Grok 4.7
Grok 4.7 is the main model from xAI, which now signs as SpaceXAI, released on Sep 21, 2026. It costs US$ 2 per million input tokens and US$ 6 per million output for requests under 200 thousand tokens; from there on, the whole request moves to US$ 4 and US$ 12. In xAI's own table, it leads on legal work and electrical engineering and trails Claude Fable 5.1 on coding and long terminal work. Its price per token is the lowest in the comparison, but the advantage shrinks once you look at cost per task.
Grok 4.7 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-4.7- Release
- 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, and the price table only lists input, cached input and output
- Long context
- once a request reaches 200 thousand tokens, the whole request pays double
- Batch API
- not supported
- Fast version
- double the price, only in Cursor and Grok Build
- Context window
- 500 thousand tokens
- Max output
- no text output limit, per the documentation
- Knowledge
- through May 2026
- Reasoning
- always on; low, medium, high (default) or xhigh, no max level
- Input and output
- text and images in, text out
- Where it runs
- xAI API (Responses and Chat Completions), Cursor on all plans, Grok Build (default model), OpenRouter, Vercel and Cloudflare
Where Grok 4.7 does well, where it falls behind and who it is for
Where it does well
- Legal work: 19.6% on the Harvey Legal Agent Benchmark in xAI's table, against 6.7% for Fable 5.1 and 2.5% for GPT-5.6 Sol.
- Electrical engineering: 64.0% on EEBench, against 56.4% for Fable 5.1 and 39.4% for GPT-5.6 Sol.
- Output price: US$ 6 per million, against US$ 10 for Sonnet 5.5 and GPT-6.1 Sol and US$ 50 for Fable 5.1.
- Low-cost band: on CursorBench 4.0, it scores 43.9% for US$ 4.69 per task at default effort, the best score in xAI's chart under US$ 5.
- A jump over 4.6 at the same price: Terminal-Bench 4.0 from 20.3% to 37.6% and HealthBench Professional from 48.5% to 56.7%.
Where it falls behind
- Long terminal work: 37.6% on Terminal-Bench 4.0, against the 57.9% xAI itself gives Fable 5.1.
- Coding: 46.3% on CursorBench 4.0, against 51.8% for Fable 5.1; on DeepSWE, 71.0% (measured at high), behind GPT-5.6 Sol's 72.7%.
- Low ceiling per task: its best on CursorBench 4.0 is 46.3%, for US$ 6.01; Fable 5.1 at low effort already scores 45.1% for US$ 5.44.
- Cache and context: cache reads at US$ 0.50, against US$ 0.20 for Sonnet 5.5, and a 500 thousand token window, half of Sonnet 5.5's.
- No batch, no turning reasoning off: there is no Batch API, the minimum effort is low, and
logprobsis silently ignored.
Who it makes sense for
- Legal and electrical engineering teams testing analysis agents, the two areas where it leads xAI's table.
- Volume work with long answers and short requests, under 200 thousand tokens, where US$ 6 output weighs on the bill.
- Anyone coding in Cursor or Grok Build: it is on all Cursor plans and is the default model in Grok Build.
- Apps that need to read what is being said on X, with X search as a server-side tool.
When to pick another model
- Agents that work for hours in the terminal: Fable 5.1 scores 57.9% on Terminal-Bench 4.0 in xAI's own table.
- Requests above 200 thousand tokens, such as a whole contract or a large repository: the price doubles for the entire request and the window ends at 500 thousand.
- Discounted batch processing: Grok 4.7 does not accept the Batch API.
- Apps with a long repeated instruction: rereading the cache costs more than on Sonnet 5.5 and GPT-6.1 Sol.
What xAI published at launch
Grok 4.7 shipped on Sep 21, 2026, at the same price and speed as Grok 4.6, according to xAI. The company says it starts from a new, larger base model and a longer reinforcement learning run, weighted toward tasks that take many hours to solve.
