The Short Version
Grok 4.7 launched 21 September 2026 at $2 per million input tokens and $6 per million output. That is the whole argument, and it is a strong one for high-volume code generation.
GPT-5.6 Sol costs materially more and holds a lead that widens as tasks get longer and less bounded.
Where Grok 4.7 Holds Up
71.0% on DeepSWE v1.1 at high effort is a genuine software-engineering result, not a marketing number. 46.3% on CursorBench 4.0 puts it in serious contention for in-editor work. If your workload is generate, review, merge, repeat, the quality is there.
Where It Falls Away
38.0% on Terminal-Bench 4.0 against a strong DeepSWE score is the tell. Grok 4.7 is good at bounded tasks and less reliable once it has to drive a terminal across many steps without a human checkpoint.
19.6% on the Harvey legal agent benchmark is the clearest limit. Multi-step professional-domain agent work is not what this model is for.
The Token Consumption Problem
Several launch-day reviews flagged that Grok 4.7 burns a lot of tokens per task. A price that is one-third of a competitor stops mattering if the model needs four times the tokens to finish.
This is the single measurement that decides the comparison for your codebase, and nobody can run it for you. Take ten real tickets, run both, and compare total spend per merged pull request.
Decision Framework
- High-volume, well-scoped code generation - test Grok 4.7 seriously. The savings are real if your tasks finish in one pass.
- Long agentic sessions, terminal work, multi-hour tasks - GPT-5.6 Sol. The Terminal-Bench gap shows up as retries, and retries cost more than the price difference.
- Legal, clinical, compliance - not Grok 4.7. A 19.6% agent benchmark is a published limitation, not a quirk.
- Latency-sensitive interactive use - the Grok 4.7 fast variant at $4 and $12 is still under most frontier pricing and roughly doubles output speed.
Verdict
Grok 4.7 is the better buy for bounded coding volume, conditional on your token-per-task measurement coming out favourably. GPT-5.6 Sol remains the safer choice for anything that runs long, runs unattended, or carries professional liability.