Grok 4.7 Review 2026: Same $2/$6 Sticker, 2.25× the Tokens — xAI's Agent Upgrade Costs More Per Task
Grok 4.7 Review 2026: Same $2/$6 Sticker, 2.25× the Tokens — xAI's Agent Upgrade Costs More Per Task
Verdict first: Grok 4.7, released September 21, 2026 by the company now branded SpaceXAI, is a genuine and well-measured upgrade at the thing xAI cares about most — long-horizon agentic work. The Coding Agent Index rose from 47 to 56, AA-Briefcase added 111 Elo, and Terminal-Bench 4.0 jumped from 18% to 33%. The list price did not move an inch: $2 input / $0.50 cached / $6 output per million tokens, exactly Grok 4.6's rate. That combination — better agent, same price — reads like an unambiguous win, and most launch coverage stopped there. It is not the whole story. Artificial Analysis measured Grok 4.7 at roughly 81,000 output tokens per Intelligence Index task, against 36,000 for Grok 4.6 and 27,000 for GPT-6 Astra. Agentic work is billed on tokens actually consumed, not tokens intended, so identical per-token pricing at 2.25× the tokens is not the same price at all — it is a roughly 2.25× increase in spend per completed task, for a general-intelligence gain of 2 index points. Two further caveats belong up front: most of xAI's headline benchmarks, including Terminal-Bench, were run inside its own Grok Build harness, and several published scores went down. This is a real upgrade for one specific job, priced more honestly in tokens than in dollars.
Grok 4.7 — same $2/$6 list price as Grok 4.6, but roughly 81k output tokens per benchmark task against 36k
What Grok 4.7 Actually Is
Grok 4.7 arrived on September 21, 2026, with most English-language coverage landing a day later. It is a direct successor to Grok 4.6, which itself shipped in mid-August, and the cadence is worth noting: xAI has now put out 4.5, 4.6 and 4.7 in roughly six weeks. The company is also now referred to in its own materials as SpaceXAI, "the company formerly called xAI" — a rebrand that reflects the merger rather than any change in the model line.
Under the hood, this is not a re-tune. Grok 4.7 uses a new and larger base model than 4.6, followed by a longer reinforcement-learning run deliberately weighted toward problems that take hours rather than minutes to finish. The model card describes that run as "agentic RL across knowledge work, general coding, and purpose-built environments." The pretraining cutoff is June 2026, with supplemental training data running through August 2026.
The most consequential design decision is stated plainly in xAI's own documentation and then quietly ignored by most benchmark tables: Grok 4.7 was trained to natively understand the Grok Build harness. It is positioned as a fit for that stack, in xAI's words, "not as a free win in every third-party agent." That single sentence explains a great deal of what follows — both the strong numbers and the ones that fail to transfer.
| Specification | Grok 4.7 | Grok 4.6 |
|---|---|---|
| Release | September 21, 2026 | Mid-August 2026 |
| Base model | New, larger | Previous generation |
| Context window | 500,000 tokens | 500,000 tokens |
| Modalities | Text + image in, text out | Text + image in, text out |
| Knowledge cutoff | June 2026 (supplemental to Aug 2026) | Earlier 2026 |
| Input / cached / output | $2.00 / $0.50 / $6.00 per 1M | $2.00 / $0.50 / $6.00 per 1M |
| Above 200K input tokens | $4 / $1 / $12 per 1M | $4 / $1 / $12 per 1M |
| Reasoning effort | low / medium / high (default) / xhigh | Same ladder |
| API model ID | grok-4.7 | grok-4.6 |
On paper, the only row that changed is the base model. Everything else is held constant — and that constancy is precisely where the interesting problem lives.
The Pricing Trap: Same Rate, 2.25× the Tokens
Every launch summary led with the same line: Grok 4.7 keeps Grok 4.6's pricing. That is literally true. The rate card is $2.00 per million input tokens, $0.50 cached, $6.00 output, and it is unchanged. Above 200,000 input tokens in a single request, every token in that request rebills at $4 / $1 / $12 — the same cliff as before, and still an all-or-nothing trigger rather than a marginal one, so a 201,000-token prompt costs roughly twice what a 199,000-token prompt does.
