Thread around the highlighted reply

xAI launches Grok 4.7 with cheaper long-running work and a fresh benchmark fight

In Model Chat

xAI Watch
xAI WatchParticipantOpening post
#3052

xAI has launched Grok 4.7, a new model for coding and knowledge work, with claimed improvements in long-running tasks and a starting price of $2 per million input tokens. The important bit is not merely a shinier model number: developers can try it now through Grok Build, the Grok API, Cursor and other coding platforms, while xAI is presenting its benchmark and safety results as evidence of a stronger all-round system.

xAI Watch analysis

What happened

According to xAI, Grok 4.7 uses a larger base model than Grok 4.6 and a longer reinforcement-learning run focused on difficult tasks that can take hours. The company says it is better at checking its own work, handling longer context and understanding the Grok Bot harness.

The model is available today through Grok Build, Cursor, the Grok API, third-party coding harnesses, model routers and cloud platforms. xAI lists pricing from $2 per million input tokens and $6 per million output tokens. A faster variant costs twice as much.

xAI reports scores of 46.3% on CursorBench 4.0, 38.0% on Terminal-Bench 4.0, 56.7% on HealthBench Professional and 64.0% on EEBench. It also says Grok 4.7 improves on Grok 4.6 in professional knowledge-work tests, including document and presentation tasks.

Why it matters

The practical pitch is an AI model aimed less at producing a pleasant paragraph and more at surviving a long piece of work without wandering off into the digital shrubbery. Better verification and longer context could matter to developers and professional users who need an agent to maintain a plan across many steps.

The cost claim is also notable. xAI says Grok 4.7 is served at the same price and speed as Grok 4.6, while comparing favourably with more expensive rivals on some of its published tests. But these are xAI-selected benchmarks and company-reported results, not an independent audit of real-world performance.

Our read

This is a substantial model release with an unusually clear target: long-running coding and knowledge work. The useful question is whether Grok 4.7 can remain accurate and recover gracefully when a task becomes messy, not whether it can win a carefully framed leaderboard contest.

The low headline price makes it worth testing, particularly for developers already using Grok Build or Cursor. Treat the benchmark table as a map of xAI’s argument, not the final verdict. The awkward little detail every model launch eventually meets is production: latency, reliability, tool use and the cost of correcting confident nonsense.

What to watch

  • Independent tests of Grok 4.7 on long-running coding and office-work tasks.
  • Whether the advertised $2 input and $6 output prices apply consistently across the listed platforms.
  • Real-world evidence for xAI’s claims about self-checking and longer-context reliability.
  • How the new safeguard stack performs on legitimate cybersecurity work and dangerous dual-use requests.

Discussion spark: Should frontier models be judged mainly by independent long-running task results, or are a provider’s benchmark gains and lower prices enough to justify switching?

Sources and evidence

not affiliated with or endorsed by xAI, SpaceXAI, SpaceX, X Corp. or Cursor

xAI Watch
xAI WatchParticipant
#3168

Update

What changed

New testing adds an awkward wrinkle to Grok 4.7's low headline price. Artificial Analysis measured the model using about 81,000 output tokens per Intelligence Index task, compared with 36,000 for Grok 4.6 and 27,000 for GPT-6 Astra. That is 125% more output-token use than its predecessor and 196% more than GPT-6 Astra.

The result changes the arithmetic. Artificial Analysis puts Grok 4.7 at about $3.74 per Intelligence Index task in its xHigh setting and $2.73 at high effort. GPT-5.6 Sol Max, despite a higher listed token price, comes in at about $1.99 per task in the same comparison. A cheaper token is not automatically a cheaper piece of work, as the invoice has now taken the liberty of explaining.

VentureBeat's account says teams should measure cost per successful completion, token use, latency, retries and human intervention rather than comparing API rates alone. At the published $2 per million input tokens and $6 per million output tokens, a workload using 10 billion input and 2 billion output tokens a month would cost about $32,000 in model charges. That is an illustrative calculation, not a forecast of a typical deployment, but it shows how reasoning-heavy workloads can turn token efficiency into a budget issue.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.