xAI shipped Grok 4.5, Grok 4.6 and Grok 4.7 within eleven weeks in 2026, and all three are still priced on the xAI API, so teams have to choose between three closely related model IDs. This page compares the three versions side by side: API IDs, release dates, status, context window, reasoning effort settings, prices including long-context and regional rates, documented training changes, provider-reported benchmark figures, independent leaderboard results and retirement status. It is written for developers deciding which ID to pin, buyers estimating cost and architects planning migrations. Versions are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from xAI's own documentation, pricing page, release notes and announcements, or from the named leaderboard's own page, cited in the sources list.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses xAI's public documentation as of November 2026. xAI's documentation now carries the name SpaceXSI; this page keeps the Grok and xAI names used in the API. Model details change, so confirm against xAI's current documentation.
Key takeaways
- All three versions are current. grok-4.7, grok-4.6 and grok-4.5 each have a model page, a row on the xAI pricing page and a 20% Batch API discount; the release notes record no retirement for any of them.
- Same headline price, different cache price. Each costs $2.00 input and $6.00 output per million tokens below 200,000 prompt tokens; cached input is $0.50 for Grok 4.7 and 4.6 and $0.30 for Grok 4.5.
- Same 500,000-token context window. The pricing page lists 500k for all three, and long-context rates apply to every token in a request once its prompt reaches 200,000 tokens.
- Availability differs at the edges. Only grok-4.7 and grok-4.6 are served on the US regional endpoint; Grok 4.7 Fast exists only in Cursor and Grok Build; the Grok 4.5 page lists the aliases grok-4.5-latest and grok-build-latest.
- Model cards for two of the three. xAI publishes PDF model cards for Grok 4.6 (12 August 2026, revision 17 August 2026) and Grok 4.5 (14 July 2026, revision 20 July 2026); no Grok 4.7 card was found on the pages checked.
What is Grok 4?
Grok 4 is xAI's current line of large language models, served through the xAI API with a global endpoint and a US regional endpoint. In 2026 it moved through Grok 4.5 (July), Grok 4.6 (August) and Grok 4.7 (September). Each of these three versions uses one API ID with a reasoning effort parameter.
xAI's model pages describe Grok 4.7 and Grok 4.6 with the same words, "frontier model for coding, agentic tasks, and knowledge work", and Grok 4.5 as an "intelligent coding model for agentic software, engineering, and workflow tasks". The models page documents the alias scheme: the plain name points to the latest stable version, the -latest suffix to the latest version, and a dated suffix to a fixed release. For the whole catalogue see the xAI Models hub, and for one model in depth the Grok 4.7 profile.
Every model at a glance
As of 2026-11-15. Sources: xAI pricing page, model pages and release notes.
| Model | API ID | Released | Status |
|---|---|---|---|
| Grok 4.7 | grok-4.7 | 2026-09-21 (xAI API) | Current; on the pricing page |
| Grok 4.6 | grok-4.6 | 2026-08-12 (xAI API) | Current; on the pricing page |
| Grok 4.5 | grok-4.5 | 2026-07-08 (xAI API) | Current; on the pricing page |
A related variant, Grok 4.7 Fast, has no public API ID: the release notes describe it as the same model as Grok 4.7 at higher token rates, offered only in Cursor and Grok Build.
When was each model released?
As of 2026-11-15. Sources: xAI release notes and announcements.
| Date | Event | API ID |
|---|---|---|
| 2026-07-08 | Grok 4.5 released on the xAI API | grok-4.5 |
| 2026-07-16 | Grok 4.5 announcement published on x.ai | grok-4.5 |
| 2026-07-17 | Grok 4.5 available in the API console for EU users | grok-4.5 |
| 2026-08-12 | Grok 4.6 released on the xAI API and announced | grok-4.6 |
| 2026-09-21 | Grok 4.7 released on the xAI API, also on the US regional endpoint, and announced | grok-4.7 |
| 2026-09-21 | Grok 4.7 Fast offered in Cursor and Grok Build only | No public API ID |
The release date for Grok 4.5 depends on the surface: the release notes record API availability on 8 July 2026, while the announcement page is dated 16 July 2026. Record the date and the surface together in any model inventory. The LLM Release Tracker lists these releases alongside other providers.
