marketplaces and governed by your tenant add-in policy. The second is Microsoft’s. Grok reaches Word, Excel and PowerPoint through the Microsoft Frontier Program, behind a dedicated admin setting that’s off by default. Frontier customers in the European Union, the European Free Trade Association and the United Kingdom are excluded during the preview, with no end date given. It isn’t Cowork and not Copilot Studio. Different owner. Different gate. Different regional terms. Same app names, right? And only one of the two is closed to European users today. A policy written about one of them doesn’t cover the other, and anyone who merges them gets both the gate and the geography wrong.
- xAI add-ins: June 16 to July 24, 2026, marketplace install, tenant add-in policy
- Frontier Program: September 12, 2026, a Microsoft admin setting that’s disabled by default
- Only the Frontier route carries an EU, EFTA and UK exclusion, and it has no published end date
- Ask which of the two a team means before you approve or forbid it
- Source for the Frontier route: Microsoft’s post “Expanding model choice in Copilot with Grok,” published September 12, 2026 and updated September 14. No xAI page mentions it
Grok inside the rival catalogs June to August
Grok 4.6 has been in GitHub Copilot since August 14, 2026, Amazon Bedrock since August 19, the Gemini Enterprise Agent Platform since August 21 and Microsoft Foundry since August 26. Grok has been on Databricks since June 18, 2026 and on Bedrock since June 17. Add-ins put Grok next to the documents. Catalog listings put it inside the platform contract. Either route can bring Grok into a company without a decision from whoever owns AI policy.
- Four surfaces carried Grok 4.6 in twelve days, and two catalogs carried Grok from June
- Procurement may already permit it
- Decide the policy before a team discovers the listing
Automations July 16
Grok can run recurring tasks on its own, on a schedule or from an email trigger. Every run opens as a full conversation thread you can read afterward.
- Scheduled automations for all users
- Email triggers are SuperGrok-only
- Every run leaves an inspectable thread
Automations, Routines and /goal are three different things Clarification
Three scheduling concepts. Three surfaces. Three owners, and three very different blast radii. Automations is the consumer feature for recurring tasks, on a schedule or an email trigger. The /goal command is the long-running execution mode inside the Grok Build CLI. Routines are the third. xAI introduced them in its September 3, 2026 post on designing Grok Bot, which describes them as giving a Bot a standing responsibility that runs on a schedule or in response to an event. A policy document that says only “scheduled tasks” now covers none of them properly.
- Automations: recurring, consumer, scheduled or email-triggered
- Routines: a standing responsibility held by a Grok Bot, scheduled or event-driven
- The
/goalcommand: sustained engineering execution in the CLI - Name the one you mean and don’t substitute one for another in a plan
Why this matters Strategic
Shipping into Excel, Outlook, Sheets, Slides and Docs puts Grok on the incumbents’ own ground, where Microsoft and Google ship their own assistants. Adoption can now happen with nobody making a purchasing decision.
- Expect overlap with Copilot and Gemini entitlements
- Set an add-in approval policy before demand arrives
- Audit which connectors each add-in requests
Grok’s enterprise story now extends beyond X
Four of the seven published plan tiers carry no price at all. One live x.ai page does show a figure: $59 per user per month on x.ai/enterprise. Don’t quote it. That page is boilerplate from another product and never mentions Grok. And on the pricing page and the developer security FAQ, the only attestation named is SOC 2. Start there, because price and compliance are what a buyer asks first. Then ask the third one: your API traffic is stored for 30 days by default, and turning that off costs you features. Grok Business and Enterprise add company knowledge, administration and security, and since September 3, 2026 Grok Bot carries access, network and audit controls of its own. So the question is no longer just whether staff may use a consumer assistant, right? And it’s no longer only a question about buying from xAI directly.
