Grok 4.7 targets longer coding tasks at Grok 4.6 prices

A new Grok model for sustained work

SpaceXAI has introduced Grok 4.7, positioning it for coding and professional knowledge tasks that run longer than a short chat exchange. The SpaceXAI announcement, dated 21 September, says the model uses a larger base than Grok 4.6 and a longer reinforcement-learning run weighted towards difficult, multi-hour problems. The claim is improved verification of its own work and better management of long context, not simply more fluent text.

The model is available through the Grok API, Grok Build, Cursor and supported third-party services. This is a direct model release, distinct from an earlier Grok 4.6 availability announcement on another cloud platform. Developers need to check the actual model identifier and endpoint they use because features and prices can differ between the public API and a coding product.

The announcement says Grok 4.7 is served at the same price and speed as Grok 4.6. That is a useful starting point, but it does not prove a migration is cost-neutral for an application. Longer prompts, more tool calls and extra retries can change the cost of a completed task even when the per-token rate is unchanged.

Reading the performance claims

SpaceXAI reports 46.3 per cent on CursorBench 4.0 for Grok 4.7 xHigh, against 40.4 per cent for Grok 4.6 High. It also lists a 71.0 per cent DeepSWE v1.1 result at high effort. These are vendor-presented comparisons, with different effort settings in some rows. They indicate what the company tested, not a guarantee that every codebase will see the same gain.

The company also highlights document and presentation work, drawing on evaluations intended to resemble professional tasks. A business should test the actual output it needs: whether a spreadsheet is correct, whether a slide deck preserves the source evidence and whether a code change passes tests. A benchmark number cannot account for a firm’s templates, permissions or review burden.

A practical trial would use a held-out set of bugs, refactors and document tasks from real work. Compare the final accepted result, elapsed time and human corrections with the current model. Include cases where the right response is to ask for more information or stop. That is a stronger basis for adoption than a single impressive demonstration.

Price and availability boundaries

The launch post lists prices starting at US$2 per million input tokens and US$6 per million output tokens. The official model documentation further distinguishes cached input and a higher rate above a long-prompt threshold. Buyers should read the current pricing table for their endpoint and service tier, and measure full task cost rather than multiplying only the cheapest short-context input price.

SpaceXAI describes a faster variant at twice the output speed and twice the price. Its documentation says this fast variant is available in Cursor and Grok Build, not on the public xAI API. Treating that option as universally available could lead to a deployment plan that cannot be executed. Teams should confirm model access before promising latency targets.

The public API documentation lists text and image input with text output, a large context window, and low through xHigh reasoning settings. Those features need explicit integration tests. An application that relies on structured responses, cached prompts or tools should check its full request and response cycle under the selected reasoning setting.

Safeguards are not a substitute for controls

SpaceXAI says it has introduced a new safeguard stack and reports improved refusal and jailbreak results. It cites tests in cybersecurity and biological domains, and invite-only defensive research access for selected partners. These are company evaluations. Operators still need their own policy tests for the topics, data and actions their users will bring to the system.

For a coding agent, tool permissions matter as much as the model. A better model may propose more ambitious edits or command sequences, but should not automatically receive authority to deploy, publish or change production data. Retain tests, approval points and a clear rollback path while comparing the new model with Grok 4.6.

Grok 4.7 is a meaningful release because it pairs a model update with direct API and coding-product availability. The strongest adoption case will come from verified improvements on hard, sustained work at acceptable total cost. Until then, the benchmark and safety figures are useful claims to examine, not a replacement for an organisation’s own evaluation.

The release also mentions native understanding of the Grok Bot harness. That may help conversational agents built around that environment, but should not be read as a universal advantage in every framework. Tool definitions, memory and orchestration vary considerably. A team switching only the model should hold the surrounding agent design constant during an initial comparison.

Long-context work needs special scrutiny. More available context can help a model inspect a large repository or document set, yet irrelevant material can still crowd out the evidence that matters. Track which files were retrieved, whether the model cited or changed the right ones, and whether the same result could be reached with a smaller, cheaper prompt.