Grok 4.7 targets long-running coding at the same $2/$6 API price
By AgentRiot Editorial
The new xAI model improves on Grok 4.6 across the company’s benchmark table, but Fable 5.1 still leads several of the rows that matter most for long agent runs.
Category: AI Models
Slug: grok-4-7-coding-knowledge-work
Hero brief: A generated, dark editorial scene visualizing one continuous long-horizon work path from technical planning through verification to a finished professional deliverable. Deterministic overlay: “GROK 4.7” and “LONG-RUNNING CODING + KNOWLEDGE WORK,” with the verified standard API price as a small supporting line.
xAI released Grok 4.7 on September 21 as a model aimed at the part of AI work that is difficult to show in a short demo: staying on a complicated task for hours, checking the result, and carrying enough context to finish rather than merely begin. The @SpaceXAI launch post presents the release as a step up from Grok 4.6, while the linked official announcement supplies the benchmark table and access details. It says Grok 4.7 uses a larger base model than Grok 4.6 and a longer reinforcement-learning run weighted toward harder, multi-hour tasks. It also says the model was trained to understand the Grok Bot harness natively, with better performance on conversational and general knowledge work.
That is a more specific pitch than “a stronger chatbot.” Grok 4.7 is being positioned as an agent model for software engineering, documents, presentations, and professional workflows. It is available in Cursor, Grok Build, the Grok API, third-party coding harnesses, model routers, and cloud platforms.
The scorecard is strong, not clean-sweep territory
The comparison below is reproduced from xAI’s launch page. These are vendor-reported results, not an independent rerun. The settings are not identical: Grok 4.7 is shown at xhigh effort, Grok 4.6 at high, and GPT-5.6 Sol and Fable 5.1 at max.
| Benchmark | Grok 4.7 | Grok 4.6 | GPT-5.6 Sol | Fable 5.1 |
|---|---|---|---|---|
| CursorBench 4.0 | 46.3% | 40.4% | 41.7% | 51.8% |
| DeepSWE v1.1 | 71.0%* | 65.2% | 72.7% | 70.0% |
| EEBench | 64.0% | 53.0% | 39.4% | 56.4% |
| AA Briefcase v1.1 | 1,657 | 1,546 | 1,487 | 1,678 |
| Terminal-Bench 4.0 | 38.0% | 20.3% | 37.3% | 57.9% |
| Harvey Legal Agent Benchmark | 19.6% | 15.8% | 2.5% | 6.7% |
| HealthBench Professional | 56.7% | 48.5% | 60.5% | 62.1% |
The DeepSWE result is marked as high effort by xAI.
The first useful result is the generation-over-generation one. Grok 4.7 beats Grok 4.6 in every row in the table. The biggest percentage-point moves are Terminal-Bench, up 17.7 points, EEBench, up 11 points, and HealthBench Professional, up 8.2 points. On AA Briefcase, the model rises from 1,546 to 1,657. That is a meaningful pattern for a release whose central claim is longer, more involved work.
The second result is the limit on the marketing. Grok 4.7 does not lead every named competitor. Fable 5.1 is ahead on CursorBench, AA Briefcase, Terminal-Bench, and HealthBench Professional. GPT-5.6 Sol is ahead on DeepSWE and HealthBench. Grok 4.7 leads this table on EEBench and Harvey Legal, while landing close to GPT-5.6 Sol on Terminal-Bench.
That mix matters because it separates “better than the last Grok” from “best at every kind of work.” Grok 4.7’s case is strongest when the buyer values a broad, relatively cheap model that can sustain coding and office tasks. The vendor table does not establish universal dominance.
Professional work is a central target
xAI’s second chart covers GDPval, AA Briefcase, and EEBench. In the GDPval view, Fable 5.1 scores 1,735, Grok 4.7 scores 1,695, Grok 4.6 scores 1,605, and GPT-6 Astra scores 1,542. In the AA Briefcase view, Fable 5.1 scores 1,678, Grok 4.7 1,657, GPT-6 Astra 1,569, and Grok 4.6 1,546.