| Capability | Business | Enterprise | Vault (add-on to Enterprise) |
|---|---|---|---|
| Team management and billing | Included | Included | Included |
| Google Drive grounding | Permission-aware | Permission-aware | Permission-aware |
| SSO and SCIM | Not included | Included | Included |
| Advanced audit and security | Standard controls | Advanced | Advanced |
| Dedicated data plane | Not included | Optional add-on | Included |
| Customer-managed keys | Not included | Optional add-on | Included |
| SOC 2 (Type I and II) | Named on the pricing page | Named on the pricing page | Named on the pricing page |
Published plan prices Four tiers unpriced
Seven tiers on the pricing page, read October 1, 2026, and only three of them show a figure. Free at $0 per month, SuperGrok at $30 and SuperGrok Plus at $100. The other four are columns with buttons and no number: SuperGrok Lite with a Get Lite button, SuperGrok Heavy with Get Heavy, Business with Get Business, and Enterprise routing to Contact. SuperGrok Lite sits between Free and SuperGrok, which is the position a cheap paid tier takes, and xAI has published no price for it. The unpriced tiers don’t split neatly into consumer and business either. Two of the four are consumer. Read the page structure, not just the columns. The line about talking to sales for custom rate limits, dedicated infrastructure, SSO, compliance requirements and volume pricing doesn’t sit with Business. It lives in a separate “Need a custom plan?” section attached to Enterprise. Quote it as a Business route and you’ve misread the page.
- Seven tiers: Free, SuperGrok Lite, SuperGrok, SuperGrok Plus, SuperGrok Heavy, Business and Enterprise
- Four unpriced: SuperGrok Lite, SuperGrok Heavy, Business and Enterprise
- The talk-to-sales line belongs to Enterprise, not Business
- No per-seat Business price is published anywhere on the pricing page
- Get any Business or Lite price from xAI in writing
- Don’t reuse the SuperGrok $30 as a business rate
A first-party page is serving an Enterprise figure that isn’t a price $59 · not a Grok price
On October 1, 2026 x.ai/enterprise carries a heading reading For Enterprise, a section headed About the Enterprise Plan, and a figure rendered as $59 per user per month. Meanwhile x.ai/pricing publishes no Enterprise price at all and routes to Contact. A buyer will find that figure on a live x.ai path, and it must not be quoted as a Grok Enterprise price. The same page shows placeholder values throughout. It claims your company could save 0 hours a month per person, $0 in savings and 0 percent return. And its feature copy talks about calendars, Slack and scheduling assistants named Amy and Andrew. Grok isn’t mentioned once. It reads as template or staging content served on a live x.ai path. If a vendor conversation or an internal business case cites the figure, ask for the number in writing on paper that xAI will stand behind.
Zero data retention, 30 days by default, and a residency hole Procurement
By default xAI stores every API request and response on its servers, encrypted at rest, for 30 days, for auditing. xAI says it never trains on your API inputs or outputs without explicit permission. You can turn storage off with zero data retention, but it restricts certain features, so find out which ones before you promise it to a security reviewer. Then there’s the part a residency clause can trip over. xAI offers a US regional endpoint, and Files, Collections and network paths sit outside that residency guarantee. So a contract that says US-only data handling can be true of the model calls and false of the documents you uploaded to answer them. Ask about those three by name.
- API requests and responses stored encrypted for 30 days by default
- Zero data retention is available and restricts certain features: get the list
- xAI states it does not train on API inputs or outputs without explicit permission
- The US regional endpoint at
us.api.x.aidoes not cover Files, Collections or network paths - All of this lives on a developer FAQ page that is rewritten in place, so confirm it in a contract
Company knowledge
Grok can search Google Drive while respecting the permissions already set there.
- Citation previews
- Highlighted relevant passages
- No access to unshared files
Data commitment
xAI says business data isn’t used to train its models.
- Still review product terms
- Map regional data needs
- Control connector access
Collections API
Agentic search across large document stores, for data rooms, legal review and financial modeling.
- Curate the source set
- Check retrieval coverage
- Preserve citations
Enterprise controls
SSO, SCIM, administration and audit features for larger managed deployments.
- Lifecycle users correctly
- Assign product ownership
- Monitor usage
Vault Add-on, not a tier
An isolated data plane with application-level encryption and customer-managed keys, aimed at the most sensitive customers. Write down what it isn’t. Vault isn’t one of the seven plan tiers on the pricing page, which are Free, SuperGrok Lite, SuperGrok, SuperGrok Plus, SuperGrok Heavy, Business and Enterprise. Treat it, and the dedicated data plane and customer-managed keys beside it in the table above, as options layered on Enterprise. Not as another plan to buy.
- Not one of the seven published plan tiers
- An option on top of Enterprise, priced through sales like Enterprise itself
- Evaluate operational model, logging and recovery
- Verify contractual scope in writing
Third-party safeguards evaluation September 1
A lab tested Grok 4.6 and got two numbers that pull against each other by design. LatchBio measured 59.2% refusal on adversarial tasks and 64.8% completion on legitimate ones. Those figures are LatchBio’s, published by xAI. It’s a safeguards check, not a product change. Refusing more bad requests and completing more good ones fight each other, so read the pair together and never one alone.
