Grok 4.6
Grok 4.6 was the main model from xAI, which now signs as SpaceXAI, from Aug 12, 2026 until Sep 21, 2026, when Grok 4.7 came out. It costs US$ 2 per million input tokens and US$ 6 per million output under 200 thousand tokens, exactly the price of its successor. It is still on the API with no published retirement date, but xAI itself tells developers to use 4.7, which beats it on all seven tests in the launch table. Today it only makes sense where xAI announced 4.6 and has not announced 4.7, such as Amazon Bedrock, Google Cloud, Microsoft Foundry and GitHub Copilot.
Grok 4.6 in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- xAI (SpaceXAI)
- API identifier
grok-4.6- Release
- Aug 12, 2026
- Status
- on the API, no retirement announced; no longer the main model since Grok 4.7, on Sep 21, 2026
- Input
- US$ 2 per million tokens (US$ 4 from 200 thousand)
- Output
- US$ 6 per million tokens (US$ 12 from 200 thousand)
- Cached input
- US$ 0.50 per million (US$ 1 from 200 thousand); automatic caching, no published cache write price
- Long context
- once a request reaches 200 thousand tokens, the whole request pays the higher rates
- Batch API
- not supported
- Context window
- 500 thousand tokens
- Max output
- no text output limit, per the release notes
- Knowledge
- date not published by xAI
- Reasoning
- always on; low, medium, high (default) or xhigh
- Input and output
- text and image in, text out
- Rate limits
- 150 requests per second and 50 million tokens per minute
- Where it runs
- xAI API, including the US endpoint (10% more expensive); dedicated announcements for Amazon Bedrock, Google Cloud's Gemini Enterprise Agent Platform, Microsoft Foundry and GitHub Copilot; at launch, also Cursor, Grok Build, OpenRouter, Vercel and Cloudflare
Where Grok 4.6 does well, where it falls behind and who it is for
Where it does well
- Office work: 1,753 points on GDPVal-AA v2 in xAI's launch table, ahead of Claude Fable 5 (1,741) and GPT-5.6 Sol (1,728).
- Legal work: 15.8% on Harvey LAB, against 11.3% for Fable 5 and 2.5% for GPT-5.6 Sol, in the same table.
- Output price: US$ 6 per million, against US$ 10 for Claude Sonnet 5.5 and GPT-6.1 Sol.
- More announced platforms: xAI published 4.6 launches on GitHub Copilot (Aug 14), Amazon Bedrock (Aug 19), Google Cloud (Aug 21) and Microsoft Foundry (Aug 26). There is no equivalent announcement for 4.7.
- Jump over Grok 4.5: DeepSWE from 54% to 65.9% and AA-Briefcase from 1,313 to 1,577, at the same input and output price; only the cache went up, from US$ 0.30 to US$ 0.50.
Where it falls behind
- Loses to a successor at the same price: in the Grok 4.7 table it trails on all seven tests; on Terminal-Bench 4.0 it scores 20.3% against 37.6%.
- Terminal and coding: 26% on Terminal-Bench v3.0, against 34.6% for GPT-5.6 Sol and 34.1% for Fable 5; 65.9% on DeepSWE, against 73% for GPT-5.6 Sol.
- Behind Fable 5 in its own table: the Anthropic model leads 5 of the 10 rows, including CursorBench v3.2 (70.5% against 69.9%) and FrontierCode (63.6% against 61.3%).
- Context and cache: a 500 thousand token window, half of Sonnet 5.5's; the price doubles from 200 thousand, and cache reads cost US$ 0.50, against US$ 0.20 for Sonnet 5.5.
- Shrinking documentation: the 4.6 guide now redirects to the 4.7 one, the knowledge cutoff was never published, and the documentation only offers reasoning summaries for 4.7.
Who it makes sense for
- Companies using Grok through Amazon Bedrock, Google Cloud, Microsoft Foundry or GitHub Copilot that have not yet found 4.7 in their platform's catalog.
- Workflows already validated in production on 4.6, while 4.7 is tested side by side on the same tasks.
- Teams that need processing in the US: the regional endpoint only serves 4.7 and 4.6, at a 10% premium.
- Legal and office work on a platform where only 4.6 is available: those were the tests it led at launch.
When to pick another model
- New projects on the xAI API: 4.7 costs the same, runs at the same speed according to xAI, and is the model the documentation recommends.
- Agents that work for hours in the terminal: 20.3% on Terminal-Bench 4.0, just over half of 4.7's 37.6%.
- Requests above 200 thousand tokens, such as a full contract or a large repository: the price doubles for the whole request and the window ends at 500 thousand.
- Systems that cannot tolerate an unannounced model swap: when xAI retired older models on May 15, 2026, requests started landing on another model, billed at that model's price.