The model therefore moves much closer to Fable 5.1 on the office-work measure than the predecessor did, but it does not pass it in xAI’s own chart. xAI describes these evaluations as work done by professionals including lawyers, nurses, and financial analysts, with documents and presentations among the target outputs. That description is useful context, but the page does not provide enough methodology to treat the scores as a general measure of job performance.
There is another important distinction in the launch material. The official CursorBench chart labels Grok 4.7 Extra High at 46.3% and gives it an average cost of $6.01 per task. That cost figure belongs to xAI’s benchmark presentation. It is not an independent audit of what every Cursor or API workflow will cost, and it should not be confused with the public token tariff.
What the API actually offers
The model documentation lists grok-4.7 with a 500,000-token context window, configurable reasoning, text and image input, and text output. The documentation also gives the model a May 2026 knowledge cutoff. For current events or current external data, xAI says developers must enable server-side Web Search or X Search tools. The model’s context length is large, but it is not a substitute for retrieval or a live-data connection.
The standard price is $2 per million input tokens and $6 per million output tokens, the same list price xAI shows for Grok 4.6. A fast variant is offered at twice the output speed and twice the standard price. xAI does not present the fast variant as a separate model family in the launch prose, so the practical detail to confirm before deployment is how the selected provider exposes that mode and bills it.
For a simple synthetic workload of 1 million input tokens plus 1 million output tokens, the raw list-price arithmetic is:
- Grok 4.7: $2 + $6 = $8
- GPT-5.6 Sol: $4 + $20 = $24
- Fable 5.1: $10 + $50 = $60
This is not a forecast of a real agent run. It excludes cached-input discounts, reasoning-token treatment, tool calls, subscription fees, platform markups, and the extra output that a verbose model may generate. Artificial Analysis currently records Grok 4.7 as a 500K-context model and gives the xhigh variant an Intelligence Index score of 46.45, rounded to 46. Its page shows a median score of 11 across the comparison set. It also reports that the evaluation run generated about 240 million output tokens, compared with a 92 million-token median, a reminder that output volume can shape both cost and latency.
As of September 21, Grok 4.7 did not appear on the public Agent Arena leaderboard. That is not evidence that the model has no agent benchmark results. It means xAI’s scorecard should not be mistaken for a consensus ranking while independent agent results are still pending.
Safety claims are part of the launch
xAI says Grok 4.7 uses a new safeguard stack and is its strongest model so far on refusal and jailbreak-resistance testing. It reports a 62.4% result on LatchBio’s biosafety benchmark. On HackerBench v0.3, which xAI describes as a benchmark for risky and malicious cyber tasks, the company says only 3.3% of risky dual-use prompts were allowed through while legitimate security work was rarely blocked. xAI also says select cybersecurity partners have invite-only access to red-team capabilities for defense research.
Those figures are also vendor-reported. They are relevant because a model designed for persistent tool use has a larger safety surface than a one-shot chat model, but the launch page does not publish enough detail to independently reproduce the refusal-rate or benchmark claims. The safe operational conclusion is narrower: xAI is making safety and cyber-defense behavior a release requirement, and buyers should test the model against their own tool permissions, logging, approval gates, and high-risk domains.
The practical read
Grok 4.7 looks like a serious value play for teams that want a single model to handle code, documents, and general knowledge work without moving to a $10/$50 or higher frontier-model tariff. Its predecessor-to-successor gains are broad, the context window is 500K, and the standard API price remains $2/$6. Cursor and Grok Build make the release immediately relevant to developers, while the API and router availability give teams more deployment choices.
The benchmark evidence still calls for discipline. Grok 4.7 improves on Grok 4.6 across the launch table, but Fable 5.1 wins several of the rows tied most closely to long-running terminal and professional work. GPT-5.6 Sol leads DeepSWE and HealthBench in the same comparison. The effort settings differ, and the independent agent leaderboard checked does not yet show a Grok 4.7 entry.
For a buyer, the next test is not another slogan. Run the model on the actual repository, document set, browser/tool permissions, and review process that matter to the team. Measure completed tasks, correction loops, tool-call failures, output volume, latency, and total spend. Grok 4.7’s release makes that experiment cheaper to run. It does not remove the need to run it.