Grok 4.7 ties GPT-6 Astra's cheapest setting, and costs $2.73 per task against $0.82
SpaceXAI's new Grok 4.7 scores 46 on the Artificial Analysis Intelligence Index, the same mark the leaderboard gives GPT-6 Astra at its lowest effort setting. Running the same test set cost $2.73 with Grok 4.7 and $0.82 with Astra, so the model SpaceXAI sells at $2 per million input tokens runs up more than three times the bill.
Grok 4.7 runs in Grok Build, the harness Terminal-Bench used for its 37.6% score. Image: SpaceXAI
Grok 4.7 arrived on September 21 at $2 per million input tokens and $6 per million output tokens, against $4 and $20 for GPT-5.6 Sol and $10 and $50 for Claude Fable 5.1 in the company's own comparison. SpaceXAI says the model is "served at the same price and speed as Grok 4.6", and it sells a fast variant at twice the output speed for twice the price.
Cheap tokens do not make a cheap task
Cost per task is what Artificial Analysis spends running its own test set, so the per-million rate is only part of the bill. A model that thinks for longer or takes more steps can run up a larger total from a lower rate.
Push Grok 4.7 to its xhigh setting and the score stays at 46 while the cost per task rises to $3.74. GPT-6 Astra at maximum effort reaches 53 for $3.26, and Claude Fable 5.1 reaches 53 for $5.98. The same leaderboard lists Grok 4.7 with a 500k context window, where both of the others are on 1M.
Terminal-Bench 4.0 ranks it 17th
Terminal-Bench's own leaderboard has Grok 4.7 at xhigh effort, running in Grok Build, resolving 37.6% of tasks with a margin of 3.5 points. That is 17th of 27 entries, with Z.ai's GLM-5.3 one place above it on 41.8% and GPT-6 Astra on Codex at the top on 58.2%.
SpaceXAI publishes 38.0% for the same benchmark, which sits inside the leaderboard's margin, so the two numbers agree. Terminal-Bench also prices each run, and it spent $3,683 on Grok 4.7, across 324 of the 330 trials, against $3.3k on GPT-6 Astra at maximum effort and $2.3k on the same model at high effort, which still scored 57.9%.
The company's comparison table runs Grok 4.7 against Grok 4.6, GPT-5.6 Sol and Claude Fable 5.1 rather than GPT-6 Astra, which cracked a 1941 Enigma message in September. Against that set, SpaceXAI reports 46.3% on CursorBench 4.0, 71.0% on DeepSWE v1.1 at high effort, 64.0% on EEBench, 19.6% on the Harvey Legal Agent Benchmark and 56.7% on HealthBench Professional.
Astra does turn up in one SpaceXAI chart. On GDPval, the company puts Grok 4.7 at xhigh on 1,695 Elo against 1,542 for GPT-6 Astra at maximum effort, with Claude Fable 5.1 ahead of both on 1,735.
SpaceXAI says the safeguards are its best yet
SpaceXAI says Grok 4.7 "is the strongest model we've tested on refusals and jailbreak resistance", and that it tops LatchBio's biosafety benchmark at 62.4%. On HackerBench v0.3, the company's own test for risky and malicious cyber tasks, it says the model lets through 3.3% of risky dual-use prompts while rarely blocking legitimate security work.
Both figures come from SpaceXAI, and HackerBench is SpaceXAI's own benchmark. The company says it has started giving selected cybersecurity partners invite-only access to the model's red-team capabilities for defence research.
Where you can run it
Grok 4.7 is live in Cursor and in Grok Build, and SpaceXAI says it also runs through the Grok API, third-party coding harnesses, model routers and cloud platforms. The company describes it as a larger base model than Grok 4.6, trained with a longer reinforcement learning run on a harder mix of tasks, weighted towards problems that take many hours.
The $2 input price is the clearest thing Grok 4.7 holds over GPT-6 Astra. Artificial Analysis's own bill suggests it does not always survive the task.