What xAI published at launch
Grok 4.6 came out on Aug 12, 2026 with a focus, in xAI's words, on long-running agents and visual and interactive work. The company says it had a longer supplemental training run than Grok 4.5, with reinforcement learning on knowledge work, coding and tasks such as kernel optimization and web development.
The launch table puts 4.6 at high effort against GPT-5.6 Sol and Claude Fable 5 at max. It leads 3 of the 10 rows: GDPVal-AA v2, AA-Briefcase and Harvey LAB. Fable 5 leads 5, and GPT-5.6 Sol leads 2. Competitor figures, xAI notes, are the best of what those makers published and public leaderboards.
| Grok 4.6 high | Grok 4.5 high | GPT-5.6 Sol max | Fable 5 max | |
|---|---|---|---|---|
| AA Intelligence Index composite index |
61 | 56 | 61 | 62 |
| GDPVal-AA v2 office work, Elo |
1,753 | 1,526 | 1,728 | 1,741 |
| CursorBench v3.2 coding |
69.9% | 66.7% | 67.2% | 70.5% |
| DeepSWE v1.1 software engineering |
65.9% | 54% | 73% | 70% |
| FrontierCode v1.1 Extended coding |
61.3% | 56.6% | 60.6% | 63.6% |
| APEX-Agents agents |
57.5% | 47.1% | 56.7% | 59.2% |
| Terminal-Bench v3.0 terminal work |
26% | 15.7% | 34.6% | 34.1% |
| APEX-SWE software engineering |
56.4% | 53.6% | not measured | 58.8% |
| AA-Briefcase multi-hour office work, Elo |
1,577 | 1,313 | 1,502 | 1,574 |
| Harvey LAB (Vals) legal work |
15.8% | 12.9% | 2.5% | 11.3% |
Figures from the Grok 4.6 announcement. The dot marks the best in each row, matching xAI's bold. "Not measured" is the cell where xAI published no score for GPT-5.6 Sol.
Grok 4.6 or Grok 4.7: what changes
Grok 4.7 came out on Sep 21, 2026 at the same price and, according to xAI, at the same speed, starting from a new, larger base model. In its launch table, it beats 4.6 on all seven tests, with the widest gap on long terminal work.
Two caveats before comparing. The columns use different effort levels: 4.7 appears at xhigh and 4.6 at high, and the 4.7 DeepSWE score was measured at high. And xAI itself changed the 4.6 figures from one announcement to the next: DeepSWE went from 65.9% to 65.2%, and AA-Briefcase, which was 1,577, became 1,546 on version 1.1 of the test.
| Grok 4.6 high | Grok 4.7 xhigh | |
|---|---|---|
| Terminal-Bench 4.0 multi-hour terminal work |
20.3% | 37.6% |
| EEBench electrical engineering |
53.0% | 64.0% |
| CursorBench 4.0 coding |
40.4% | 46.3% |
| DeepSWE v1.1 software engineering |
65.2% | 71.0%* |
| AA Briefcase v1.1 multi-hour office work, Elo |
1,546 | 1,657 |
| Harvey Legal Agent Benchmark legal work |
15.8% | 19.6% |
| HealthBench Professional clinical reasoning |
48.5% | 56.7% |
| GDPval office work, Elo |
1,605 | 1,695 |
Main table and professional work chart from the Grok 4.7 announcement. * The Grok 4.7 DeepSWE score was measured at high effort. The chart names the last test only GDPval, with no version.
Is Grok 4.6 still worth using?
For a new project, no. The price per token is identical, xAI recommends 4.7 for everything that is not image, video or voice, and the only table comparing the two shows 4.7 ahead on every row.
4.6 is still on the API, in the price table and on the US endpoint, and xAI has not published a retirement date. But there is a sign of end of cycle: the model guide in the documentation already redirects to the 4.7 one. And there is a precedent: in the May 15, 2026 retirement, the old identifiers kept answering, but were served by another model and billed at its price. Anyone who depends on 4.6 should keep an eye on the release notes.
- Stay on 4.6 if you use Grok through a cloud or GitHub Copilot and 4.7 does not show up there yet, or if you have a validated workflow and have not tested 4.7.
- Switch to 4.7 for any direct use of the xAI API: change the
modelfield togrok-4.7, and the price stays the same. - Check when switching that, on the Responses API, 4.7 always returns encrypted reasoning (
reasoning.encrypted_content). xAI says clients that ignore the field keep working.
| Grok 4.6 xAI | Grok 4.7 xAI | Sonnet 5.5 Anthropic | Muse Spark 1.3 Meta | |
|---|---|---|---|---|
| Input | US$ 2 | US$ 2 | US$ 2 | US$ 1.25 |
| Output | US$ 6 | US$ 6 | US$ 10 | US$ 4.25 |
| Cache read | US$ 0.50 | US$ 0.50 | US$ 0.20 | US$ 0.15 |
| Context window | 500 thousand | 500 thousand | 1 million | 1,048,576 |
| Long context | doubles from 200 thousand | doubles from 200 thousand | same price | same price |
xAI, Anthropic and Meta price tables, checked on Oct 1, 2026.