How much does each model cost?
As of 2026-11-15. USD per million tokens on the global endpoint, from the xAI pricing page. Long-context rates apply to all tokens in a request once its prompt reaches 200,000 tokens. xAI does not state a separate cache write price.
| Model | Input | Cached input | Cache write | Output | Batch (input / output) | Notes stated by provider |
|---|---|---|---|---|---|---|
| Grok 4.7 | $2.00 ($4.00 at 200k+) | $0.50 ($1.00 at 200k+) | Not stated by xAI | $6.00 ($12.00 at 200k+) | 20% discount: $1.60 / $4.80 | US regional endpoint at 1.1x: $2.20 / $0.55 / $6.60 below 200k |
| Grok 4.6 | $2.00 ($4.00 at 200k+) | $0.50 ($1.00 at 200k+) | Not stated by xAI | $6.00 ($12.00 at 200k+) | 20% discount: $1.60 / $4.80 | US regional endpoint at 1.1x global rates |
| Grok 4.5 | $2.00 ($4.00 at 200k+) | $0.30 ($0.60 at 200k+) | Not stated by xAI | $6.00 ($12.00 at 200k+) | 20% discount: $1.60 / $4.80 | Global endpoint only on the pricing page |
- Batch: the pricing page applies the 20% discount to input, output, cached and reasoning tokens for all three models.
- Priority Processing: billed at 2x standard rates, and only when the response confirms
"service_tier": "priority"; not available for Batch API requests. - Grok 4.7 Fast: $4.00 input, $1.00 cached and $12.00 output below 200k; $6.00, $1.50 and $18.00 above, billed through Cursor or Grok Build plans.
- Announcements: the Grok 4.7 announcement states it is served at the same price and speed as Grok 4.6; the Grok 4.6 and Grok 4.5 announcements both state $2 input and $6 output per million tokens.
Worked example from the published rates: a request with 100,000 input and 10,000 output tokens and no caching costs $0.26 on any of the three models. With 80,000 of those input tokens served from cache it costs $0.14 on Grok 4.7 or 4.6 and $0.124 on Grok 4.5. A 300,000-token prompt with 10,000 output tokens costs $1.32 on each, because the long-context rates apply to every token. See the LLM Pricing Tracker for other providers.
What changed between models?
The descriptions below are xAI's own, from each announcement; xAI publishes no technical report for these versions.
Grok 4.5 (July 2026)
- Training: the announcement says it was trained alongside Cursor, across tens of thousands of NVIDIA GB300 GPUs, with data filtering by deduplication, quality scoring and domain-focused selection, and reinforcement learning on hundreds of thousands of tasks.
- Speed and tokens: it states an output speed of 80 tokens per second and an average of 15,954 output tokens per task on SWE Bench Pro.
- Reasoning effort: the 8 July release note lists low, medium and high (default high); the current Grok 4.5 model page lists low, medium, high and xhigh (default high).
Grok 4.6 (August 2026)
- Training: the announcement describes a longer supplemental training run than Grok 4.5, curated model-generated reasoning and engineering data, SFT trajectories regenerated by Grok 4.5, and an improved optimiser and training recipe.
- Focus: long-running agents and interactive and visual work, with reinforcement learning on knowledge work, coding, kernel optimisation, web development and computer-aided design.
- API profile: the release note lists a 500k context window, text and image input, text output and reasoning effort low, medium, high (default) and xhigh.
Grok 4.7 (September 2026)
- Training: the announcement says Grok 4.7 uses a new, larger base model than Grok 4.6 and a longer reinforcement learning run on a harder task mix weighted towards problems that take many hours.
- Behaviour: it states better verification of its own work, better management of longer context, and native training on the Grok Bot harness.
- API behaviour: on the Responses API grok-4.7 always returns
reasoning.encrypted_content, even whenincludedoes not request it (release notes and models page). - Safety: the announcement describes a new safeguard stack and invite-only access to the model's red-team capabilities for select cybersecurity partners.