- Source: “Biosecurity at the frontier,” xAI, September 1, 2026
- Figures produced by LatchBio, published by xAI
- A safeguards study, not a capability or product release
- Useful evidence for a security review, not a compliance certificate
Skills and apps
Reusable work behavior and connected tools turn the product into a workflow layer for the organization.
- Govern creation and sharing
- Review third-party connectors
- Retire stale skills
SOC 2 is the certification, and it’s on two different pages Compliance
Two xAI pages talk about compliance: the pricing page and the developer security FAQ at docs.x.ai/developers/faq/security. On both, SOC 2 is the only attestation named, and the two don’t quite match. The pricing page names SOC 2 Type I and Type II across multiple tiers including Free, and names nothing else: no ISO 27001, no HIPAA, no FedRAMP, no GDPR. The developer security FAQ says something narrower, that xAI is SOC 2 Type 2 compliant, and it goes further on the other two. It references GDPR, and it routes HIPAA to a BAA questionnaire rather than refusing it outright. So a reviewer who reads only the pricing page will conclude xAI has no GDPR or HIPAA position, and that’s wrong. Read both pages, and get the answer that matters into a contract. The safety page separately links an xAI Frontier AI Framework dated June 30, 2026. That’s a policy document, not an attestation, so don’t hand it to a reviewer as one.
- Pricing page: SOC 2 Type I and Type II, and no other certification named
- Developer security FAQ: SOC 2 Type 2, a GDPR reference, and a BAA questionnaire route for HIPAA
- No ISO 27001 or FedRAMP claim on either page
- Ask for the report itself, not the badge
- Both pages are rewritten in place, so confirm directly
Three price levers that move the bill API commercials
The headline $2 and $6 isn’t the whole bill. Three switches sit behind it, and two of them push the number up. The docs price page lists a US regional endpoint at 1.1 times standard rates and Priority Processing at 2 times standard rates. The Batch API is the one that pulls down, and the flagship can’t use it: grok-4.7 has no Batch API support at all. So the cheapest way to run xAI’s best model is the full-price way. Model pages state region us-east-1, while the status page lists four API regions as operational: US East, US West, EU West and central US. Those two facts don’t agree, and nothing on xAI reconciles them. Before you sign off a budget, find out which levers a team has switched on and which region it’s actually hitting.
- US regional endpoint: 1.1x standard rates
- Priority Processing: 2x standard rates
- Batch API: the one lever that lowers the bill, and not available on
grok-4.7 - Model pages say us-east-1;
status.x.ailists four regions. Ask xAI which is the contractual answer - Read from a docs page rewritten in place, so confirm directly
Grok Bot for Enterprise September 3
xAI says enterprises need to govern Bots at scale, and that this release adds the access, network and audit controls to do it. Its isolation claim has two halves. Each user’s work in Grok Bot runs in its own secure, isolated environment, separate from every other user. And a Bot has no access by default, reaching only the accounts you sign it into. xAI names Legora, Supermicro and ServiceTitan as customers and says thousands of organizations have adopted Grok Bot since launch. All of that is vendor statement, not audited finding. The post also offered free usage for two weeks from September 3. That expired on September 17, 2026, and no price is stated for the period after it.
- Access, network and audit controls, as described by xAI
- Isolation and adoption figures are vendor claims, not findings
- The free-usage offer expired on September 17, 2026
- A full security architecture document is linked from the post, and the isolation claim rests on it rather than on the post
- Read that document before relying on the isolation claim
Both case studies are xAI measuring itself Vendor self-report
There are two of them now, and neither has a third party in it. On September 4, 2026 xAI pointed Grok Bot at its own vendor spend and published more than $100,000 in direct savings. On September 22 it published its own support desk numbers: a 175 percent rise in tickets absorbed with zero new hires, roughly 200 employees not hired, and cost per resolution of $0.20 to $0.30 against a stated industry range of $1 to $4. Read both the way you read the LatchBio figures, then one notch more carefully. The vendor ran the study, on its own data, and published its own result. Useful descriptions of a workflow. Not evidence of what the same agent would do in your company.
