Grok 4.7 is now available for coding and knowledge work, with SpaceXAI keeping its base API price at $2 per million input tokens and $6 per million output tokens. The September 21 launch pairs a larger base model and longer reinforcement-learning run with new safety controls, but the most useful buyer question is not whether one benchmark crown changed hands; it is whether longer tasks finish with acceptable total token use.
Grok 4.7 changes the price-performance calculation
SpaceXAI says the model scores 46.3% on CursorBench 4.0, up from 40.4% for Grok 4.6 in its comparison. It also reports a 38.0% Terminal-Bench 4.0 score, versus 20.3% for the previous version. SiliconANGLE independently reported the launch and the same pricing, while VentureBeat highlighted the counterweight: a low token rate does not guarantee a low bill if an agent consumes substantially more tokens to finish.
The practical meaning of Grok 4.7 is that buyers can test a more capable long-horizon model without accepting a higher published base rate, but they still need task-level cost, completion and safety measurements of their own. That distinction matters for Indian software services teams, where a small rate advantage can disappear once retries, tool calls and review time are included.
| Item | Published detail | Buyer implication |
|---|---|---|
| Input | $2 per million tokens | Same base rate as Grok 4.6 |
| Output | $6 per million tokens | Measure completion cost, not rate alone |
| Context | 500,000 tokens in API docs | Long jobs still need context discipline |
What the safeguards claim does—and does not—prove
SpaceXAI says a new safeguard stack allowed 3.3% of risky dual-use prompts through on its HackerBench v0.3 evaluation and posted a 62.4% LatchBio biosafety result. Those are company-reported benchmark outcomes, not a blanket assurance for production. Security teams should preserve human approval for privileged actions, log tool calls and test refusal behaviour against their own threat model.
The release is available through the Grok API, Cursor, Grok Build and third-party model routes. Teams comparing orchestration approaches can also read Lapaas Voice’s coverage of the Salesforce enterprise AI harness and its analysis of agent cost intelligence.
A sensible evaluation should begin with a fixed set of real tickets, repositories and document tasks. Record successful completion, reviewer edits, latency and the full token bill, then compare the result with the model already in production. Keep the model behind a limited-permission service account during testing. This converts a broad launch claim into a procurement decision that can be defended, repeated and reversed if quality falls.
Procurement teams should also separate the standard and fast variants in testing because the fast option doubles the token rate. A model that returns sooner can still be economical when developer waiting time is costly, but that result depends on workload and staffing. The launch therefore creates a three-part comparison: output quality, end-to-end spend and elapsed time. None can be inferred reliably from the headline token price alone.
Frequently asked questions
What is Grok 4.7?
It is SpaceXAI’s September 21 model for coding, agentic tasks and knowledge work, offered through its API and selected products.
How much does Grok 4.7 cost?
The standard API starts at $2 per million input tokens and $6 per million output tokens below the long-prompt threshold; the documentation lists higher rates above 200,000 prompt tokens.
Should companies trust the benchmark table?
Treat it as a useful vendor disclosure, not an independent procurement verdict. Re-run representative tasks and record accepted completions, latency, tokens and reviewer corrections.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