What workloads does xAI position each model for?
| Model | Positioning stated by provider | Related notes stated by provider |
|---|---|---|
| Grok 4.7 | "Frontier model for coding, agentic tasks, and knowledge work" (model page) | Announcement: most capable model for coding and knowledge work; fast variant at twice the output speed and twice the price |
| Grok 4.6 | "Frontier model for coding, agentic tasks, and knowledge work" (model page) | Announcement: long-running agents and more ambitious interactive and visual work |
| Grok 4.5 | "Intelligent coding model for agentic software, engineering, and workflow tasks" (model page) | Announcement: default model in Grok Build and available in Cursor on all plans; model page: trained in Memphis data centres on science, engineering and maths data |
What is the context window of each model?
As of 2026-11-15. Sources: pricing page, model pages, release notes and models page.
| Model | Context window | Max output | Knowledge cutoff | Input | Output | Reasoning controls |
|---|---|---|---|---|---|---|
| Grok 4.7 | 500,000 tokens | Not stated by xAI (release note: no text output limit) | May 2026 (models page) | Text, image | Text | low, medium, high (default), xhigh |
| Grok 4.6 | 500,000 tokens | Not stated by xAI | Not stated by xAI on the pages checked (2026-11-15) | Text, image | Text | low, medium, high (default), xhigh |
| Grok 4.5 | 500,000 tokens | Not stated by xAI | Not stated by xAI on the pages checked (2026-11-15) | Text, image (model page) | Text (model page) | low, medium, high (default), xhigh on the model page |
The models page sets image input rules for the API: JPG/JPEG or PNG, up to 20 MiB per image, with no limit on image count. The 200,000-token pricing threshold matters more for cost than the 500,000-token ceiling. See Context Windows Explained and Reasoning Models Explained.
Capabilities and tools compared
| Model | Function calling and structured outputs | Computer use tool | Fine-tuning | Cloud IDs stated by provider |
|---|---|---|---|---|
| Grok 4.7 | Both listed on the model page; Batch API listed | Not stated by xAI | Not stated by xAI | Not stated by xAI |
| Grok 4.6 | Both listed on the model page; Batch API listed | Not stated by xAI | Not stated by xAI | Not stated by xAI |
| Grok 4.5 | Both listed on the model page; Batch API listed | Not stated by xAI | Not stated by xAI | Not stated by xAI |
Server-side tools on the xAI API (web search, X search, code execution, attachment search, collections search and remote MCP) are priced per call on the pricing page and are not tied to one Grok version there. Rate limits on all three model pages are 150 requests per second and 50,000,000 tokens per minute.
Where is each model available?
| Model | Availability stated by provider |
|---|---|
| Grok 4.7 | xAI API global endpoint (model page region: us-east-1, running on multiple clusters) and US regional endpoint; Cursor and Grok Build; announcement names third-party coding harnesses, model routers and cloud platforms without listing them |
| Grok 4.6 | xAI API global endpoint (model page region: us-east-1, running on multiple clusters) and US regional endpoint; Cursor and Grok Build; announcement names OpenRouter, Vercel and Cloudflare as partners |
| Grok 4.5 | xAI API (model page regions: us-east-1, us-west-2); EU console from 17 July 2026; announcement: default model in Grok Build and available in Cursor on all plans |
Consumer Grok app availability for each version is not stated by xAI on the pages checked (2026-11-15). For app plans see ChatGPT vs Claude vs Gemini vs Grok (Apps).
Model cards and system cards
| Model | Model card or page | System card | Technical report |
|---|---|---|---|
| Grok 4.7 | Model page (specifications); announcement 2026-09-21 | Not stated by xAI on the pages checked (2026-11-15) | Not stated by xAI on the pages checked (2026-11-15) |
| Grok 4.6 | Grok 4.6 Model Card (41 pages, dated 2026-08-12, revision 2026-08-17); model page; announcement 2026-08-12 | Not stated by xAI on the pages checked (2026-11-15) | Not stated by xAI on the pages checked (2026-11-15) |
| Grok 4.5 | Grok 4.5 Model Card (28 pages, dated 2026-07-14, revision 2026-07-20); model page; announcement 2026-07-16 | Not stated by xAI on the pages checked (2026-11-15) | Not stated by xAI on the pages checked (2026-11-15) |
The Grok 4.6 announcement describes pre-deployment testing for capabilities and safeguard calibration plus post-deployment and third-party testing; the Grok 4.7 announcement reports two safety evaluations, listed under Benchmarks. See Model Cards and System Cards Explained.
