Grok 4.7 chases the default agent slot, not the benchmark crown

A launch aimed at agent distribution

xAI released Grok 4.7 on September 21, 2026, and the most revealing detail is not a benchmark number but where the model shows up. According to a dev.to analysis, Grok 4.7 is available through the xAI API, Cursor, xAI's Grok Build, a GitHub Copilot rollout, and third-party gateways. That placement frames the launch: xAI is competing for the slot your coding agent routes to by default, not for the top line on a general leaderboard.

A split benchmark story

Artificial Analysis scored Grok 4.7 at 46 on its Intelligence Index when running at the highest "xhigh" reasoning effort — two points above Grok 4.6 at high effort, and below the leading score of 53 cited in the write-up. On its own, that is not a frontier-model victory.

The picture changes inside agent products. On Artificial Analysis's Coding Agent Index, measured with the Grok Build harness, the model reaches 56, up from 47 for Grok 4.6. Per the API documentation summarized by dev.to, the model ID is grok-4.7, with a 500K-token context window, text and image input, four reasoning effort levels, support for both Responses and Chat Completions APIs, and a May 2026 knowledge cutoff. A faster-serving variant appears in Cursor and Grok Build at roughly twice the token price, but is not offered through the public API.

The harness is part of the score

The same model produced three different Terminal-Bench 4.0 results at launch: 38.0% in the xAI model card using the Grok Build harness (results attributed to Harbor), 33% in Artificial Analysis's native Grok Build setup, and roughly 26% under the evaluator's standardized harness.

These numbers answer different questions, the dev.to piece argues. Native-harness scores approximate what a product stack delivers; standardized scores isolate model capability under a common protocol. Because an agent is a stack — model, system prompt, tool schema, repository context, planning loop, error recovery, verification policy — a bare-model score cannot predict the behavior of the system you will actually deploy. The practical advice is to benchmark the exact combination you plan to ship.

Gains concentrate in long-horizon work

The strongest improvements show up in tasks that require sustaining a plan and producing a deliverable. On AA-Briefcase, which evaluates realistic professional work products, Grok 4.7 reaches 1657 Elo, 111 points above Grok 4.6 high; on GDPval-AA it reaches 1695 Elo, a 90-point gain. With Grok Build, all three components of the Coding Agent Index improve: DeepSWE from 65% to 73%, Terminal-Bench 4.0 from 18% to 33%, and SWE-Atlas-QnA from 58% to 63%.

Artificial Analysis also reports a lower hallucination rate on AA-Omniscience — 29%, down from 34% — while raw accuracy stayed roughly flat, suggesting better restraint on that evaluation rather than a general fix for hallucinations. Reasonable candidate workloads include repository-scale implementation and debugging, terminal tasks needing tool use and recovery, and research that ends in a structured document, spreadsheet, or presentation.

Output tokens are the hidden cost variable

xAI's list pricing is $2 per million input tokens, $0.50 for cached input, and $6 per million output tokens below the 200K-token long-context threshold, rising to $4 and $12 above it. Those rates look competitive, but per-token price is only one term in the bill. Artificial Analysis reports Grok 4.7 at xhigh consumed roughly 81K output tokens per Intelligence Index task, versus 36K for Grok 4.6 at high effort (the article's chart shows 38K for Grok 4.6 at xhigh) and about 27K for GPT-6 Astra max.

The dev.to analysis therefore proposes an accepted-task cost metric: total spend across attempts, including reasoning tokens, retries, timeouts, and failed outputs, divided by the number of results that pass acceptance testing. A cheap-listed model that loops or needs retries can cost more than an expensive one that finishes cleanly.

The post also notes that gateway provider BeatAPI lists the public grok-4.7 model at 35% of xAI's published rates across both standard and long-context rows, exposed through OpenAI, Anthropic, and Gemini-compatible request formats — while cautioning that the same accepted-task economics still determine the final bill.

Why it matters

Grok 4.7 illustrates a shift in how model launches should be read. General-purpose index scores matter less than performance within the harness an agent actually runs, and vendors increasingly ship model-plus-harness as a single product. For agent developers, three consequences follow: benchmark the exact stack you deploy rather than the bare model; expect first-party harnesses to amplify (or mask) capability in ways leaderboards do not capture; and model costs on a per-accepted-task basis, since reasoning-heavy models can quietly triple output-token consumption. xAI is betting that distribution and integration depth, not a single crown score, decide which model becomes the default inside your tooling.