SpaceXAI says Grok 4.7 scores 38 percent on Terminal-Bench 4.0

An agent cheap enough to leave running all night still fails most of the jobs it is handed. SpaceXAI released Grok 4.7 on 21 September 2026 as a coding model built for tasks that run for hours. Its own announcement calls it "twice as fast, at half the price of comparable models", and prices the API at $2 per million input tokens and $6 per million output tokens.

What the training changed

SpaceXAI says Grok 4.7 is the same base model as Grok 4.6, retrained through a longer reinforcement learning run that deliberately weighted harder problems, including tasks the company describes as taking "many hours" to complete. The company credits the gain to two capabilities: self-verification, so the agent catches its own wrong turns before building on them, and long-context management, so a growing interaction history stays usable.

SpaceXAI has not explained how either capability was measured, or whether context handling comes from architecture, summarisation, retrieval or something else. Amanda Caswell at The New Stack noted on 21 September that "the company did not disclose whether the context gains came from architectural changes, summarization, retrieval, or better retention across long sequences".

The scores it is claiming

On the vendor's own numbers, Grok 4.7 scores 38.0 percent on Terminal-Bench 4.0, up from 20.3 percent for Grok 4.6. On CursorBench 4.0, an in-editor multi-step coding benchmark, it moves from 40.4 to 46.3 percent. On AA Briefcase v1.1, a multi-hour professional work test, it goes from 1,546 to 1,657. SpaceXAI also reports 71.0 percent on DeepSWE v1.1, 64.0 percent on EEBench and 19.6 percent on the Harvey Legal Agent Benchmark.

For a reference point outside SpaceXAI's own numbers, The New Stack cites the independent Terminal-Bench leaderboard, on which Anthropic's Claude Fable 5.1 scores 57.9 percent. Grok 4.7 still trails Fable 5.1 by nearly 20 points on that test.

The harness is now part of the model

SpaceXAI trained Grok 4.7 to understand the Grok Bot harness directly. That is its terminal coding agent, which feeds context, tool definitions and results back into the model. Training on the harness instead of prompting it at runtime is meant to cut the overhead in tool use.

That direction mirrors OpenAI, which opened its Codex harness as the Agents API last week to move the surrounding infrastructure into a managed service. When a model is trained around a specific tool format and execution environment, swapping models without sacrificing agent performance becomes harder, and the decision about which harness to adopt starts to lock in a set of models along with it.

The jump is larger than the release explains

A jump from 20.3 to 38.0 percent on the same benchmark, from the same base model, is close to a doubling, and SpaceXAI has attributed it to two vague-sounding levers with no numbers behind either. A separate recent benchmark of private codebases had the best model failing more than 60 percent of the time, so the reliability floor for unattended agents is still low. Read the deltas as vendor claims that need independent reproduction.

For a team weighing an unattended overnight run, the practical question is whether Grok 4.7 clears the threshold at which the run is worth kicking off. At $6 per million output tokens with self-verification claims that can be tested against a public benchmark, that is a decision worth an afternoon of setup.

Source

SpaceXAI, Grok 4.7, 21 September 2026. Reporting and independent Terminal-Bench context from Amanda Caswell, The New Stack, Grok 4.7 was built to work for hours. It still fails most of the time., 21 September 2026.