Musk’s xAI launches Grok 4.7 with 2 trillion parameters and SpaceX data – and it still can’t beat Claude
Elon Musk’s xAI today released Grok 4.7, calling it “a notable improvement over Grok 4.6 at the same price and speed” and its most capable model yet for coding and knowledge work. It’s bigger, it’s cheap, and it was partly trained on data from SpaceX. What it isn’t, by the benchmarks xAI itself published, is the best model on the market.
What’s new
Grok 4.7 uses a larger base model than Grok 4.6, with around 2.1 trillion parameters, up roughly 40% from its predecessor’s 1.5 trillion, according to Decrypt. xAI also gave it a longer reinforcement-learning run aimed at tasks that take hours to complete, and says the model “works longer on difficult tasks” and “checks its own work more carefully.”
The more unusual ingredient is SpaceX. xAI folded in supplemental training data from Musk’s rocket company, including Starlink satellite telemetry, manufacturing records and engineering failure logs.
xAI also says Grok 4.7 handles long context better and is better at producing documents and presentations.
The benchmarks
xAI’s own numbers show Grok 4.7 improving on Grok 4.6 across the board, including:
- DeepSWE v1.1 (software engineering): 71.0% at high effort
- CursorBench 4.0: 46.3%
- HealthBench Professional: 56.7%
- GDPval (professional knowledge work): 1695 Elo
Against the competition, though, it keeps finishing second. On GDPval, Anthropic’s Claude Fable 5.1 scores 1735. On independent tracker Artificial Analysis’s Intelligence Index, Grok 4.7 scores 46 to the 53 of both Fable 5.1 and OpenAI’s GPT-6, according to TBreak. And on Terminal-Bench 4.0, an agentic coding test, the gap is stark: 26% for Grok 4.7, against 55% for Fable 5.1 and 60% for GPT-6.
That won’t surprise anyone who was listening to Musk. Ahead of launch, he set expectations at “roughly on par with” Anthropic’s Claude Opus 5, not the newer 5.1 models, Decrypt reports.
The real pitch: price
Where Grok 4.7 competes is cost. It’s priced at $2 per million input tokens and $6 per million output tokens, the same as Grok 4.6. A “fast” variant runs at twice the output speed for twice the price.
That fits xAI’s long-running strategy of being good enough, cheap and everywhere rather than leading the frontier. Grok 4.7 is available now through the Grok API, Cursor, xAI’s Grok Build coding tool, and a range of third-party coding tools, model routers and cloud platforms.
Safety
xAI says Grok 4.7 ships with a new set of safeguards with strong resistance to jailbreaks. On its HackerBench v0.3 test, it lets through just 3.3% of risky dual-use hacking prompts. It also reports a score of 62.4% on the LatchBio biosafety benchmark, and says more capable “red team” access will be offered only to select cybersecurity partners.
What’s next
Musk has already sketched out the road ahead: Grok 4.8 as a meaningful upgrade, Grok 4.9 to match “Astra/Fable class” models from OpenAI and Anthropic, and Grok 5 as a potential frontier leader. He hasn’t said when any of them will arrive.