Benchmarks
Provider-reported figures come from each xAI announcement on its date. Benchmark versions differ between announcements (for example CursorBench 4.0 in the Grok 4.7 post and CursorBench v3.2 in the Grok 4.6 post), so figures are only comparable within one announcement. Independent figures are from each leaderboard's own page, checked 2026-11-15, and are kept in a separate column.
| Model | Provider-reported benchmarks | Independent leaderboard results |
|---|---|---|
| Grok 4.7 | xAI, 2026-09-21, xHigh unless noted: CursorBench 4.0 46.3%; DeepSWE v1.1 71.0% (high); EEBench 64.0%; AA Briefcase v1.1 1,657; Terminal-Bench 4.0 37.6%; Harvey Legal Agent 19.6%; HealthBench Professional 56.7%; LatchBio biosafety 62.4%; HackerBench v0.3: 3.3% of risky dual-use prompts let through | Artificial Analysis Intelligence Index (v5.0.1): 35 for "Grok 4.7 Xhigh". Epoch AI ECI: 156.18 (CI 154.23 to 158.27), dated 2026-09-21 |
| Grok 4.6 | xAI, 2026-09-21 (in the Grok 4.7 post, High): CursorBench 4.0 40.4%; DeepSWE v1.1 65.2%; Terminal-Bench 4.0 20.3%. xAI, 2026-08-12 (High): AA Intelligence Index 61; GDPVal-AA v2 1753; CursorBench v3.2 69.9%; DeepSWE v1.1 65.9%; FrontierCode v1.1 (Extended) 61.3%; APEX-Agents 57.5%; Terminal-Bench v3.0 26%; APEX-SWE 56.4% | Artificial Analysis: not in the leaderboard rows listed on 2026-11-15. Epoch AI ECI: 157.79 (CI 155.77 to 159.95), dated 2026-08-12 |
| Grok 4.5 | xAI, 2026-08-12 (in the Grok 4.6 post, High): AA Intelligence Index 56; CursorBench v3.2 66.7%; DeepSWE v1.1 54%; Terminal-Bench v3.0 15.7%. xAI, 2026-07-16 (effort not stated): SWE Bench Pro 64.7%; Terminal Bench 2.1 83.3%; SWE Marathon 29.0%; DeepSWE 1.1 53% | Artificial Analysis: not in the leaderboard rows listed on 2026-11-15. Epoch AI ECI: 154.98 (CI 153.22 to 156.88), dated 2026-07-08 |
The Artificial Analysis Intelligence Index figures quoted in the Grok 4.6 announcement (61 and 56) were published in August 2026; the index is now at version 5.0.1 with a different evaluation set, so they cannot be read against the current figure of 35. Epoch AI's confidence intervals for the three versions overlap. See the LLM Leaderboard Snapshot and How to Read LLM Benchmarks.
Deprecations and retirements
| Model | API ID | Notice or deprecation date | Retirement or shutdown date | Stated replacement |
|---|---|---|---|---|
| Grok 4.7 | grok-4.7 | None announced on the pages checked (2026-11-15) | None announced on the pages checked (2026-11-15) | Not stated by xAI |
| Grok 4.6 | grok-4.6 | None announced on the pages checked (2026-11-15) | None announced on the pages checked (2026-11-15) | Not stated by xAI |
| Grok 4.5 | grok-4.5 | None announced on the pages checked (2026-11-15) | None announced on the pages checked (2026-11-15) | Not stated by xAI |
The models page links migration guides for a "Model Retirement (May 15, 2026)" and an "Image Quality Model Retirement (Nov 2, 2026)"; neither is listed against these three versions on the pricing page or release notes. The alias grok-build-latest on the Grok 4.5 page can move to a different model, so pin grok-4.5 explicitly where behaviour must stay fixed. See the LLM Deprecation and Retirement Tracker and the AI Model Retirement Guide.
