xAI shipped Grok 4.7 on 21 September 2026 at the same list price as Grok 4.6: $2 per million input tokens and $6 per million output. A free upgrade is a good deal. The more interesting number is the one almost nobody quotes: a 450,000-token maximum output, which is three and a half times what any frontier model from OpenAI or Anthropic will write in a single response.
What you get
Grok 4.7 carries a 500,000-token context window and accepts text, images and files. xAI reports improvements over Grok 4.6 across every benchmark in its own comparison table, and specifically leads on EEBench and the Harvey legal agent benchmark. It trails Claude Fable 5.1 on coding, which xAI does not hide.
Pretraining runs to June 2026 with supplemental data through August 2026. For anything where recency matters, that is among the more current knowledge cutoffs available right now.
One pricing detail that catches people out: the $2/$6 rate applies under 200,000 prompt tokens. At or above that threshold the whole request bills at $4/$12, not just the tokens above the line. If you routinely work near 200K, crossing it doubles the cost of the entire call.
Who the output ceiling is for
Most work never approaches a 128,000-token response. A long, detailed answer is a few thousand tokens. So for the majority of users, Grok's output ceiling is a spec-sheet curiosity.
It stops being a curiosity for bulk generation: translating or rewriting a long document in one pass, producing large volumes of structured output, generating extensive test fixtures or synthetic data. In those cases the alternative is chunking the work and stitching it back together, which costs you coherence at every seam. One long response avoids that.
Worth noting our own practical limit: on Deepest, interactive chat responses are capped well below any of these ceilings, because a 128,000-token response takes long enough to generate that it stops being a conversation. The headline output limits matter for API and batch work more than for chat.
Grok 4.7 against the week it launched
It arrived into a crowded 48 hours. Claude Opus 5.5 landed at $4/$20 and GPT-6 Sol at $2/$10. On input, Grok 4.7 and Sol are level at $2. On output, Grok is the cheapest of the three at $6, against Sol's $10 and Opus 5.5's $20.
So Grok 4.7 is the cheapest frontier option of the three on output, with the largest context-to-output envelope, and by xAI's own admission not the strongest coder. That is a coherent position rather than a weakness: it is priced and shaped for volume and long-form work rather than for topping coding leaderboards.
Should you switch from 4.6?
Yes, if you are on 4.6 with no evaluation pinned to it. Same price, better scores on xAI's own table, more recent training data. There is no argument for staying except the cost of re-running your own evals, which is a real consideration for production systems and not one to wave away.
On Deepest we retired Grok 4.6 when 4.7 arrived, and anyone who still had it selected now resolves to 4.7 automatically rather than hitting an error.
Frequently Asked Questions
Is Grok 4.7 good for coding?
It improved over 4.6, but xAI's own comparison shows Claude Fable 5.1 ahead on coding and long-horizon tasks. If coding is your main use, test Grok against a Claude model on your actual codebase rather than taking either vendor's word for it.
What does the 200K pricing threshold mean in practice?
Requests under 200,000 prompt tokens bill at $2/$6 per million. At or above that, the entire request bills at $4/$12. It is a cliff, not a gradient, so a prompt that creeps from 199,000 to 201,000 tokens doubles in cost. Worth watching if you work with large document sets.
Why do prices differ between xAI and resellers?
List price is what the provider publishes; what you actually pay depends on where you buy. Grok 4.7 lists at $2/$6 from xAI and currently routes cheaper through OpenRouter, which is where Deepest sources it. Always check the rate at your actual point of purchase rather than assuming the vendor's published number is what you will be billed.
