This week everyone in AI was slashing prices. SpaceXAI, Elon Musk's AI company, took a different road: on September 21, 2026 it released Grok 4.7, a bigger, smarter model — at exactly the same price as its predecessor.
The company calls it its most capable model yet for coding and knowledge work, built on a new, larger base model than Grok 4.6 and trained with a longer reinforcement learning run weighted toward tasks that take hours to finish rather than seconds. The stated goal: a model that sticks with hard problems longer, double-checks its own output, and handles long context windows without losing the thread.
What's new under the hood
Four changes separate Grok 4.7 from the last version: the new base model itself, a longer and harder RL training run, upgraded self-verification and long-context handling, and native training on the Grok Bot harness, which SpaceXAI says improves conversational tasks and general knowledge work — documents and presentations included, not just code.
The numbers that matter
SpaceXAI's own benchmark table stacks Grok 4.7 against Grok 4.6, OpenAI's GPT-5.6 Sol, and Anthropic's Fable 5.1:
- CursorBench 4.0 (long-running coding tasks): 46.3%, up from 40.4% on Grok 4.6, ahead of GPT-5.6 Sol's 41.7% — though still behind Fable 5.1's 51.8%
- Terminal-Bench 4.0: 38%, nearly double Grok 4.6's 20.3% — a big jump in agentic coding reliability
- DeepSWE v1.1: 71.0% at high effort; EEBench: 64%; Harvey legal benchmark: 19.6%, ahead of the other models in its table
- AA Briefcase (documents and presentations): 1,657, up from 1,546 — just short of Fable 5.1's 1,678
It doesn't lead everywhere — HealthBench Professional trailed the other two flagships — but the direction is clear.
Same price, plus a Fast variant
Developers pay the same $2 per million input tokens and $6 per million output tokens as Grok 4.6 — and SpaceXAI claims that means roughly twice the value at half the price of comparable models. A Grok 4.7 Fast variant offers twice the output speed at twice the price in Cursor and Grok Build. Grok Build also offers free access to try the model.
It's live today in Cursor and Grok Build, and through the Grok API, third-party coding harnesses, model routers, and cloud platforms including OpenRouter, Vercel, and Cloudflare.
Safety claims
The release also includes a new "safeguard stack" — SpaceXAI's best-calibrated safeguards to date, per the company — scoring 62.4% on LatchBio's biosafety benchmark, with select cybersecurity partners getting invite-only access to its red-teaming capabilities for defensive research.
Why it matters
For teams building on Grok, this is the rare upgrade that doesn't force a budget conversation: same price, same speed, but measurably smarter. A Terminal-Bench jump from 20.3% to 38.0% is the kind of number that changes what you'd trust a model to do unsupervised in a terminal — and that's the benchmark that matters most for agentic coding tools like Cursor and Grok Build, where xAI already has distribution locked in.
Explore more AI tools and launches at aipost.tech.
Comments
Post a Comment