The announcement table compares Grok 4.7 at xhigh effort with GPT-5.6 Sol and Claude Fable 5.1 at max. It wins two of the seven tests and loses four to Fable 5.1. Two caveats: the 71.0% on DeepSWE was measured at high, and the figures xAI gives Fable 5.1 do not match Anthropic's (57.9% against 55.8% on Terminal-Bench 4.0, 70.0% against 67.4% on DeepSWE, 6.7% against 19.09% on Harvey).
| Grok 4.7 xhigh | Grok 4.6 high | GPT-5.6 Sol max | Fable 5.1 max | |
|---|---|---|---|---|
| Harvey Legal Agent Benchmark legal work |
19.6% | 15.8% | 2.5% | 6.7% |
| EEBench electrical engineering |
64.0% | 53.0% | 39.4% | 56.4% |
| DeepSWE v1.1 software engineering |
71.0%* | 65.2% | 72.7% | 70.0% |
| CursorBench 4.0 software engineering |
46.3% | 40.4% | 41.7% | 51.8% |
| AA Briefcase v1.1 multi-hour office work, Elo |
1,657 | 1,546 | 1,487 | 1,678 |
| Terminal-Bench 4.0 multi-hour terminal work |
37.6% | 20.3% | 37.3% | 57.9% |
| HealthBench Professional clinical reasoning |
56.7% | 48.5% | 60.5% | 62.1% |
Figures from the Grok 4.7 announcement, with each column's effort as stated by xAI. The dot marks the best in each row. * Grok 4.7's DeepSWE score was measured at high effort.
What was left out of the main table
The main table compares against GPT-5.6 Sol, but OpenAI had already launched GPT-6 Astra on Sep 3, 2026. It only appears in the professional work charts of the same announcement, and there one of the wins changes hands: on EEBench, GPT-6 Astra scores 69.3, above Grok 4.7's 64.
Opus 5.5, GPT-6 Sol and GPT-6.1 Sol came out after Grok 4.7 and are not in any xAI chart.
| Grok 4.7 xhigh | Grok 4.6 high | GPT-6 Astra max | Fable 5.1 max | |
|---|---|---|---|---|
| GDPval office work, Elo |
1,695 | 1,605 | 1,542 | 1,735 |
| AA Briefcase multi-hour office work, Elo |
1,657 | 1,546 | 1,569 | 1,678 |
| EEBench electrical engineering |
64.0% | 53.0% | 69.3% | 56.4% |
Tabbed charts from the Grok 4.7 announcement. The chart calls the first test just GDPval, with no version.
Cost per task: where the low price shrinks
For CursorBench 4.0, xAI published the score and average cost per task at each effort level and says Grok 4.7 is at the frontier of price and performance. In the low-cost band, the chart bears that out: at default effort it scores 43.9% for US$ 4.69.
Above that, the math changes. Grok 4.7 tops out at 46.3% at xhigh, for US$ 6.01, with 70,141 output tokens per task. Fable 5.1, whose output costs more than eight times as much per token, scores 45.1% at low for US$ 5.44, with about half the tokens, and passes Grok from medium up.
Opus 5.5 came out a day later and is not in that chart. In Anthropic's chart for the same test, it scores 56.0% for US$ 3.97 per task at high effort. These are charts from different makers, but the Fable 5.1 points match in both.
| Score | Cost per task | Output tokens | |
|---|---|---|---|
| Grok 4.7 low |
33.1% | US$ 1.58 | 15,677 |
| Grok 4.7 medium |
41.6% | US$ 3.49 | 36,683 |
| Grok 4.7 high (default) |
43.9% | US$ 4.69 | 56,382 |
| Grok 4.7 xhigh |
46.3% | US$ 6.01 | 70,141 |
| Fable 5.1 low |
45.1% | US$ 5.44 | 34,795 |
| Fable 5.1 medium |
46.8% | US$ 7.05 | 45,411 |
| Opus 5 high (default) |
44.7% | US$ 9.00 | 61,405 |
| GPT-5.6 Sol max |
41.7% | US$ 8.23 | 42,944 |
CursorBench 4.0 score-versus-cost chart from the Grok 4.7 announcement: average cost per task in dollars and output tokens per task, at list price.
