Grok 4.6: self-checking during the run does not replace sign-off

On 12 August, SpaceXAI introduced Grok 4.6. The stated focus is long-running agents: the model is meant to stay with complex tasks across many steps, researching a topic, analysing information, working across a codebase, turning an idea into a finished application. One sentence in the announcement is more interesting than any benchmark row: on longer trajectories, the company says it started to see more self-testing and verification, with the model checking its own work before moving on. That is good news for the quality of the intermediate steps. It is not sign-off.

What is Grok 4.6?

Grok 4.6 builds on Grok 4.5, according to SpaceXAI, with a focus on long-running agents and on more ambitious interactive and visual work. It is available today in Cursor and Grok Build, plus the API and partners such as OpenRouter, Vercel and Cloudflare. Grok Build and Cursor include double the usage for the first week.

On training, the announcement names a longer supplemental run than for Grok 4.5, with curated model-generated data for reasoning and advanced technical concepts, high-quality engineering data, and an improved optimiser. The SFT trajectories were regenerated using Grok 4.5, with problematic traces filtered out by model-based checks. The agentic reinforcement learning that followed covers knowledge work, general coding and domain-specific environments such as kernel optimisation, web development and computer-aided design.

What does “more self-checking” actually mean?

It means the model tests what it just built more often along the way. SpaceXAI phrases this carefully, as an observation from its own test projects rather than a guaranteed property. Read it that way.

The structural point is untouched by it. Here, the check and the work come out of the same system. A model reviewing its own output reliably finds the errors it recognises as errors. A misread requirement survives every self-test, because the test was written from the same misreading. What improves is the hit rate against obvious defects. What does not improve is the question of whether the result was the right thing.

There is even a side effect worth planning for. A run that checked itself along the way arrives at your desk pre-checked. It looks more finished. The residue of errors is therefore subtler on average than before, and subtler errors need more review, not less.

What does the benchmark sentence say, and what not?

The prose contains exactly one comparison: Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, a composite score of nine benchmarks. A composite is an average, and averages hide the spread.

The eval table in the post, not the prose, shows that spread. On DeepSWE v1.1 it lists 65.9% for Grok 4.6 High against 73% for GPT-5.6 Sol Max. On CursorBench v3.2 the two sit the other way round, 69.9% against 67.2%. For your decision, none of these rows is the answer, because none of them contains your repository. They only say the choice depends on the task profile, not on a leaderboard.

What does a run cost?

Pricing starts at 2 US dollars per million input tokens and 6 US dollars per million output tokens, per the announcement, with a fast variant at twice the price. Nothing more is stated, and that is the real gap: a price per token says nothing about the cost of a completed task. How many tokens a run consumes is not given, and for a model built explicitly for long trajectories and extra self-testing, consumption is the number that decides the bill. Self-checking is not free. It costs output tokens.

How to work that out for yourself is in the piece on Grok 4.5 and the cost per completed task. The method does not change, only the model version in the measurement series.

What does this mean for your work?

Keep an acceptance condition that sits outside the run. A test you wrote, a check in CI, a review against the requirement. What the model checks by itself along the way is part of its work, not an assessment of it.

Use the strength where it is actually claimed. SpaceXAI describes the pattern of turning a broad product idea into a working first version, with research into unfamiliar domains, structure for the application, core interactions, and several rounds of feedback and refinement. For first passes on visual and interactive projects, SpaceXAI explicitly sees progress over Grok 4.5. That is a good place to spend it: early, disposable, quick to judge. A draft you can look at and throw away in two minutes needs no formal sign-off. Code that goes to production does.

The line stays the same as in agentic engineering: reliability comes from the architecture around the model. Grok 4.6 moves part of the checking into the run. That makes the run better. The person who says the result is good enough is still you.

Sources

• SpaceXAI: Introducing Grok 4.6