Grok 4.7: Same $2/$6 Price, and Where the Scores Moved

SpaceXAI released Grok 4.7 on 21 September 2026. The model id is grok-4.7. The rate card did not move: $2 per million input tokens and $6 per million output, the same as Grok 4.6. The launch line, "twice as fast, at half the price of comparable models," is the comparison with GPT-5.6 Sol at $4/$20 and Claude Fable 5.1 at $10/$50 in SpaceXAI's own table. Against 4.6, they say it is the same price and the same speed.

Two facts decide whether you switch. Artificial Analysis scores it 46 on Intelligence Index v4.3.2, two points above Grok 4.6. That index was rebuilt. The 46 is not a fall from the 61 we reported for 4.6 in August, which was v4.1.1. The second fact is the harness. The model card reports 38.0% on Terminal-Bench 4.0 inside Grok Build. The Decoder, reading Artificial Analysis's standardized chart, reports 26%. Same benchmark name. Different loop.

This is the day after launch. What follows is the model card, the API price page, Artificial Analysis, XBOW's early-access eval, and the first notes from Cursor's forum and Hacker News. It is not our own run on a client workload. We'll test it the same way we always do: one real task, same harness, cost per completed job.

What Grok 4.7 is, in numbers

• Model id grok-4.7. Text and image in, text out. No published text output cap on the API. Proprietary weights. SpaceXAI says the base is larger than Grok 4.6. The model card does not give a parameter count.
• 500,000-token context window. In Cursor the standard window is 256k and the max is 500k.
• The model card puts the pretraining cutoff at June 2026, with supplemental data generated as late as August 2026. The API docs page still says May 2026. Use the card until those two pages match.
• Reasoning effort: low, medium, high (default), xhigh. Cursor says the steps between efforts are wider than on Grok 4.6.
• API list below 200,000 prompt tokens: $2.00 input, $0.50 cached input, $6.00 output per million. At or above 200,000, the whole request bills at $4.00 / $1.00 / $12.00.
• Fast is the same model on faster hardware, at twice the token rates. It exists in Cursor and Grok Build only. It is not on the public API, and Grok Build's free tier does not include it. Above 200,000 prompt tokens, Fast is $6.00 / $1.50 / $18.00.
• Cursor's cliff is 256k, not 200k. At or below 256k input: $2 / $0.50 / $6, and Fast at $4 / $1 / $12. Above 256k, standard doubles and Fast goes to $6 / $1.50 / $18. Colin confirmed that table on launch day. Fast is the default speed on Pro and higher. The India Start plan is fixed at medium effort and standard speed.
• The US endpoint https://us.api.x.ai/v1 adds 10%: $2.20 / $0.55 / $6.60 below 200k, and $4.40 / $1.10 / $13.20 above. Right now that endpoint serves grok-4.7 and grok-4.6 only.
• API tools: function calling, web search, X search, code execution. Cursor adds the agent tool set, including shell, browser, and file edits.
• Set prompt_cache_key on the Responses API, or the x-grok-conv-id header on Chat Completions. Without it, multi-turn loops often miss the cache and pay full input price. Responses always return reasoning.encrypted_content. Pass those items back unchanged.
• Available on day one in the API, Grok Build (where it is the default), and Cursor on every plan. GitHub Copilot is rolling it out gradually to Pro, Pro+, Max, Business, and Enterprise, in VS Code, Visual Studio, the Copilot CLI, the cloud agent, JetBrains, Xcode, Eclipse, and the Copilot app. Gateways listed in the model card: OpenRouter, Vercel, Cloudflare, Snowflake, and Databricks Mosaic. It is also the default in the Grok add-ins for Word, PowerPoint, and Excel. Consumer web, the mobile apps, and Grok inside X come later.
• The model card says 4.7 had supplemental training on anonymized Cursor workflow data. CursorBench 4.0 is built from real Cursor sessions. Read that score with the training note attached. We flagged the same issue for Grok 4.6 on an earlier CursorBench.

OpenRouter currently lists x-ai/grok-4.7 at $1.60 input and $4.80 output, with cache reads at $0.40, about 20% under the API card. A priority tier there is $3.20 / $9.60. Their first-day mix showed cache hit rates around 85% and an effective input price near $0.68 per million. That is a router discount plus a traffic mix. It is not the invoice from api.x.ai. Recheck the listing before you quote a client a unit cost.

