Grok 4.7: 38% on Terminal-Bench in xAI's harness, 26% in an independent one, and on by default in Copilot

The same benchmark, twelve points apart

xAI's model card for Grok 4.7 reports 38.0% on Terminal-Bench 4.0 at its highest reasoning effort. Artificial Analysis, running the same 66 task benchmark against the same effort setting, records 25.8%. Applied to 66 tasks that is roughly 25 tasks solved against 17, a difference of about eight tasks. Both figures were published on launch day, 21 September 2026, and on that same day GitHub made the model generally available in Copilot, where a model nobody has configured is switched on by a policy rather than by a person.

What the gap does not establish: that either number is wrong, or that anyone inflated anything. xAI's card names the harness behind its figure and warns in the same paragraph that absolute scores "remain sensitive to the agent harness". Artificial Analysis publishes its harness too. These are two measurements of the same model taken with different equipment, and the difference belongs to the equipment at least as much as to the model. The question for a UK security or platform lead is narrower: which of those two setups resembles the way your developers will actually run the model, and did anyone in your organisation decide they could?

What xAI claims and what the independent record shows

The launch post is short. It carries a benchmark table, two price lines, a subtitle that promises "Twice as fast, at half the price of comparable models", and a body sentence that the model is "Served at the same price and speed as Grok 4.6". The 30 page model card is longer and more careful. Artificial Analysis publishes a third view, built from its own runs against the first party API. The table below sets the claims against that independent record, fetched on 22 September 2026, the day after launch.

xAI's launch claims against the Artificial Analysis record for Grok 4.7 (xhigh), fetched 22 September 2026. Sources: x.ai/news/grok-4-7, the Grok 4.7 model card, artificialanalysis.ai/models/grok-4-7.

Claim	Independent record	What the record does not establish
Terminal-Bench 4.0: 38.0% at xhigh	25.8% at xhigh, 24.7% at high	Which of the two harnesses predicts your own results
"Twice as fast, at half the price of comparable models"	39.2 output tokens per second, ranked 152 of 206 models. GPT-6 Astra (max) 60.7, Claude Fable 5.1 (max) 65.7, GPT-5.6 Sol (max) 82.0	What the speed claim is measured against: the page does not say
"Served at the same price and speed as Grok 4.6"	Price identical at $2 and $6 per million tokens. Speed: Grok 4.6 (high) 58.6 tokens per second, against 39.2 for Grok 4.7 (xhigh) and 46.5 at high	Whether the measured speed gap survives launch week capacity
Card: reaches results with "fewer steps and fewer output tokens than other frontier models"	240M output tokens across the Intelligence Index against a median of 92M, described as "very verbose"	Token use on CursorBench 4.0, where xAI's own chart measures it
GDPval Elo 1,695 at xhigh	GDPval-AA 1,695.21, the same number	Anything beyond this one agreed figure

That last row matters for fairness. Where xAI quotes GDPval it is quoting Artificial Analysis runs, and the numbers match to the decimal: 1,695 against 1,695.21 for Grok 4.7, 1,735 against 1,734.67 for Claude Fable 5.1, 1,542 against 1,541.89 for GPT-6 Astra. This is not a vendor that ignores the independent scoreboard. On one benchmark, the one it ran itself in its own agent, its number and the independent number come apart.

On the headline aggregate, Artificial Analysis puts Grok 4.7 (xhigh) at 46.4 on its Intelligence Index v4.3.2, ranked 20th of 206 models in its class, against 52.7 for GPT-6 Astra (max) and 53.4 for Claude Fable 5.1 (max). Grok 4.6 (high) scored 44.3, so the independent gain generation on generation is about two points. Terminal-Bench 4.0 is one of the ten evaluations inside that index, so the lower terminal score is part of the 46.