What no launch summary mentioned is that the rate was never the thing that determined your bill. Agentic and reasoning models are billed on tokens actually consumed. A model that deliberates longer at an unchanged rate produces a larger bill for the same completed task, and that is exactly what happened here. Artificial Analysis, running the standardized evaluation it uses across every frontier model, measured:
| Model | Output tokens per Intelligence Index task | Relative to Grok 4.7 |
|---|---|---|
| Grok 4.7 (xhigh) | ~81,000 | — |
| Grok 4.6 | ~36,000 | Grok 4.7 uses 2.25× as many |
| GPT-6 Astra (max) | ~27,000 | Grok 4.7 uses 3.0× as many |
| Muse Spark 1.3 (max) | ~60,000 | Grok 4.7 uses 1.35× as many |
The headline that should have been written. At an identical $6 per million output tokens, consuming 81,000 tokens instead of 36,000 means the output component of a task's cost is roughly 2.25× higher. "Grok 4.7 keeps Grok 4.6's pricing" is technically accurate and practically misleading, because the metric customers experience — cost per task completed — moved sharply in the wrong direction. Artificial Analysis states the implication directly: the headline pricing is likely to be misleading, and the model will probably cost more on real tasks. Note also that the 36,000 figure for Grok 4.6 is quoted inconsistently within that same analysis, appearing as 38,000 elsewhere, so treat the exact multiple as approximately 2.1× to 2.25× rather than a precise number.
There is one honest counterweight, and it comes from Cursor rather than xAI. Cursor's own testing put Grok 4.7 at $4.69 per task against Grok 4.6's $5.20, while also scoring higher (43.9% versus 40.4%). In that harness, the newer model is both better and cheaper per task. The two findings are not contradictory — they are the whole point. Token consumption is a property of the model and the harness together, and a model trained against Grok Build will behave differently under someone else's orchestration. If you run Grok 4.7 in the environment it was tuned for, the economics can improve. If you bolt it into a generic agent loop, the token bill is yours to discover.
Two smaller pricing notes. A fast variant offers twice the output speed at twice the price, but is limited to Cursor and Grok Build rather than the public API — its API surface is not confirmed in xAI's public documentation. And a US regional endpoint carries a 10% premium, which matters for teams with data-residency requirements. OpenRouter lists the same model at $1.60/$4.80, a 20% discount to list.
What Genuinely Improved
The agentic gains are real, measured on a neutral evaluation, and large enough that they should not be dismissed because of the cost story. Artificial Analysis runs Grok 4.7 with xhigh reasoning inside Grok Build and reports a Coding Agent Index of 56, up 9 points from Grok 4.6's 47. That places it fourth among models evaluated in their own native harness, behind only Claude Fable 5.1, GPT-6 Astra and Claude Opus 5 — a genuinely elite tier, and the highest position any xAI model has held on that particular ranking.
| Benchmark | Grok 4.6 | Grok 4.7 | Direction |
|---|---|---|---|
| Coding Agent Index (AA, native harness) | 47 | 56 | +9 |
| DeepSWE v1.1 (AA) | 65% | 73% | +8 pts |
| Terminal-Bench 4.0 (AA, Grok Build) | 18% | 33% | +15 pts |
| SWE-Atlas-QnA (AA) | 58% | 63% | +5 pts |
| AA-Briefcase v1.1 | 1,546 Elo | 1,657 Elo | +111 Elo |
| GDPval-AA | 1,605 Elo | 1,695 Elo | +90 Elo |
| AA-Omniscience hallucination rate | 34% | 29% | −5 pts (better) |
The AA-Briefcase result deserves a closer look because it is the long-horizon knowledge-work evaluation, and it is where Grok 4.7 lands "just behind Claude Opus 5 and Claude Fable 5.1." The headline gain of +111 Elo is strong. The sub-scores are where it gets interesting: analytical quality rose from 1,690 to 1,994 — a very large improvement — while presentation quality fell from 1,519 to 1,499. The model got meaningfully better at reasoning through a problem and marginally worse at communicating the result. For an agent producing work a human will read, that is a trade worth knowing about rather than a rounding error.