The benchmarks, with the harness attached

The model card is the place to start, because it names the harness and the lab that ran each row. The launch graphic compresses several of these into one table and prints 64.0% for EEBench, where the card's xhigh figure is 66.0%.

• CursorBench 4.0: 46.3% at xhigh, 43.9% at high. The launch table has Grok 4.6 at 40.4%, Sol at 41.7%, and Fable 5.1 at 51.8%. Version 4.0 added longer tasks, so these scores are not comparable with the CursorBench 3.2 number in our Grok 4.6 piece.
• DeepSWE v1.1, every model in mini-SWE-agent, graded by Datacurve: 71.0% at high. Sol max 72.7%, Fable 5 max with fallback 69.7%, Grok 4.6 high 65.2%.
• Terminal-Bench 4.0, Grok Build harness, graded by Harbor: 38.0% at xhigh. Fable 5.1 max 57.9%, Sol max 37.3%, Grok 4.6 high 20.3%. Agents get up to eight hours. The card says absolute scores stay sensitive to the harness.
• SWE-Marathon v1.1, each model in its own agent, graded by Abundant: 46.0% at high. Opus 5 max 50.0%, Fable 5 max 45.0%, Sol max 42.5%, Grok 4.6 high 31.9%. A trial counts only if every verifier passes.
• FrontierSWE V2, Proximal's Proximus harness, partial credit, up to 20 hours: 29.0% at xhigh. Fable 5.1 max 56.3%, Sol max 32.2%, Grok 4.6 xhigh 25.3%, Kimi K3 max 25.9%. This is mean partial credit, not a solve rate.
• EEBench, Grok Build, graded by Atopile: 66.0% at xhigh. GPT-6 Astra max 69.3%, Opus 5 max 61.6%, Grok 4.6 xhigh 60.0%, Grok 4.6 high 53.0%, Sol max 39.4%.
• CADGenBench generation split, graded by Mecado: 44.4% at high, against Grok 4.6 high at 40.9%, Sol at 37.1%, and Opus 5 at 36.6%.
• Harvey Legal Agent Benchmark, Vals AI's Valkyrie harness, internet off: 19.6% at xhigh, against 15.8% for Grok 4.6 high, 6.7% for Fable 5.1, and 2.5% for Sol. Every model is low on this set. The gap is the useful part.
• HealthBench Professional: 56.7% at xhigh, up from 48.5% for Grok 4.6. Astra max 63.4%, Fable 5.1 max 62.1%, Sol max 60.5%. Grok 4.6 was the grader for the Grok 4.7 number.

Artificial Analysis ran the model themselves at xhigh. Intelligence Index v4.3.2 is 46 for both xhigh and high on their release page. The Decoder puts Fable 5.1 and GPT-6 at 53 on that same index. Outside agentic knowledge work, AA says 4.7 broadly matches 4.6 high, with Terminal-Bench 4.0 up 4.5 points, GDP.pdf up 3.0, AA-LCR down 3.7, and AutomationBench-AA down 1.1.

Their Coding Agent Index is a different product. Grok 4.7 with Grok Build scores 56, up 9 from Grok 4.6 xhigh at 47, and ranks fourth among native harnesses, behind Fable 5.1, GPT-6 Astra, and Claude Opus 5. The components they published: DeepSWE from 65% to 73%, Terminal-Bench 4.0 from 18% to 33%, SWE-Atlas-QnA from 58% to 63%. That 33% is not the model card's 38%, and neither is the standardized Intelligence Index row. Incrypted, citing Vals AI's own post, reports the opposite movement on Vals's composite: 54.2% and 24th, against 59.2% and 14th for Grok 4.6, with the gains concentrated in legal and medical tasks.

On AA-Briefcase, long professional work, 4.7 scores 1657 Elo, up 111 from Grok 4.6 high. The gain is analytical quality, 1994 Elo against 1690. Presentation quality slipped, 1499 against 1519. GDPval-AA is 1695 Elo, up 90 from 1605. On the launch chart Fable 5.1 max is 1735 and GPT-6 Astra max is 1542. Hallucination on AA-Omniscience fell from 34% to 29%. Accuracy stayed flat, 47% against 48%.