The price and the fine print
The US$ 2 and US$ 6 apply to requests under 200 thousand tokens. Once a request reaches 200 thousand, the entire request moves to US$ 4 input, US$ 1 cached input and US$ 12 output, and cached tokens count toward that threshold. Reasoning is billed as output.
- Cache with no guarantee: without
prompt_cache_key(Responses) or thex-grok-conv-idheader (Chat Completions), xAI warns that you often pay the full input price. - Fast and region: the Fast version costs double and only exists in Cursor and Grok Build; the US regional endpoint charges 10% more.
- Encrypted reasoning: the Responses API always returns
reasoning.encrypted_content, which the client must send back unchanged on the next turn. - Violation fee: on the Responses API, a request blocked by the usage policy before generation pays US$ 0.05.
| Grok 4.7 xAI | Sonnet 5.5 Anthropic | GPT-6.1 Sol OpenAI | Opus 5.5 Anthropic | |
|---|---|---|---|---|
| Input | US$ 2 | US$ 2 | US$ 2 | US$ 4 |
| Output | US$ 6 | US$ 10 | US$ 10 | US$ 20 |
| Cache read | US$ 0.50 | US$ 0.20 | US$ 0.10 | US$ 0.20 |
| Context window | 500 thousand | 1 million | 1.05 million | 1 million |
| Long context | everything doubles from 200 thousand | same price | 2x input and cache, 1.5x output above 272 thousand | same price |
xAI, Anthropic and OpenAI pricing pages, checked on Oct 1, 2026.
Examples of using Grok 4.7
An agent reads the case documents, builds the timeline and points to precedents for the lawyer to check. It is the kind of task the Harvey Legal Agent Benchmark measures, where Grok 4.7 leads xAI's table with 19.6%, a score that still requires reviewing everything.
A routine writes the documentation for every module of a system, with short requests and long answers. Under 200 thousand tokens per request, output is the line that weighs most, and that is where Grok 4.7 charges less.
An assistant follows what is being said about a brand on X with the built-in search and summarizes it for the team. The tool costs US$ 5 per thousand posts read, on top of tokens.
Where Grok 4.7 stands out and where it falls behind, by the numbers
Grok 4.7 has published results on 24 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Dec 25, 2025 and Dec 20, 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 24 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 92.7% | #21 / 73 | #21 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 98.1% | #26 / 75 (tied with 1) | #26 / 168 | 100.0% GPT-5.5 |
| Dtbench | 96.0% | #12 / 64 (tied with 3) | #12 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 49.38 | #23 / 64 | #23 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 18.0% | #30 / 76 (tied with 1) | #30 / 112 | 32.3% GPT-5.6 Sol |
| SciCode | 57.8% | #15 / 74 (tied with 1) | #15 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 53.0% | #45 / 64 | #49 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 61.4% | #25 / 43 (tied with 1) | #25 / 83 | 95.0% GPT-6 Astra |
| WebDev Arena | 1,638 | #12 / 67 | #12 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 90.2% | #26 / 43 | #27 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 56.0% | #19 / 59 | #20 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 41.0% | #28 / 58 (tied with 1) | #32 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 29.0% | #30 / 58 | #30 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 34.0% | #40 / 64 | #41 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 17.1% | #43 / 54 (tied with 3) | #47 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 10,537 | #6 / 43 | #6 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 54.6% | #20 / 48 | #20 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 47.6% | #11 / 43 | #11 / 43 | 54.6% Claude Opus 5.5 |
| Gdp Pdf | 22.8% | #20 / 36 | #20 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 20.8% | #30 / 32 (tied with 1) | #32 / 34 | 83.3% Claude Opus 5.5 |
| Blueprint Bench 2 | 32.5% | #13 / 29 | #13 / 31 | 54.4% Gemini 4 Argon |
| FrontierSWE | 29.5% | #11 / 20 | #11 / 20 | 65.5% GPT-6 Astra |
| CursorBench | 46.3% | #7 / 16 | #7 / 16 | 57.8% Claude Opus 5.5 |
| Weirdml V3 | 8.6% | #7 / 10 | #7 / 11 | 42.2% GPT-6 Astra |
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 4.7 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 1.2 | Meta AI | 154.90 | 1,509 | 78.0 | Aug 5, 2026 |
| DeepSeek V4 Flash 0731 | DeepSeek | 154.49 | n/a | 74.2 | Jul 31, 2026 |
| Gemini 3.5 Flash | Google DeepMind | 154.38 | 1,526 | 74.6 | May 19, 2026 |
| Grok 4.7 | xAI | 154.37 | 1,453 | 77.4 | Sep 21, 2026 |
| Gemini 3.6 Flash | Google DeepMind | 154.24 | 1,530 | 73.6 | Jul 21, 2026 |
| Muse Spark 1.1 | Meta AI | 154.00 | 1,516 | 75.3 | Jul 9, 2026 |
| Qwen3.7-Max | Alibaba | 153.83 | 1,530 | 73.1 | May 19, 2026 |
Grok 4.7 or Gemini 3.5 Flash?