Examples of using Grok 4.6
A team that standardized on GitHub Copilot picks Grok 4.6 in the model picker for multi-step tasks in the repository. In some companies, an administrator has to enable the model in Copilot settings first.
A company that keeps its data on AWS uses 4.6 through Amazon Bedrock to build reports and spreadsheets from internal documents. It is the kind of deliverable GDPVal-AA v2 measures, where it scored 1,753 at launch.
Since both cost the same per token, you can run the same tasks on 4.6 and 4.7 for a week, compare accuracy and tokens spent, and only then switch the identifier in production.
Where Grok 4.6 stands out and where it falls behind, by the numbers
Grok 4.6 has published results on 28 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Nov 15, 2025 and Nov 10, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 28 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 94.0% | #11 / 79 (tied with 1) | #11 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 99.2% | #15 / 81 (tied with 1) | #15 / 168 | 100.0% GPT-5.5 |
| Dtbench | 97.3% | #5 / 68 (tied with 2) | #5 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 48.45 | #25 / 68 | #25 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 19.7% | #26 / 78 | #26 / 112 | 32.3% GPT-5.6 Sol |
| WeirdML | 67.3% | #21 / 52 | #21 / 112 | 93.6% GPT-6 Astra |
| SciCode | 56.5% | #21 / 75 (tied with 1) | #21 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 66.0% | #32 / 68 (tied with 1) | #32 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 67.1% | #22 / 49 | #22 / 83 | 95.0% GPT-6 Astra |
| ALE-Bench | 1,508 | #15 / 62 | #15 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,618 | #15 / 73 | #15 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 87.5% | #31 / 49 (tied with 1) | #31 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 49.3% | #27 / 63 | #29 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 33.0% | #39 / 63 (tied with 1) | #43 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 34.0% | #22 / 61 (tied with 2) | #22 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 51.0% | #28 / 70 | #28 / 74 | 100.0% Claude Fable 5.1 |
| SimpleBench | 75.9% | #7 / 36 | #7 / 74 | 81.9% Claude Fable 5 |
| Frontiermath Tier 4 V2 | 31.7% | #27 / 58 (tied with 2) | #27 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 9,047 | #10 / 48 | #10 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 65.3% | #7 / 49 | #7 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 48.0% | #10 / 43 | #10 / 43 | 54.6% Claude Opus 5.5 |
| Gdp Pdf | 17.2% | #26 / 37 | #26 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 40.0% | #17 / 34 | #17 / 34 | 83.3% Claude Opus 5.5 |
| Blueprint Bench 2 | 33.2% | #11 / 30 | #11 / 31 | 54.4% Gemini 4 Argon |
| Ebr Bench | 30.5% | #13 / 24 | #13 / 27 | 76.2% GPT-6 Astra |
| DeepSWE | 67.5% | #9 / 26 | #9 / 26 | 74.1% GPT-6 Astra |
| FrontierSWE | 25.3% | #14 / 20 | #14 / 20 | 65.5% GPT-6 Astra |
| CursorBench | 41.4% | #11 / 16 | #11 / 16 | 57.8% Claude Opus 5.5 |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare Grok 4.6 with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| Muse Spark 1.3 | Meta AI | 156.76 | 1,528 | 81.6 | Sep 2, 2026 |
| Claude Sonnet 5 | Anthropic | 156.36 | 1,491 | 76.0 | Jun 30, 2026 |
| GPT-5.6 Luna | OpenAI | 156.33 | 1,478 | 73.6 | Jul 9, 2026 |
| Claude Opus 4.7 | Anthropic | 156.32 | 1,532 | 76.5 | Apr 16, 2026 |
| Grok 4.6 | xAI | 156.31 | 1,475 | 78.0 | Aug 12, 2026 |
| GPT-6 Luna | OpenAI | 156.31 | n/a | 72.0 | Sep 22, 2026 |
| GLM-5.3 | Z.ai (Zhipu AI) | 155.91 | n/a | 76.1 | Aug 14, 2026 |
Grok 4.6 or GPT-6 Luna?