38% on Terminal-Bench inside Grok Build, 33% on Artificial Analysis's coding-agent run, and about 26% on their standardized chart are three measurements. Quote the harness with the number.

Token price is not task cost

Output at $6 is 70% below Sol's $20 and 88% below Fable's $50. Procurement notices that gap, and it should. It is the price of a token, not the price of a finished job.

Artificial Analysis measured about 81,000 output tokens per Intelligence Index task at xhigh. Grok 4.6 xhigh used about 38,000, Grok 4.6 high about 36,000, and GPT-6 Astra max about 27,000. Cost per index task on their release page is $2.73 at high and $3.74 at xhigh, for the same index score of 46. If that pattern holds on your workload, xhigh is a more expensive way to buy the same composite. XBOW saw the same shape on exploit crafting: xhigh was not distinguishable from high, while low was clearly worse.

Cache reads stayed at $0.50 per million, a quarter of the input price. On a long coding loop, cache reads are most of the bill. A commenter on the Hacker News thread pointed out that $0.50 is already above Sol's cache rate. The 200,000-token API cliff still doubles every token in the request, not only the overflow. Cursor's cliff is 256k. A prompt that is cheap in the editor can be expensive on the API once it crosses 200k, and a prompt that is safe in Cursor can still blow the API budget.

Speed depends on who measured it. Artificial Analysis put answer speed near 188 tokens per second on long prompts, and about 7.1 minutes per Intelligence Index task. Their release page lists 55 tokens per second at high and 39 at xhigh. On the Cursor forum, people said even the Fast tier felt slow while everyone was trying the model on day one. Treat launch-day latency as a queue, and remeasure before you promise a client a response time.

The rate card is still $2/$6. At xhigh, Artificial Analysis counted about twice the output tokens Grok 4.6 used, for two points on the index.

Trained for one harness

The launch post describes a longer reinforcement-learning run, weighted toward problems that take many hours, and native training on the Grok Bot harness. The model card adds what the blog compresses. Supplemental training ran longer than for 4.6. Grok 4.6 generated the supervised trajectories, and model checks filtered the bad ones. The domains are the same family as 4.6: knowledge work, general coding, and purpose-built environments for kernel optimisation, web development, and CAD. On top of that, 4.7 was trained on anonymized Cursor sessions.

XBOW had early access and split the results by orchestration. In their production exploit-crafting harness, 4.7 was slightly down from 4.6 at a fixed iteration budget, and it often needed more iterations to reach the same result. Each completion was shorter. It preferred short shell commands over longer Python scripts. One old failure mode disappeared: 4.6 would sometimes reason until the context filled and never take an action, in about 0.85% of runs. XBOW counted that stall in none of the 4.7 runs.

On full offensive-security workflows, the picture flipped with the orchestrator. Build-based systems averaged 68 true-positive findings, against 42 for 4.6 in that style of setup. XBOW's own external harnesses had been the stronger way to run 4.6. With 4.7, leaving Build looked like a small step back. They still place it below the strongest systems they test. The New Stack drew the same conclusion from the training description: a model trained against one tool schema gives some of the gain back when you drop it into a different loop. SWE-Marathon scores native harnesses for that reason. If you call grok-4.7 from a generic tool loop, budget for the generic-loop number.

Day-one notes from people using it

Cursor's feedback thread is still a first afternoon. One user said 4.7 is easier to talk to than 4.6, which had been inventing terms in the reasoning trace and then using them as if you had been there. Another asked it to revise a handbook of about 30 modules and got metaphors that made the text worse, then worse again as he corrected it. A third is watching token burn, because 4.6 at extra-high Fast had only used 32% of his cycle. Nobody on that thread has a week of production use.

The Hacker News thread, a few hundred comments in by the next morning, splits the same way. One engineer said recent Groks have been strong on frontend work and weaker than Claude on backend Ruby, and that he prefers the plainer English. Another said Grok is terse past the point of being clear: it names something in the trace and then assumes you know the name. A third said 4.7 feels slower and more expensive than 4.6 on his agent workflows, and that he cannot yet tell whether it clears the line where he still uses Sol or Opus. A fourth runs Grok in code review because it finds bugs Fable and Astra miss. Those are named accounts on day one.