Gemini 3.5 Flash is the model from another lab closest to Grok 4.7 in capability. Across the 19 tests both took, this is the score:
| Test | Grok 4.7 | Gemini 3.5 Flash |
|---|---|---|
| Apex Agents | 54.6% | 27.5% |
| Vending Bench 2 | 10,537 | 5,396 |
| Gdp Pdf | 22.8% | 14.0% |
| Frontiermath Tier 4 V2 | 17.1% | 26.8% |
| CritPt | 18.0% | 13.1% |
| Frontiermath Tiers 1 3 V2 | 53.0% | 62.8% |
| SimpleQA Verified | 56.0% | 66.2% |
| ARC-AGI 2 | 61.4% | 72.1% |
| Game puzzles | 29.0% | 32.0% |
| Chess puzzles | 41.0% | 45.0% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Grok 4.7 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.
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 4.7? 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 4.7
How much does Grok 4.7 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. The Fast version costs double and only exists in Cursor and Grok Build. There is no batch discount.
Is Grok 4.7 better than Claude and GPT?
In xAI's table, it wins two of seven tests (legal work and electrical engineering), loses four to Claude Fable 5.1, including coding and terminal work, and one to GPT-5.6 Sol. In the electrical engineering chart, GPT-6 Astra is also ahead of it. The independent scores are at the top of this page.
Is Grok 4.7 really cheaper?
Per token, yes: US$ 6 output against US$ 10 for Sonnet 5.5 and GPT-6.1 Sol. Per task, it depends on effort. On CursorBench 4.0, it scores 46.3% for US$ 6.01 at xhigh, while Fable 5.1 scores 45.1% for US$ 5.44 at low, using about half the output tokens. And above 200 thousand tokens per request, its price doubles.
Where can I use Grok 4.7?
On the xAI API (identifier grok-4.7), through the Responses and Chat Completions APIs, in Cursor on all plans, in Grok Build, where it is the default model, and through the OpenRouter, Vercel and Cloudflare gateways. xAI notes that access may vary by region and account type.
Where does Grok 4.7 rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Grok 4.7 ranks #38 out of 267 models, with an ECI of 154.37. 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 4.7 better than Gemini 3.5 Flash?
Across the 19 tests both models took and Epoch AI compiled, Grok 4.7 comes out ahead on 10, behind on 9 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 4.7 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 4.7 are: Vending Bench 2 (#6 of 43); WebDev Arena (#12 of 67); Dtbench (#12 of 64). 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 4.7 announcement (table and charts)
- xAI, Grok 4.7 guide (documentation)
- xAI, grok-4.7 model page (documentation)
- xAI, pricing (documentation)
- xAI, cache pricing and long context (documentation)
- xAI, reasoning (documentation)
- Anthropic, Claude Fable 5.1 announcement (Anthropic's figures for Fable 5.1)
- Anthropic, Claude Fable 5.1 system card (DeepSWE and Harvey)
- Anthropic, Claude Sonnet 5.5 announcement (Opus 5.5 cost per task on CursorBench 4.0)
- Anthropic, pricing (documentation)
- OpenAI, API pricing (GPT-6.1 Sol)
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