- Procurement, September 4, 2026: $14,220 from 43 unused SaaS seats, $85,662 annually in unused SKUs, an office supply order cut from $14,629 to $6,143
- Customer support, September 22, 2026: trained on over one million customer interactions, 99 percent of refund requests resolved without a human
- xAI’s own data, xAI’s own measurement, xAI’s own publication, both times
- No independent verification of any line item
- Useful as a workflow template, not as a savings forecast
- Any pilot should establish its own baseline before it starts
Nine habits for using Grok without confusing speed with truth
Grok pays off when a team keeps four things apart on purpose: live signal, verified evidence, blunt critique and bounded execution. And when nothing you ship leans on a moving alias, an undated absence claim, or a model that can walk off a price table without notice.
Use X for signal
Map the reaction and the language. Don’t call it representative truth.
Verify the factual core
Move every claim that matters to a primary record.
Time-stamp live answers
Say when the signal was observed, because it will change.
Ask for direct critique
Use Grok’s tone to stress-test an idea, then go back to the evidence.
Use /goal for outcomes
Give Build a finish line, tests and a boundary on what it may do.
Inspect external state
Look at the code, the file, the post or the media. Not the claim that it’s done.
Fetch the file, not the folder
On xAI, a bare directory path 404s while the pages inside it resolve. data.x.ai does it with model cards. docs.x.ai/developers/migration does it with the retirement guides. Both traps read as a dead section. An absence claim needs a direct fetch of the exact document, every time.
Name region, surface and date
API, consumer app, cloud catalog and geography are four separate claims, and they landed on four different days.
Pin the model ID
Never ride an alias or a default. grok-voice-latest repriced itself on August 5, 2026 with no action by the customer. grok-imagine-image-quality silently drops to low quality on November 2, 2026. And the Speech-to-Text default moves from grok-voice-transcribe-1.0 to 2.0 on no announced date.
Fast social intelligence is noisy by construction
The things that make Grok good are the same things that make it risky. Live feeds can be gamed. A confident tone can hide uncertainty. And an integrated agent can act before a weak claim gets corrected. Two more risks sit on xAI’s own pages. A bare directory path can return 404 while every page inside it resolves, which makes an absence easy to assert and hard to prove. And xAI rewrites undated pages without notice, in every direction: a price gets reconciled upward, a model vanishes from the price table with no retirement date, and a flagship gets replaced on a page that looks identical the next morning.
Selection bias
Visible X conversation isn’t the same thing as broad public, customer or employee sentiment.
Coordinated manipulation
Bots and campaigns can manufacture salience, repetition and the look of consensus.
Confident tone
Directness can make an uncertain reading feel settled. Ask for sources and calibration.
A bare xAI directory 404s while every page inside it resolves
This trap has two instances now, and it reads like a dead section both times. data.x.ai/ and data.x.ai/model-cards return 404 while three model card files beneath them serve normally, all three fetched directly on October 1, 2026. docs.x.ai/developers/migration and docs.x.ai/developers/deprecations return 404 while the retirement guides at docs.x.ai/developers/migration/may-15-retirement and docs.x.ai/developers/migration/imagine-image-quality-nov-2 both resolve and are linked from the models page. So a check aimed at a folder reports the contents as gone. Fetch the file, not the folder, and be as precise about the shape of an absence as about the absence.
Model cards are per model, and they sit on two hosts
Six cards exist. Grok 4.7 (September 21, 2026, revision 2026-09-21), Grok 4.6 (August 12, 2026, revision 2026-08-17) and Grok 4.5 (July 14, 2026, revision 2026-07-20) are on media.x.ai. Grok 4 Fast, Grok Code Fast 1 and Grok 4 are on data.x.ai. All six are text models. There’s no card for any voice model, for grok-build-0.1 or for any Imagine model. A card for the current model says nothing about the one you deployed in July. Match the card to the model ID your product actually calls.
An announced model with no ID to buy
x.ai/api has carried three different flagships inside four weeks. grok-4.3 at $1.25/$2.50 in early September, then grok-4.6 at $2.00 and $6.00, and on October 1, 2026 grok-4.7, with grok-4.6 pricing no longer shown. xAI changes that page with no notice and no date, so a figure copied from it two weeks earlier is quietly wrong. The live hazard runs the other way. Both the August 12 Grok 4.6 post and the September 21 Grok 4.7 post say there’s a fast variant at twice the price, and no model ID with the word fast in it exists on any xAI pricing surface. Don’t budget for a model you can’t name. And if a vendor conversation quotes the fast-variant multiplier, ask for the ID.