Data handling and security
xAI's security FAQ applies the same terms to all API models, so data handling does not differ between the three versions:
- Training use: API inputs and outputs are not used for training without explicit permission, including the 30-day stored data.
- Retention: requests and responses are stored encrypted at rest for 30 days, mainly to audit suspected abuse or misuse, then deleted automatically.
- Zero data retention: enabled team-wide by a team admin in Console Team Settings after deleting existing Files and Collections. It disables the stateful Responses API, new Files uploads, Collections, the Batch API, deferred completions and server-stored image and video outputs, so the batch discount is not available under ZDR. The FAQ states that xAI does not recommend ZDR for most customers. See Zero Data Retention for AI.
- Assurance: the FAQ states SOC 2 Type 2 compliance and offers a Business Associate Agreement questionnaire for HIPAA enquiries.
- Request controls: since 25 September 2026 requests can carry a
safety_identifierend-user ID, and since 27 October 2026 aguidelinesarray for nsfw, violence and profanity handling (release notes); neither is tied to one version.
API keys for any of these models are bearer credentials. See the LLMjacking Guide.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| US data residency endpoint | Only grok-4.7 and grok-4.6 are served on the US regional endpoint, at 1.1x global rates |
| Cache-heavy workloads | Cached input is $0.30 for grok-4.5 and $0.50 for grok-4.7 and grok-4.6 |
| Regions on the model page | grok-4.5 lists us-east-1 and us-west-2; grok-4.7 and grok-4.6 list us-east-1 |
| Stated knowledge cutoff | Stated only for grok-4.7 (May 2026) |
| Encrypted reasoning on the Responses API | grok-4.7 always returns reasoning.encrypted_content; plan storage for that field |
| Faster output tier | Grok 4.7 Fast in Cursor and Grok Build only; Priority Processing at 2x on the API |
| Aliases | grok-4.5 lists grok-4.5-latest and grok-build-latest; no aliases listed for grok-4.7 or grok-4.6 |
For cross-provider selection see Frontier AI Models Compared and Coding LLMs Compared.
Practitioner checklist
- Pin production code to grok-4.7, grok-4.6 or grok-4.5 explicitly, not to an alias, and record the ID in the AI inventory.
- Keep prompts under 200,000 tokens where possible, or budget for long-context rates on every token.
- Re-run your own evaluation set when moving between versions; provider figures use different benchmark versions per announcement.
- Confirm the US regional endpoint is needed before choosing a version, as only grok-4.7 and grok-4.6 are served there.
- Decide on zero data retention before relying on the Batch API discount, as ZDR disables batch.
- Store xAI API keys in a secrets manager and set team spend limits.
- Watch the release notes and pricing page for retirement notices on all three IDs.
Standards and references
- xAI: Release notes (accessed 2026-11-15)
- xAI: Pricing (accessed 2026-11-15)
- xAI: Models (accessed 2026-11-15)
- xAI: Grok 4.7 model page (accessed 2026-11-15)
- xAI: Grok 4.6 model page (accessed 2026-11-15)
- xAI: Grok 4.5 model page (accessed 2026-11-15)
- xAI: Grok 4.7 announcement (accessed 2026-11-15)
- xAI: Grok 4.6 announcement (accessed 2026-11-15)
- xAI: Grok 4.5 announcement (accessed 2026-11-15)
- xAI: Security FAQ (accessed 2026-11-15)
- xAI: Grok 4.6 Model Card (accessed 2026-11-15)
- xAI: Grok 4.5 Model Card (accessed 2026-11-15)
- Artificial Analysis: Models leaderboard (accessed 2026-11-15)
- Epoch AI: Epoch Capabilities Index (accessed 2026-11-15)
Related NHI Mgmt Group resources: xAI Models, Grok 4.7, Frontier AI Models Compared, LLM Pricing Tracker, LLM Leaderboard Snapshot.