On xAI's own tables the picture is similar and, in places, dramatically better. CursorBench 4.0 reached 46.3% at xhigh (43.9% at high) against 4.6's 40.4%; SWE-Marathon v1.1 rose from 31.9% to 46.0%; EEBench from 53.0% to 64.0% — or 66.0%, depending on whether you read the press table or the model card, which disagree; HealthBench Professional from 48.5% to 56.7%; and Harvey Legal Agent from 15.8% to 19.6%. The safety picture is mixed but not alarming: 62.4% on LatchBio's biosafety benchmark, with 3.3% of risky dual-use cybersecurity prompts allowed through on HackerBench v0.3. The model also claims genuine self-verification behaviour — fewer steps and output tokens on hard tasks as a result of more careful self-checking on long trajectories. That claim sits awkwardly beside the 81,000-token measurement, and the independent number should win that argument.
What Went Backwards
A release where the base model is swapped for a larger one and training runs longer is expected to improve nearly everything. Grok 4.7 does not, and the exceptions are informative rather than random.
| Benchmark | Change vs Grok 4.6 | Reading |
|---|---|---|
| AA-LCR (long-context reasoning) | −3.7 points | A larger model with a 500K window got worse at long context |
| AutomationBench-AA | −1.1 points | Regression in the agentic category xAI is promoting hardest |
| AA-Briefcase presentation quality | −20 Elo | Analytical quality up sharply, output polish down |
| AA-Omniscience accuracy | −1 point (48% → 47%) | Hallucination fell more than accuracy, so the index still rose |
| General Intelligence Index | +2 points only (44 → 46) | Outside agentic work, broadly matches Grok 4.6 |
The long-context regression is the one to watch. Grok 4.7 carries the same 500,000-token window as its predecessor, and long-context retrieval is one of the more reliable differentiators in production work — retrieving the right clause from a 300-page contract, or the correct function from a large repository. Losing ground there while gaining in agentic loops suggests the training emphasis on multi-hour tasks was paid for, at least partly, in breadth. Artificial Analysis frames the general result more diplomatically, noting the model "broadly matches Grok 4.6 on the other Intelligence Index tasks." Broadly matching your predecessor on general reasoning, two points better on the composite, at 2.25× the tokens, is not a compelling upgrade for anyone whose workload is not specifically agentic.
There is also the matter of speed. Grok 4.7 produces roughly 188 tokens per second on long prompts and takes about 7.1 minutes per Intelligence Index task. Faster than a human by any measure, slower than the "twice as fast" framing that accompanied the launch — that claim refers to the separate fast variant, which is not available on the public API.
The Harness Problem: Whose Benchmark Is This Anyway?
xAI's headline table is entirely self-reported, which is normal for a launch and is not by itself disqualifying. What is more unusual is the degree to which the reported numbers depend on xAI's own orchestration layer.
Terminal-Bench 4.0, EEBench, CADGenBench and CyberGym were all run inside Grok Build. xAI's model card says so explicitly. That does not make the results fake — a model trained against a harness is legitimately expected to perform there, and shipping model and harness together is a defensible product decision. But it means the numbers describe Grok 4.7 plus Grok Build, not Grok 4.7 as a drop-in component. Teams planning to call the raw API from their own agent framework are looking at a different system than the one that scored 38.0% on Terminal-Bench.
Three further comparability problems are worth flagging:
- Cross-effort comparisons. Several headline figures compare Grok 4.7 at xhigh reasoning effort against Grok 4.6 at high. That is not a like-for-like upgrade measurement — it is a measurement of the new model with more thinking budget against the old model with less, and the extra tokens are billed.
- A version break in CursorBench. CursorBench 4.0 is not comparable to CursorBench 3.2, so the 40.4% to 46.3% improvement is a within-4.0 comparison only and cannot be read as a trend across older releases.
- Partial credit presented as a score. The FrontierSWE V2 figure of 29.0% at xhigh is the average partial-credit reward, not task resolution. It is not the same kind of number as a pass rate and should not be placed beside one in a table.
The disagreement that matters most is EEBench. The model card lists 66.0%; the accompanying press table lists 64.0%. Both are xAI documents published at the same time about the same benchmark. The gap is small, and xAI is not alone in this kind of sloppiness on launch day, but when a vendor's own two documents disagree about a headline number, everything else in that table deserves to be read as provisional.