Grok Business pricing is unconfirmed
The $30 per seat per month figure that circulates for Grok Business is UNCONFIRMED and is published nowhere. x.ai/pricing, read October 1, 2026, shows SuperGrok at $30 per month and SuperGrok Plus at $100 per month. Business shows no per-seat price at all, only a Get Business button. SuperGrok Lite, the tier sitting just above Free, shows no price either. The invitation to talk to sales sits in a “Need a custom plan?” section attached to Enterprise, not to Business. The $30 is the SuperGrok consumer price, and that’s how it travels. Don’t quote a Business seat price. Get it from xAI sales in writing.
xAI is selling a plan that isn’t on its own pricing page
Team Bots went into public beta on September 28, 2026, and xAI says it’s available on Teams and Enterprise plans. There is no Teams tier. The pricing page lists Free, SuperGrok Lite, SuperGrok, SuperGrok Plus, SuperGrok Heavy, Business and Enterprise. A buyer told they need a Teams plan can’t find one to buy, which makes a renewal conversation harder than it needs to be. Ask which published tier the word Teams maps to, and get the answer in the order form rather than in an email.
A silent quality downgrade has a date
On November 2, 2026 grok-imagine-image-quality retires and its requests get served by grok-imagine-image-2.0 at low quality, a cent cheaper per image at every resolution. No error. No outage. Cheaper invoice, worse images. Put a visual check and a decision in the calendar before the date, because nothing will alert you on the day.
A deprecation with no date is worse than one with a date
On September 18, 2026 xAI announced that Grok Voice Transcribe 2.0 will become the default in the Speech-to-Text API and that 1.0 will be deprecated in the coming weeks. That’s it. No switchover date, nothing in any release note, nothing to diarize. A dated shutdown you can raise a change request against. An undated one you can only watch for, and the only place it will appear is in your output. If a pipeline depends on 1.0, treat the migration as already started. Prices are at least published: $0.10 per hour batch and $0.20 per hour streaming.
Voice alias reroute, executed
On August 5, 2026 grok-voice-latest began serving grok-voice-think-fast-2.0. That’s 60 percent more per minute of audio, $0.05 to $0.08, with nobody on your side doing anything. The warning was published a week ahead and it landed on the day, which is the argument for keeping a watchlist. The price table now names the 2.0 model inline on the speech-to-speech row, so the current model is documented even though the alias target itself is still not restated on any page.
A model can leave a page with no notice at all
In early September grok-voice-think-fast-1.0 was marked [Deprecated] at $0.05 per minute of audio. By October 1, 2026 it’s simply absent from the price table. No deprecation marking. No retirement date. No dated article. Its model page renders, and the Pricing heading on it has nothing underneath. A dated deprecation is something you can diarize, raise a change request against and argue about with a vendor. A silent removal is none of those. It looks tidier on the page and it’s strictly worse for you. Treat any model that disappears from a vendor price table as a migration you’ve already started, whether or not it still answers when you call it.
Don’t count anything on the news index
The x.ai/news index has served a current window while leaving out real articles that exist at their own URLs, including xai-joins-spacex and anthropic-compute-partnership. It still isn’t stable. Two reads on October 1, 2026 returned different totals, and the second read dropped an article from its date window that the first one listed and that resolves perfectly well at its own URL. So the index can’t support a count, and it can’t support a claim that something wasn’t published. Never enumerate xAI releases from it. Link the dated article, and when you need to prove a gap, name the URLs you fetched.
Beta language is sticky
Grok Bot went from three plans to eight on August 26, 2026, gained an X integration on August 29 and gained enterprise access, network and audit controls on September 3. None of those posts uses the word beta, and none says it left beta either. Then on September 28, forty-eight days after launch, xAI shipped Team Bots and wrote public beta in plain words. So the beta label is still live on the product line, even with governance controls and a thousands-of-organizations adoption claim attached. Widen the availability claim to the named plans. Stop short of calling it generally available, because xAI hasn’t.
