xAI kept the price flat and rewrote the model underneath. Grok 4.7 doubles its predecessor's terminal-coding score, cuts hallucinations, and uses a lot more tokens getting there. We lined up the vendor table against independent testing to see which upgrades hold up.
Two models, one price tag
When a company charges the same for a new model as the old one, the interesting question isn't the price. It's what changed to justify calling it a new release.
For Grok 4.7, xAI's answer is a fresh base model. Grok 4.7 doesn't reuse the base behind Grok 4.6, and xAI trained it with a longer reinforcement learning run aimed at problems that take hours rather than seconds. Both models cost $2 per million input tokens and $6 per million output tokens, and both list a 500,000-token context window.
What that buys you depends heavily on the task, and the two sets of published benchmarks tell slightly different versions of the story.
Grok 4.7 | Grok 4.6 | |
|---|---|---|
Price (input / output) | $2 / $6 per 1M tokens | $2 / $6 per 1M tokens |
Base model | New, larger | Previous generation base |
Training | Longer RL run on harder tasks | Shorter RL run |
Context window | 500K tokens | 500K tokens |
Terminal-Bench 4.0 | 38.0% | 20.3% |
CursorBench 4.0 | 46.3% | 40.4% |
EEBench | 64.0% | 53.0% |
Hallucination rate | 29% | 34% |
Omniscience accuracy | 47% | 48% |
Terminal-Bench doubles, and that's the headline
xAI's own table puts Grok 4.7 ahead of Grok 4.6 on all seven benchmarks it published. The size of each gap is the useful part.
Terminal-Bench 4.0 nearly doubles, from 20.3% to 38.0%. It tests whether a model can drive a command line and finish a task in a terminal, which is the kind of work where a model either gets the sequence right or leaves the job half-finished. A jump that size in one release is rare.
EEBench climbed 11 points to 64.0%, and there Grok 4.7 takes the top spot in xAI's table. Legal work moved from 15.8% to 19.6% on the Harvey benchmark. CursorBench 4.0 rose about six points.
Coding-adjacent tasks trended the same way. On DeepSWE v1.1, Grok 4.7 scored 71.0% at high effort against 65.2% for 4.6 at the same setting.
What independent testing says
xAI's numbers are vendor-reported, so Artificial Analysis ran its own evaluation. The picture that comes back is more measured.
Grok 4.7 scores 46 on the Intelligence Index, two points above Grok 4.6. That's real progress, but it sits well behind Claude Fable 5.1 and GPT-6, both at 53. On long-horizon professional work the gap narrows: AA-Briefcase put Grok 4.7 at 1,657 Elo, up 111 points and just behind Claude Opus 5 and Fable 5.1. GDPval-AA rose 90 points to 1,695.
Coding agents moved too. Paired with Grok Build, Grok 4.7 scored 56 on the Coding Agent Index, up nine points. Under that setup DeepSWE improved by eight points to 73%, and Terminal-Bench reached 33%, up from 18%.
Then there's the token count, which is where the flat price gets complicated. Grok 4.7 uses about 81,000 output tokens per Intelligence Index task. Grok 4.6 used around 38,000 at the same xhigh effort. GPT-6 Astra used 27,000. Double the tokens at the same per-token rate means double the bill for a comparable job.
Hallucinations down, accuracy flat
The most useful reliability change is easier to miss than a benchmark jump.
Artificial Analysis measured Grok 4.7's hallucination rate at 29%, down from 34% for Grok 4.6. That's the share of questions where the model invents an answer instead of admitting it doesn't know. Accuracy barely moved, from 48% to 47%, and the combined Omniscience Index rose from 30 to 32.
Read those together and the story is clear. Grok 4.7 doesn't know more facts than 4.6. It just makes up fewer of them, and it says "I don't know" more often. For anyone using a model to pull numbers out of documents or answer factual questions, guessing less is the upgrade that matters, even when the headline accuracy number looks frozen.
Where each version still makes sense
Grok 4.6 didn't get worse. It uses roughly half the output tokens of 4.7 on the same tasks, so at identical token prices it can come out cheaper for work that doesn't need long agent runs. If your prompts are short and your tasks finish in one pass, 4.6 remains a reasonable pick.
Grok 4.7 earns its keep on jobs that run for a while. Terminal work, multi-step coding, anything where the model has to plan, act, check itself, and correct course. That's the exact territory where the new base model and the longer training run show up, and where the doubled terminal score comes from.
One caveat on the big claim: reports have described Grok 4.7 as a 2.1 trillion parameter model. Musk hasn't confirmed that number, and xAI's docs don't publish parameter counts. Don't treat it as settled.
Both models are live now. Grok 4.7 is available through the xAI API, Cursor, and Grok Build, and it's the default in Grok Build. The practical move is to test it on your longest-running tasks first, since that's where the gains cluster, and keep an eye on token usage, which is where the flat price can turn into a higher bill.