The price claim holds per token and gets more complicated per task. Grok 4.7 lists at $2 input and $6 output per million tokens, against $4 and $20 for GPT-5.6 Sol and $10 and $50 for both GPT-6 Astra and Claude Fable 5.1. On the Artificial Analysis cost per Intelligence Index task, which prices the tokens each model actually consumed, Grok 4.7 (xhigh) cost $3.74 per task, against $7.63 for Claude Fable 5.1 (max), $3.26 for GPT-6 Astra (max) and $1.99 for GPT-5.6 Sol (max). Against Anthropic's model the "half the price" line holds on that measure, at 49% of the cost. Against OpenAI's two it does not: Grok 4.7 cost 15% more per task than GPT-6 Astra and 88% more than GPT-5.6 Sol, because verbosity eats the cheaper rate. The same arithmetic runs through time: Artificial Analysis estimates 1,126 seconds per index task for Grok 4.7 against 449 for GPT-6 Astra, 751 for Claude Fable 5.1 and 353 for GPT-5.6 Sol.

Why one benchmark gives two answers

Three things routinely move a benchmark score without anyone behaving badly, and only one of them is in play here.

• The harness. xAI evaluated Grok 4.7 on Terminal-Bench 4.0 "with the Grok Build harness", with the runs "conducted by Harbor". Artificial Analysis uses the mini-swe-agent harness, over the full 66 task set, with pass at one scoring averaged over three repeats, a cap of 500 agent steps, and, in its own words, "no context compaction or summarization: the agent always sees its full transcript". Grok Build is xAI's own coding agent and the card says Grok 4.7 was trained to understand its own harness natively. A model tuned to its house agent, measured in that agent, is measured at its best. That is a reasonable thing for a vendor to publish and a poor thing for a buyer to assume.
• The effort setting. Not the explanation here. Both figures are at xhigh. Artificial Analysis also publishes the high variant at 24.7%, so dialling effort down costs about one point, not twelve.
• The version and the date. Also not the explanation. Both are Terminal-Bench 4.0, both use the 66 task set described in the card, and both were published within a day of each other.

Terminal-Bench 4.0, as reported in the Grok 4.7 model card and as measured by Artificial Analysis. Differences are our arithmetic.

Model and effort	xAI model card	Artificial Analysis	Difference
Grok 4.7 (xhigh)	38.0%	25.8%	12.2 points lower
Grok 4.6 (high)	20.3%	21.2%	0.9 points higher
GPT-5.6 Sol (max)	37.3%	39.9%	2.6 points higher
Claude Fable 5.1 (max)	57.9%	52.0%	5.9 points lower

Read the column of differences rather than any single row. For three of the four models the two scorekeepers land within six points of each other, and they disagree in both directions. The figure xAI publishes for Anthropic's Claude Fable 5.1 is nearly six points higher than the independent one, which is the opposite of the pattern a reader looking for bad faith would expect. Only Grok 4.7 moves by twelve. The card does not state which harness produced the competitor scores, so the honest reading is that a house harness score and a generic harness score are different quantities, and the difference is largest for the model whose house harness it is.

One plausible mechanism, and this is inference rather than something either source establishes: a generic harness that never compacts context and stops at 500 steps will punish a verbose model on long tasks, and Grok 4.7 is measurably verbose. Neither xAI nor Artificial Analysis says this is the cause, and nobody has published a run of Grok 4.7 in a third harness that would settle it. If your developers work in Copilot, Cursor or another third party agent, they are not using Grok Build, which is the harness behind the 38.0%.

The model card exists, and it is better than the launch post

This briefing began from the premise that xAI had shipped a model to general availability without publishing a model card. That premise was wrong, and correcting it is worth more than the story it was meant to support. The card exists: a 30 page PDF headed "Model Card: Grok 4.7, September 21, 2026, Revision: 2026-09-21", built on launch day, served from media.x.ai with a last modified timestamp of 15:32 GMT on 21 September. The launch page, in the HTML served to us, does not link to it. Neither did the Grok 4.6 announcement link the Grok 4.6 card. Cards for Grok 4, Grok 4 Fast, Grok 4.1 and Grok 4.20 sit on a different host again, data.x.ai, under dated filenames.

