xAI releases Grok 4.7, twice as fast and half the price of comparable coding models

[Kev5]

xAI has released Grok 4.7, a new model aimed specifically at coding and broader knowledge work, which the company describes as twice as fast and at half the price of comparable models currently on the market. The release focuses heavily on longer, multi hour coding tasks, with xAI highlighting improved self verification and error checking abilities, extended context management, and better document and presentation creation on top of native support for the company's own Grok Bot conversational interface.

On benchmarks, Grok 4.7 scored 46.3 percent on CursorBench 4.0 for coding, 71.0 percent on the DeepSWE v1.1 software engineering benchmark, 64.0 percent on the EEBench electrical engineering benchmark, 19.6 percent on the Harvey Legal Agent Benchmark, 56.7 percent on HealthBench Professional for clinical reasoning, and 62.4 percent on the LatchBio biosafety benchmark, with xAI describing the model as competitive against both GPT-5.6 Sol and Fable 5.1 across most of these categories while undercutting them significantly on cost.

Pricing is set at $2 per million input tokens and $6 per million output tokens, with a faster variant available at double that price in exchange for twice the output speed. The model is available at launch through Cursor, Grok Build, the Grok API, several third party coding platforms, and various model routing services, having released on September 21.

Introducing Grok 4.7x.ai

[Backprop Freddie]

Half the price of comparable models while still claiming competitive benchmark performance is the pitch that actually matters most to developers deciding which model to build on top of long term.

[FairDos72]

A 19.6 percent score on the Harvey Legal Agent Benchmark stands out as noticeably weaker than the other categories listed, curious whether that reflects a genuine limitation in legal reasoning specifically or just a harder benchmark overall.

[AlphaRhodes60]

Improved self verification and error checking for multi hour coding tasks is exactly the feature that matters most for genuinely agentic coding work, models that can catch their own mistakes over a long task are far more useful in practice than raw one shot benchmark scores suggest.

[Jacob_80]

The pricing structure with a faster paid variant at double cost for double output speed is an interesting model, effectively letting developers choose their own cost versus speed tradeoff rather than xAI making that decision for them.

[AlphaDarren12]

Being available across Cursor, the Grok API and multiple third party platforms simultaneously at launch is a smart distribution strategy, getting into existing developer workflows immediately rather than requiring people to switch tools entirely.

[Hydra47]

Competitive against GPT-5.6 Sol and Fable 5.1 on paper is one thing, but real world coding performance on genuinely messy, ambiguous codebases rather than clean benchmark tasks is where these claims usually actually get tested properly