Regional and tier availability, and why the EU isn’t settled
Grok 4.5 launched without EU access and gained it on July 17, 2026 on xAI’s own API console. That single date is the most misread fact about Grok in Europe. Availability varies by surface. GitHub Copilot, Amazon Bedrock, the Gemini Enterprise Agent Platform, Microsoft Foundry and Databricks each carry their own regions and terms, and Copilot access may need enabling in Copilot settings on business and enterprise tenants. One surface is shut to European users outright. Grok reached Word, Excel and PowerPoint through the Microsoft Frontier Program on September 12, 2026, behind an admin setting disabled by default. Frontier customers in the European Union, the European Free Trade Association and the United Kingdom are excluded during the preview, with no end date given. That rests on Microsoft’s own Copilot Blog post. No xAI page mentions it. And it isn’t the xAI Frontier AI Framework, a safety policy dated June 30, 2026 with no Microsoft content in it. xAI model pages state region us-east-1 while status.x.ai shows four operational API regions, US East, US West, EU West and central US. Those two readings disagree and nothing on xAI resolves them, so treat the four-region list as a question for xAI rather than an entitlement. Check the current status per surface, per model and per tier. Not a remembered one.
Third-party evaluation isn’t certification, and self-reporting isn’t evaluation
The LatchBio evaluation xAI published on September 1, 2026 reports 59.2% refusal on adversarial tasks and 64.8% completion on legitimate ones for Grok 4.6. Those are LatchBio’s measurements on task sets somebody chose, published by the vendor whose model was tested. Two case studies are a further step away from independence. On September 4, 2026 xAI ran Grok Bot against its own vendor spend and published its own savings. On September 22 it published its own support desk figures the same way. Cite all three with their attributions. Present none of them as an approval or a forecast.
Agent permissions
Build plugins, external integrations and business apps widen the blast radius of a mistake, and Grok Build has been on web and mobile since August 19, 2026.
Voice and likeness
Custom voices and generated speech need explicit consent, impersonation controls and secure handling of the assets.
Integrated media errors
Video, physics, speech and audio can each fail on their own inside one polished clip. And changed defaults, like the August 28 move from medium to auto quality, shift the output without you doing anything.
Agent memory is now a security scope
Since September 16, 2026 Grok Build writes what xAI calls conventions, decisions and project facts to markdown files that live across sessions, per project workspace and globally. xAI says task state, temporary conclusions and secrets are deliberately left out. Take that as design intent, not as an assurance you’ve tested. An agent that carries notes between sessions hands you three new things. A new place for sensitive detail to land. A new artifact that travels with a repository. And a new thing to review when somebody leaves the team. Read the stored files with /memory, and decide who is allowed to.
Three cutoffs for one model
Grok 4.7’s knowledge cutoff reads May 2026 on the docs model reference. The model card says a pretraining cutoff of June 2026, with supplemental data as late as August 2026. The dedicated 4.7 model page says nothing at all. Three months of spread, three first-party pages, no reconciliation. The May date sits on the index page everybody opens first, which is exactly why it gets quoted as settled. Name the surface whenever you state a cutoff. And never let a recency assumption ride on a number xAI’s own pages disagree about.
The retirement list is readable, and it names eight models
xAI’s migration guide for May 15, 2026, last updated September 2, lists eight models switched off at 12:00 PM Pacific that day. grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning, grok-4-fast-reasoning, grok-4-fast-non-reasoning and grok-4-0709 all went to grok-4.3 at various reasoning efforts. grok-3 went to grok-4.3 with reasoning effort none. grok-code-fast-1 went to grok-build-0.1. grok-imagine-image-pro went to grok-imagine-image-quality, which is itself retiring on November 2. Read that last chain twice. A replacement model can be retired before you’ve finished migrating onto it. The guide lives at a child path. The folder above it 404s.
Rapid product churn
The release cadence is why dated claims and direct source links aren’t optional here.
Where these facts come from
Each entry below is a dated xAI article, a model card, an xAI migration guide or one Microsoft post. None is the rolling news index. SpaceX acquired xAI on February 2, 2026, so page titles now render “SpaceXAI” while footers still read “xAI Corp.,” and xAI is still the operating brand. Each slug is the one that actually resolves. Don’t rebuild a URL from a headline: x.ai/news/biosafety-at-the-frontier serves “Biosecurity at the frontier,” and x.ai/news/grok-4-6-vertex-ai serves “Grok 4.6 on Gemini Enterprise Agent Platform.”