What the Grok 4.7 model card states, and what it leaves out, on the points a buyer is likely to ask about.

Topic	Stated in the card	Not stated in the card
Model size	Nothing. The Grok 4.6 card described a "1.5T-scale model family"; the announcement says only that 4.7 uses "a new, larger base model"	Any parameter count for Grok 4.7, including the 2.1 trillion figure carried by press reports of Elon Musk's posts
Training data	Public data, internally generated data, licensed data, and "supplemental training on anonymized Cursor workflow data". Pretraining cutoff June 2026	Any SpaceX engineering, Starlink telemetry or failure log corpus, which appears only in secondary coverage
Terminal-Bench method	Grok Build harness, runs conducted by Harbor, 66 tasks, agents given up to eight hours, scores "sensitive to the agent harness"	The harness behind the competitor scores, and the number of trials per task
Where it runs	xAI API, Grok Build, Cursor on every plan tier, Microsoft Office add-ins, and gateways including OpenRouter, Vercel and Cloudflare	GitHub Copilot, although GitHub announced general availability the same day
Cyber posture	CyberGym 80.3% and CVE-Bench 37.7% at high, both unsafeguarded; HackerBench harmful compliance 3.31% at high and 4.02% at xhigh	Any independent replication of HackerBench, which is an internal xAI suite
Safety movement	Self-harm compliance 1.05%, up from 0.84% for Grok 4.6; general refusal compliance 1.10%, up from 0.93%; jailbreak compliance down	Confidence intervals, so whether those small rises are meaningful cannot be judged

The card earns credit on method. It names nine third party evaluation partners, states the harness for each benchmark it reports, discloses that Grok 4.6 was used as the grader model for the Grok 4.7 HealthBench results, and says plainly that its cyber capability tests were run without the production safeguards, so that the scores measure ability rather than refusal. It also flags an unrestricted configuration supplied to third party evaluators. Those are the disclosures a reviewer needs in order to discount a number, and most of them are absent from the launch page.

Two small inconsistencies are worth knowing before you quote a figure. The announcement table gives EEBench 64.0% for Grok 4.7; the card gives 66.0% at xhigh. The developer documentation states a knowledge cutoff of May 2026 while the card states a pretraining data cutoff of June 2026, which may be two different measurements rather than a contradiction. Neither changes a purchasing decision. Both are reasons to cite the card rather than the post.

In Copilot, the default is not a decision

GitHub's changelog entry of 21 September puts Grok 4.7 in the model picker across Visual Studio Code, Visual Studio, the Copilot CLI, the cloud agent, the Copilot app, JetBrains, Xcode and Eclipse, on Pro, Pro+, Max, Business and Enterprise plans, billed at provider list pricing under usage based billing. GitHub's own pricing reference lists it as generally available at $2.00 input, $0.50 cached input and $6.00 output per million tokens up to 200,000 input tokens, and $4.00, $1.00 and $12.00 above that threshold. The sentence that should reach your change advisory board is the one about enablement: "new models are automatically enabled unless an administrator has turned off the global default or explicitly disables this model".

Look at the label in the middle of that flow. A model that nobody has configured shows in enterprise settings as Delegate to Default Policy. The words describe an act of delegation, which sounds like a choice somebody made. It is the name for the absence of a choice. The same trap sits in the release status: GA is a statement about GitHub's rollout, not an assurance that the model suits your risk appetite, your regulated code or your data classification. Neither label is dishonest. Both read, to a busy administrator scanning a settings page, like evidence that somebody already thought about it.

GitHub Copilot model controls for Business and Enterprise plans, from GitHub's documentation on default model availability and enterprise model management.