On Reddit's Cursor thread, one person switched back to 4.6 the same day. Another said a few hours in the editor "did alright." The recurring question from subscription users is quota, which is the right question if you are not paying the API directly. A model that spends twice the output tokens can empty a Cursor pool faster at the same sticker price.

Safeguards, and what barely moved

SpaceXAI says 4.7 has a new safeguard stack and is the strongest Grok they have tested on refusals and jailbreaks. On their internal HackerBench v0.3, 3.3% of risky dual-use cyber prompts got through, and legitimate security work was rarely blocked. LatchBio's BioSecBench refusal score is 62.4% at xhigh, up from 45.6% for 4.6. The model card says dual-use biology capability did not rise: virology troubleshooting accuracy on VCT fell from 67.4% to 63.0% at high. Offensive cyber scores barely moved. CyberGym is 80.3% against 79.7% for 4.6 high. CVE-Bench reward at xhigh is 36.6%, below 4.6 high at 39.8%. Selected security partners have invite-only access to a red-team configuration. The card also says the model is not for autonomous decisions in medicine, law, finance, or safety-critical systems without a person and a domain expert. That belongs in the vendor review, next to the Elo.

Where we'd use it

For long agent jobs that already run in Grok Build or Cursor, 4.7 is the Grok we would try in place of 4.6. SWE-Marathon moving from 31.9% to 46.0%, and AA-Briefcase gaining 111 Elo, are the rows that justify a trial. The $2/$6 card is why the trial is cheap enough to run, provided the cache key is set and the prompt stays under the cliff for that surface. API and Cursor do not share a cliff.

We would not drop a stack that is winning FrontierSWE-shaped work, where Fable 5.1 is at 56.3% partial credit and 4.7 is at 29.0%. We would not treat a two-point index gain as a reason to leave Astra or Fable for general capability. 46 against 53 is a wide gap, and 4.7 spends more output tokens to stand there. We also would not put Grok on a customer-facing or compliance-sensitive path because the coding scores moved. That vendor-risk conversation is the same one we wrote for 4.6, and it happens in writing before anyone argues about Elo.

• Good fit: long coding and knowledge-work loops already inside Grok Build or Cursor, where you can verify the result and you stay under the context cliff.
• Good fit: a second opinion in code review, which is how at least one Hacker News user is already using the family, and document or spreadsheet work where analytical quality matters more than slide polish.
• Good fit: electrical-engineering and CAD-shaped tasks, where EEBench and CADGenBench are the rows 4.7 actually leads or nearly leads.
• Poor fit: a single-model stack that has to win 20-hour partial-credit SWE, clinical answers, or a standardized Terminal-Bench run. Fable and Astra still lead those.
• Poor fit: workflows that regularly cross 200k tokens on the API, or 256k in Cursor, without compaction. Poor fit anywhere the Grok brand is a procurement problem, regardless of the rate card.
• Poor fit: xhigh by default. On the Intelligence Index, high and xhigh scored the same 46, and xhigh cost $3.74 per task against $2.73.

Kimi K3 still has open weights and a million-token window. On FrontierSWE V2 the model card has K3 at 25.9% and Grok 4.7 at 29.0%, in different harnesses. If you need weights you can host, K3 still has that job. DeepSeek V4 Flash is still the price floor when the work is volume and you can live with the steering. Grok 4.7 is priced against Sol and Fable, on cost per finished task, after you count tokens, steps, and retries.

How we'd test it on your workload

Same method as every release. Take one real task from a live project. Run it on the incumbent and on grok-4.7, in the harness you will actually ship, at high and at xhigh. If the harness has to change, write that down, because XBOW's numbers flipped when they did. Count cost per completed task, including cache, how often a person had to step in, and whether xhigh beat high by enough to pay for the extra tokens. If the task is a document, read it. AA's presentation Elo went down while the analytical Elo went up, and the handbook report on the Cursor forum has the same shape.

If you're weighing an AI feature and want to know whether this price makes the feature affordable, or whether this vendor should touch the data at all, that's a scoping conversation we're glad to have. Send us the constraint you're stuck on and you'll get a written scope and a fixed quote before any commitment.