Live pages behind the figures, not listed below Rewritten in place
These pages carry no date and change without notice, so they can’t anchor a citation. Every price, context window, plan tier and security statement above comes from one of them, read October 1, 2026. Confirm directly before you rely on a figure, and keep your own dated copy.
docs.x.ai/developers/pricing: API, voice and Imagine prices, the eight price-table rows, the 1.1x regional and 2x Priority rates, no Batch API ongrok-4.7docs.x.ai/developers/modelsand the per-model pages: the May 2026 cutoff, reasoning effort settings, the empty Pricing sectionsdocs.x.ai/developers/release-notes: Grok 4.5 on the API July 8, EU console access July 17, the August 28 Imagine changesdocs.x.ai/developers/faq/security: SOC 2 Type 2, GDPR, the HIPAA BAA route, 30-day storage, zero data retention, the residency carve-outx.ai/pricing: seven tiers, four unpriced, the SOC 2 claimx.ai/api,x.ai/voice,x.ai/safetyandstatus.x.aix.ai/enterprise: the $59 figure, an observation and never a price
Two migration guides are listed anyway Fetch the file, not the folder
They sit in documentation. But they’re the only authoritative record of when an xAI model dies, and each carries its own last-updated date. Both were read October 1, 2026. Watch the shape of the trap around them. docs.x.ai/developers/migration and docs.x.ai/developers/deprecations return 404 while the child pages resolve.
Six model cards on two hosts Not a dead host
Grok 4.7, Grok 4.6 and Grok 4.5 are on media.x.ai. Grok 4 Fast, Grok Code Fast 1 and Grok 4 are on data.x.ai. That host’s two index paths, data.x.ai/ and data.x.ai/model-cards, return 404 while the files resolve. So don’t report the host as dead.
· Fetched directly October 1, 2026: the Grok 4.7 and Grok 4.6 cards and all three data.x.ai files
One route rests on Microsoft, not xAI No xAI source
Grok in Word, Excel and PowerPoint through the Microsoft Frontier Program from September 12, 2026. The admin setting that’s off by default. The exclusion of EU, EFTA and UK customers during the preview. A search across x.ai and docs.x.ai finds none of it. All three rest on Microsoft’s own post, listed below. Don’t mix it up with the xAI Frontier AI Framework, a safety policy dated June 30, 2026. Same word, different company.
· “Expanding model choice in Copilot with Grok,” Microsoft Copilot Blog, published September 12, 2026, updated September 14
One alias target rests on inference Caveat
No page states in so many words what grok-voice-latest resolves to. The August 5 reroute rests on the July 29 announcement plus the model name now printed inline on the price table.
Vendor claims, named as such Attributed inline
None of these is an independent finding. Each is attributed wherever it appears.
- Elon Musk’s “Opus-class” description of Grok 4.5: vendor positioning
- “Twice as fast, at half the price of comparable models” for Grok 4.7, and its benchmarks: vendor claims on vendor-chosen tests
- 59.2% refusal and 64.8% completion for Grok 4.6: produced by LatchBio, published by xAI
- September 4 procurement and September 22 support figures: xAI measuring its own agent on its own data
- Isolation, governance, privacy and adoption statements for Grok Bot and Team Bots: vendor claims
Arithmetic, not published figures Calculated
Two percentages are calculations on xAI’s own rates. The 60 percent voice rise is $0.05 to $0.08 per minute of audio. The $0.004 text-input charge is the same on both models, so the figure describes the audio rate, not total cost of ownership. The 17 to 20 percent Imagine cut comes from the published per-image rates and the stated $0.01 delta.
Twelve open questions Unconfirmed
Each is marked unconfirmed rather than stated, for the reason given.
- The $30 per seat Grok Business figure in circulation: the pricing page doesn’t support it
- SuperGrok Lite and SuperGrok Heavy prices: neither is published
- Whether SuperGrok Lite is new: the pricing page carries no date or changelog
- Which tier “Teams” means for Team Bots: no Teams tier on the pricing page
- X Premium Plus entitlements for Grok Build and Grok for Outlook: no x.ai page supports them
- The current target of
grok-voice-latest: no page restates it - Whether
grok-voice-think-fast-1.0still answers, and at what rate: empty Pricing section, no price-table row - When
grok-voice-transcribe-2.0becomes the default: “in the coming weeks,” no date - ID, context window and price of the fast variant announced for Grok 4.6 and Grok 4.7
- Whether Grok 4.6 is deprecated, retired or demoted: no retirement date on any xAI page
- Whether Grok Bot has left beta: no post says so, and the September 28 Team Bots beta argues against it
- The Microsoft Frontier Program route: no xAI page carries it
AI Mindset