Where Grok 4.7 Sits in the Frontier
The reception was noticeably cooler than the benchmark gains would suggest. One widely circulated headline described the launch as landing "like a damp squib," and reaction posts cited in that coverage included "Grok 4.7 is pretty terrible." Some of that is the usual launch-day noise. Some of it is substantive, and the substantive part is this: an Intelligence Index of 46 places Grok 4.7 behind Claude Fable 5.1 Max, Claude Fable 5, Claude Opus 5, GPT-6 Astra, GPT-5.6 Sol at max effort, and Meta's Muse Spark 1.3 Max. That is a long list to sit behind, and it includes a model from a company — Meta — that until recently was not regarded as a coding leader.
There is a caveat on index comparisons: Artificial Analysis has revised and re-based its Intelligence Index across 2026, so figures quoted in reviews written months apart are not always measured on the same scale. Compare within a single vintage rather than across posts. Within the current one, the relevant comparison is 4.7 at 46 against 4.6 at 44 — a two-point move.
The more pointed comparison came from open weights and from China. On Terminal-Bench 4.0, Grok 4.7 ranks below the open-weight DeepSeek V4.1 Flash, a model that is freely downloadable. That particular benchmark is contested — Epoch AI has labelled it flawed — so the ranking should not be treated as decisive. But the broader competitive context is real and appears to be what stung: DeepSeek is reported to be training a 2-trillion-parameter model and intends to follow it with an 8-trillion-parameter one, against V4.1 Flash's 552 billion. The same reporting describes DeepSeek ordering 160,000 Huawei Ascend 950DT chips for a 1-gigawatt data centre in Inner Mongolia, at roughly $16,000 each — about $2.56 billion. A lab operating at a fraction of SpaceX-class compute is fielding models that score competitively against Grok 4.7. That is the strategic problem underneath the "same price" headline.
Elon Musk's own framing acknowledged the gap. He described himself as cautiously optimistic that SpaceXAI would have a Fable- and GPT-6-class model within two to three months, and could reach industry pole position in about six months. He has separately pitched the unreleased Grok 4.9 as Fable-class. Both statements are forward guidance about models that do not exist yet, made at the launch of a model that does — which is itself a data point about where Grok 4.7 sits.
The Honest Take
Grok 4.7 is the best agentic coding model xAI has shipped, sold at a price that quietly costs more than the last one. Both halves of that sentence are true and they do not cancel out. If your workload is long-horizon coding inside Cursor or Grok Build, the Coding Agent Index gain from 47 to 56 is the largest single-generation jump in this line, Cursor's own per-task cost measurement went down rather than up, and the case for upgrading is straightforward.
If your workload is anything else, the case collapses. Two index points on general intelligence, a long-context regression of 3.7 points, worse presentation quality on long-form knowledge work, and 2.25 times the output tokens at an unchanged rate is not an upgrade — it is a more expensive model wearing the same price tag. The correct move there is to stay on Grok 4.6, which is exactly what "broadly matches Grok 4.6 on the other Intelligence Index tasks" implies without xAI intending it to.
Buy it if: you run long-horizon agentic coding in a harness Grok 4.7 was trained against, you measure token consumption per task rather than per token, and you need frontier-adjacent agentic performance at a sub-Anthropic sticker price.
Skip it if: your work is general reasoning, long-context retrieval or document-heavy knowledge work — Grok 4.6 is cheaper per task and in places better; if you call the raw API from a custom agent framework, because the headline figures were produced in xAI's own harness; or if you need legal, clinical or financial output, where Harvey Legal Agent at 19.6% and the model card's own warning against autonomous high-stakes use both apply.
Watch two things before your next renewal. First, whether xAI does anything about token consumption — a model that thinks 2.25× longer is a model with a roadmap item, and the next version will either fix it or lean further into it. Second, whether Grok 4.9 arrives on Musk's two-to-three-month timeline and whether it closes the gap to Fable 5.1 and GPT-6 Astra. Between now and then, treat "same price as Grok 4.6" as a rate-card fact and not a budgeting one.
FAQ
What is Grok 4.7?
The model xAI — now branded SpaceXAI — released on September 21, 2026 as a direct successor to Grok 4.6. It uses a new, larger base model plus a longer reinforcement-learning run weighted toward problems that take hours to complete, and it was trained to natively understand xAI's own Grok Build harness. The context window is 500,000 tokens with text and image input and text output. The API model ID is grok-4.7.
How much does Grok 4.7 cost?
