Grok 4.5: xAI's coding model, reviewed honestly

The short answer

Grok 4.5 is xAI's first model built specifically for coding and agentic work, released on 8 July 2026 at $2 per million input tokens and $6 per million output. It is not the most capable model available — it landed fourth on the Artificial Analysis Intelligence Index at launch with a score of 54 — but it costs roughly a fifth of what the models above it charge.

That trade is the entire pitch, and for a specific kind of work it is a good one: long agent runs where token consumption compounds, and high-volume coding tasks where a small capability gap costs less than a large price gap.

Figures reflect xAI's published pricing and reported benchmark results at the time of writing (26 July 2026). Leaderboard positions in this field change within weeks.

What makes it different

Trained on real developer sessions

Grok 4.5 is a mixture-of-experts model trained on trillions of tokens of real Cursor developer session data — actual people working in actual codebases, rather than general web text and public repositories. The bet is that watching how developers iterate teaches a model something that reading finished code does not: what a half-broken change looks like, and what people do next.

Benchmarked on long sessions, not trivia

xAI leaned on evaluations built to measure what a model can do inside a real repository over an extended session, rather than one-shot puzzle solving. That mirrors the wider 2026 shift — every lab has moved from knowledge quizzes toward agentic benchmarks, because agentic work is where the money is and where models still visibly fail.

Token efficiency as a feature

The most under-discussed number: Grok 4.5 reportedly uses 3–4x fewer tokens per task than the most expensive rivals while matching GPT-5.5 in the Codex harness on the Coding Agent Index (76). Combine that with a per-token price a fifth of theirs and the effective cost difference on a long agent run is large — this is a bigger deal than the headline price alone suggests.

The numbers, in context

For comparison, OpenAI reported Sol Ultra at 91.9% and base Sol at 88.8% on that same Terminal-Bench 2.1 benchmark. So Grok 4.5 is meaningfully behind the frontier on agentic coding — roughly five points behind base Sol — while costing $2/$6 against Sol's $5/$30.

Whether that trade is good depends entirely on your failure tolerance. Five percentage points on a benchmark means something different when you are generating throwaway scripts than when you are running an unattended agent against production code.

When Grok 4.5 is the right call

Use it when: you are running high-volume coding or agentic tasks where per-task cost is the binding constraint; your workflow reviews output before it lands anywhere important; or you are running long agent sessions where the token-efficiency advantage compounds.

Use something else when: the task is a one-off where a few dollars of tokens is irrelevant and you just want the best answer — use GPT-5.6 Sol or Claude Fable 5. Or when you need very long context, where Claude Opus 5's 1M-token window is the differentiator.

If you are not a developer: Grok 4.5 is not aimed at you. It is a coding and agent model, and its strengths show up in tooling rather than chat. For general everyday use, the assistant you already know well will out-perform a new model you don't — the skill gap between users is consistently larger than the capability gap between the top few models.