Control	What it does	What it does not do
Default availability for released models, left on	Enables every unconfigured generally available model for seats in scope as soon as it ships	Evaluate the model, notify anyone, or record who accepted the risk
Setting a model to Disabled	Blocks that model for every user and agent app in scope, and keeps it out of automatic enablement	Stop the same model reaching your code through Cursor, a gateway or a personal API key
Delegate to organisations	Hands the decision to each organisation owner	Guarantee a decision: an unconfigured model then inherits that organisation's default policy instead
Data resident or FedRAMP restrictions	Excludes models that do not respect those policies from default enablement	Apply to models you have explicitly enabled yourself

On one axis Grok 4.7 arrives in a stronger position than the strongest Anthropic model. GitHub's hosting documentation says xAI operates its models in Copilot "under a zero data retention API policy", with prompts and completions held only in memory for the life of the request. Claude Fable 5 and Claude Fable 5.1 are the exception to GitHub's retention agreement: Anthropic retains prompts and outputs by default to run safety classifiers, eligible enterprises can request zero retention only through a time limited exemption running to the end of 2026, and those models are therefore excluded from default enablement altogether. If your reason for scrutinising a new model is data handling, that is an argument in xAI's favour and against Anthropic's flagship, and it should be weighed as one. It is also a commitment reported by GitHub about a third party, not something you can inspect. Ask for it in your contract paperwork rather than reading it off a documentation page.

Cursor works the other way round. Its enterprise documentation states that "When new models become available, Cursor doesn't immediately enable them for all enterprise teams", and that enterprise teams opt in. Model access control there is described as an Enterprise feature, configured under Team Settings and Organization Groups, and the two layers merge on a "most-permissive (union)" rule, so a group can widen access but never tighten it below what the team allows. Two consequences follow. Below the Enterprise tier the documentation describes no administrative model restriction at all, while the Grok 4.7 card lists Cursor availability as "all users, on every plan tier". And if you run Cursor Enterprise alongside Copilot, your strictest setting must live on the team baseline, because the union rule means a permissive group wins.

One pricing detail that gets misread. The "fast variant with twice the output speed at twice the price" from the announcement is, per xAI's documentation, "the same model served on faster infrastructure", available only in Cursor and Grok Build, excluded from Grok Build's free tier, and not offered on the public xAI API. Copilot bills Grok 4.7 at the standard $2 and $6 rates, which implies the standard service rather than the fast one; GitHub does not say so directly, so treat that as inference and confirm it if latency matters to your business case.

Procurement questions for a UK buyer

Who you are contracting with, and where the processing happens, are answerable from primary sources only in part. The model card records that "SpaceXAI is a doing-business-as (DBA) name of XAI LLC", so the brand on the launch page is not the contracting entity. For direct API use, xAI's regions documentation says the global endpoint may route "between regions for capacity or reliability, so the processing location is not guaranteed", and the only regional alternative it documents is a United States endpoint, serving grok-4.7 and grok-4.6 at a 10% token premium, whose guarantee covers request handling, inference, moderation and retained request data but not files, collections or server side tools. There is no documented endpoint that keeps processing in the United Kingdom or the European Economic Area. The data processing addendum is linked from x.ai's footer, but repeated requests for it returned an HTTP 403 bot challenge, so this briefing does not characterise its terms; that is a gap you should close before signing, not one to fill with assumption.

What to do, in order

The question that exposes the gap

Benchmark scores move. Artificial Analysis will refresh its numbers, other harnesses will publish, and the twelve point gap on Terminal-Bench 4.0 will narrow, widen, or be explained by someone who runs the model three ways. That argument will resolve itself without you. The governance point will not. A frontier model became generally available inside your development tooling on the day it was announced, priced per token against a bill you have not modelled, scored by its maker in a harness your developers do not use, and switched on by a policy setting whose name describes a decision nobody made.

So: if Grok 4.7 is live for your developers today, can you name the person who approved it, and the evidence they looked at?