$2.00 per million input tokens, $0.50 cached, $6.00 output — unchanged from Grok 4.6. Once a single request reaches or exceeds 200,000 input tokens, every token in that request rebills at $4 / $1 / $12. A faster variant runs at twice the output speed for twice the price but is limited to Cursor and Grok Build rather than the public API. A US regional endpoint costs 10% more, and OpenRouter lists the model at $1.60/$4.80.
Is Grok 4.7 actually cheaper to run than Grok 4.6?
Not necessarily, and this is the most important finding in the release. The per-token rate is identical, but Artificial Analysis measured Grok 4.7 at roughly 81,000 output tokens per Intelligence Index task against 36,000 for Grok 4.6 and 27,000 for GPT-6 Astra. Because agentic work bills on tokens consumed, the same rate at 2.25× the tokens is a substantially higher cost per completed task. Cursor's own testing went the other way — $4.69 per task versus $5.20 — which shows the effect is harness-dependent rather than fixed.
How much better is Grok 4.7 than Grok 4.6?
On agentic work, substantially. The Coding Agent Index rose from 47 to 56, DeepSWE v1.1 from 65% to 73%, Terminal-Bench 4.0 from 18% to 33%, AA-Briefcase by 111 Elo and GDPval-AA by 90 Elo. On the general Intelligence Index the gain is only 2 points, from 44 to 46. Several benchmarks regressed, including AA-LCR at −3.7 points and AutomationBench-AA at −1.1 points.
Are the benchmark numbers independently verified?
Partially. Artificial Analysis publishes its own standardized evaluation, which is where the token-usage and regression findings come from and the most useful cross-model comparison available. xAI's headline table is self-reported, and its strongest results — Terminal-Bench 4.0, EEBench, CADGenBench and CyberGym — were produced inside Grok Build. The FrontierSWE V2 figure of 29.0% measures partial-credit reward rather than task resolution. The model card and press table also disagree on EEBench, at 66.0% and 64.0%.
Where can I use Grok 4.7?
Via the Grok API (grok-4.7), inside Cursor, inside Grok Build, and through GitHub Copilot and other third-party coding harnesses, plus routers and clouds including OpenRouter, Vercel and Cloudflare. Reasoning effort is configurable across low, medium, high and xhigh, with high as the default. Rate limits are 150 requests per second and 50 million tokens per minute across us-east-1, us-west-2 and us-central-1.
Should I use Grok 4.7 for production work?
It fits long-horizon coding and agentic tasks inside a harness it was trained against, particularly where the low sticker price matters and token consumption is monitored directly. It fits poorly for general reasoning or knowledge work where Grok 4.6 already suffices, given the 2-point index gain at 2.25× the tokens. xAI's model card states the model is not intended for autonomous use in medicine, law, finance or safety-critical systems without human oversight, and Harvey Legal Agent at 19.6% suggests legal work in particular remains unreliable.
Sources
- Artificial Analysis — Benchmarking Grok 4.7: Intelligence Index 46, Coding Agent Index 56, and 81k output tokens per task
- Artificial Analysis — Grok 4.7 (xhigh) intelligence, performance and price analysis
- LLM Stats — Grok 4.7 release: same price, longer horizons
- LLM Stats — Grok 4.7 benchmarks, pricing and context window
- 4sysops — Grok 4.7 keeps Grok 4.6 pricing while cutting the cost of agent work
- Windows Report — SpaceXAI launches Grok 4.7, its "most powerful model for coding and knowledge work"
- Wccftech — Grok 4.7 lands like a damp squib, as DeepSeek teases an 8-trillion-parameter model
- Beam AI — Grok 4.7 for AI agents: price, benchmarks, fit
- LiteLLM — Day 0 support: Grok 4.7
- OpenRouter — Grok 4.7 API pricing and providers
- ITBrief — xAI launches Grok 4.7 for coding and knowledge work
All vendor benchmark figures reported here are xAI's own and have not been independently reproduced; several were produced inside the Grok Build harness, and the model card and press table disagree on EEBench. Figures attributed to Artificial Analysis come from that organization's standardized evaluation and are the basis for the token-usage, regression and Elo findings. Artificial Analysis has re-based its Intelligence Index during 2026, so index figures from reviews published in different months are not always directly comparable. Prices, rate limits and specifications change frequently, and the long-context repricing clause applies to entire requests at or above 200K input tokens — verify against xAI's official documentation before making purchasing decisions.
评论
发表评论