GPT-6 Luna is the model from another lab closest to Grok 4.6 in capability. Across the 20 tests both took, this is the score:
| Test | Grok 4.6 | GPT-6 Luna |
|---|---|---|
| Game puzzles | 34.0% | 7.0% |
| Frontiermath Tier 4 V2 | 31.7% | 56.1% |
| Apex Agents | 65.3% | 44.3% |
| Gdp Pdf | 17.2% | 23.0% |
| ProofBench | 51.0% | 64.0% |
| Frontiermath Tiers 1 3 V2 | 66.0% | 78.9% |
| SimpleQA Verified | 49.3% | 41.4% |
| FrontierCode | 48.0% | 42.4% |
| ARC-AGI 2 | 67.1% | 59.3% |
| Furniture Assembly | 40.0% | 44.2% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Grok 4.6 costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
CursorBench
The score scale does not start at zero.
DeepSWE
The score scale does not start at zero.
Other models from xAI
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| Grok 4.7 | Sep 21, 2026 | 154.37 | 1,453 | 77.4 | API access |
| Grok 4.6 | Aug 12, 2026 | 156.31 | 1,475 | 78.0 | API access |
| Grok 4.5 | Jul 8, 2026 | 153.93 | 1,495 | 75.8 | API access |
| Grok 4.3 (no thinking) | Apr 17, 2026 | 149.21 | 1,425 | 62.3 | Hosted access (no API) |
| Grok 4.3 Beta | Apr 17, 2026 | 148.95 | n/a | n/a | Hosted access (no API) |
| Grok 4.20 | Feb 17, 2026 | 152.08 | n/a | n/a | API access |
| Grok 4 Fast | Sep 19, 2025 | 144.19 | 1,437 | n/a | API access |
| Grok 4 | Jul 9, 2025 | 146.43 | n/a | n/a | API access |
| Grok-3 mini | Jun 24, 2025 | 140.33 | 1,394 | n/a | API access |
Have you used Grok 4.6? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about Grok 4.6
How much does Grok 4.6 cost?
US$ 2 per million input tokens, US$ 0.50 for cached input and US$ 6 for output, on requests under 200 thousand tokens. From 200 thousand on, the whole request moves to US$ 4, US$ 1 and US$ 12. It is the same price as Grok 4.7. There is no batch discount, and the US endpoint charges 10% more.
Will Grok 4.6 be retired?
xAI has not published a date. As of Oct 1, 2026 it is still on the API, in the price table and on the US endpoint. The model guide in the documentation, however, already redirects to the Grok 4.7 one. In xAI's last retirement, on May 15, 2026, the old identifiers started being served by another model, billed at that model's price.
Should I switch from Grok 4.6 to Grok 4.7?
On the xAI API, yes: the price is the same and 4.7 beats 4.6 on all seven tests in xAI's table, as on Terminal-Bench 4.0 (37.6% against 20.3%). 4.6 only makes sense where 4.7 is not yet available, such as the platforms where xAI announced only 4.6.
Is Grok 4.6 better than GPT-5.6 Sol and Claude Fable 5?
In xAI's launch table, it wins 3 of 10 tests: office work, multi-hour office work and legal work. Of the other 7, Fable 5 leads 5 and GPT-5.6 Sol leads 2. The independent scores are at the top of this page.
Where can I use Grok 4.6?
On the xAI API (identifier grok-4.6), including the US endpoint. xAI also announced 4.6 on Amazon Bedrock, Google Cloud, Microsoft Foundry and GitHub Copilot, and at launch it was also in Cursor and Grok Build.
Where does Grok 4.6 rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Grok 4.6 ranks #27 out of 267 models, with an ECI of 156.31. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is Grok 4.6 better than GPT-6 Luna?
Across the 20 tests both models took and Epoch AI compiled, Grok 4.6 comes out ahead on 14, behind on 6 and tied on 0. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is Grok 4.6 best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Grok 4.6 are: Dtbench (#5 of 68); Apex Agents (#7 of 49); GPQA Diamond (#11 of 79). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- xAI, Grok 4.6 announcement (evals table)
- xAI, Grok 4.7 announcement (table with the Grok 4.6 column)
- xAI, grok-4.6 model page (documentation)
- xAI, models and usage recommendation (documentation)
- xAI, pricing (documentation)
- xAI, release notes (documentation)
- xAI, reasoning (documentation)
- xAI, US regional endpoint (documentation)
- xAI, May 15, 2026 model retirement (documentation)
- xAI, Grok 4.6 in GitHub Copilot
- xAI, Grok 4.6 on Amazon Bedrock
- xAI, Grok 4.6 on Gemini Enterprise Agent Platform
- xAI, Grok 4.6 on Microsoft Foundry
- Anthropic, pricing (Sonnet 5.5)
- OpenAI, API pricing (GPT-6.1 Sol)
- Meta, Meta Model API pricing (Muse Spark 1.3)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
Read also: Grok 4.7: the cheapest model at the table wins two tests out of seven
← See the full ranking