SpaceXAI has released Grok 4.7, positioning the new model around a deceptively simple idea: difficult AI work should not end just because the first plausible answer has appeared. The company says its newest model is designed to work longer on demanding tasks, verify its own output more carefully and handle extended context more effectively, with coding and professional knowledge work at the center of the release.
The launch was announced on September 21 in SpaceXAI’s official Grok 4.7 release and covered the same day by Unite.AI. Grok 4.7 is available through Grok Build, Cursor, the Grok API, third-party coding harnesses, model routers and cloud platforms. Standard API pricing starts at $2 per million input tokens and $6 per million output tokens, matching Grok 4.6.
The important story is not simply another model-number increase. SpaceXAI is explicitly optimizing around long-running work: software engineering, terminal tasks, documents, presentations and professional assignments that require the model to sustain reasoning, inspect intermediate work and continue operating instead of prematurely declaring the task finished.
A larger base model and a longer reinforcement-learning run
SpaceXAI says Grok 4.7 uses a new and larger base model than Grok 4.6. It does not publish a parameter count in the launch announcement, so the release should not be described using speculative parameter figures that circulated before launch.
The company says it also extended reinforcement learning and changed the task mix, weighting training toward harder problems that can take many hours to complete. That emphasis helps explain the release language around persistence and self-checking: Grok 4.7 is intended to remain productive deeper into complicated workflows rather than optimizing only for short prompt-response exchanges.
SpaceXAI specifically claims improvements in verifying its own work and managing longer context. Those are increasingly important capabilities as frontier models move from answering questions toward operating as agents inside coding environments, terminals and professional software.
Grok 4.7 is being trained for the harness, not just the prompt
Another notable detail is that SpaceXAI says Grok 4.7 was trained to natively understand the Grok Bot harness. That matters because modern coding and agentic performance depends on more than raw language-model capability.
An agent has to understand how to operate inside its environment: when to use tools, how to interpret results, how to preserve state across a longer workflow and when to continue rather than stop. Training a model around the harness it will actually use is one way of reducing the gap between benchmark intelligence and practical agent behavior.
SpaceXAI says this training also improves conversational tasks and general knowledge work, suggesting that the harness-aware approach is not being treated purely as a software-engineering optimization.
The coding gains are substantial over Grok 4.6
SpaceXAI’s published benchmark table reports a meaningful generation-over-generation improvement. On CursorBench 4.0, which the company says stresses longer-running coding work, Grok 4.7 at xhigh effort scores 46.3%, compared with 40.4% for Grok 4.6 at high effort.
On DeepSWE v1.1, Grok 4.7 records 71.0% at high effort versus 65.2% for Grok 4.6. On Terminal-Bench 4.0, which measures multi-hour terminal work, the reported increase is larger: 38.0% for Grok 4.7 versus 20.3% for its predecessor.
These are SpaceXAI-reported results and should be read with the usual benchmark caveat: individual evaluations measure specific environments and do not guarantee the same advantage on every real-world repository or engineering workflow. The direction of the company's optimization, however, is clear. Grok 4.7 is being pushed toward tasks that require sustained execution rather than one-shot code generation.
It does not lead every coding benchmark
The launch data is more interesting because it does not present Grok 4.7 as universally dominant. In SpaceXAI’s table, Grok 4.7’s 46.3% CursorBench score is below Fable 5.1’s reported 51.8%, while its 71.0% DeepSWE score is close to but below the 72.7% figure listed for GPT-5.6 Sol.
Terminal-Bench shows a different pattern: Grok 4.7’s reported 38.0% is slightly above GPT-5.6 Sol’s 37.3% but well below Fable 5.1’s 57.9%. Because the compared models also use different effort settings, these figures should not be collapsed into a simplistic overall ranking.
SpaceXAI’s stronger commercial argument is price-performance. The company prices Grok 4.7 at $2 per million input tokens and $6 per million output tokens, compared in its table with $4/$20 for GPT-5.6 Sol and $10/$50 for Fable 5.1.
Knowledge work is becoming as important as coding
Grok 4.7 is not marketed only as a developer model. SpaceXAI says it is better at creating documents and presentations and highlights evaluations designed around tasks performed by professionals.
On AA Briefcase v1.1, a multi-hour office-work benchmark, SpaceXAI reports a score of 1,657 for Grok 4.7 versus 1,546 for Grok 4.6. Its launch materials also highlight GDPval, which evaluates work representative of professional occupations, and say Grok 4.7 improves over its predecessor on both GDPval and AA Briefcase.
This reflects a broader change in frontier-model competition. The target is increasingly not “write a paragraph” or “solve this coding problem,” but “complete a substantial piece of work that might otherwise occupy a professional for hours.”
Legal, clinical and engineering benchmarks show a mixed but stronger profile
SpaceXAI reports 19.6% for Grok 4.7 on the Harvey Legal Agent Benchmark, up from 15.8% for Grok 4.6. On EEBench, covering electrical engineering, the new model reaches a reported 64.0% versus 53.0% for its predecessor.
On HealthBench Professional, Grok 4.7 improves from Grok 4.6’s 48.5% to 56.7%, but SpaceXAI’s own comparison lists higher figures for GPT-5.6 Sol and Fable 5.1. That is a useful reminder that generation-over-generation improvement and category leadership are separate questions.
For enterprises evaluating these systems, the practical decision is therefore likely to remain workload-specific. A model can offer an attractive blend of price, latency and performance without leading every benchmark category.
SpaceXAI is keeping Grok 4.6 pricing
Despite the larger base model and longer training run, SpaceXAI is launching standard Grok 4.7 at the same starting API price as Grok 4.6: $2 per million input tokens and $6 per million output tokens.
The company also offers a fast variant with twice the output speed at twice the price. This gives developers a latency-cost tradeoff without requiring a different model generation.
SpaceXAI describes Grok 4.7 as twice as fast at half the price of comparable models, but that is the company’s comparative positioning rather than a universal market measurement. Actual cost depends on token usage, reasoning effort, caching, workload shape and the competing model configuration used for comparison.
Longer work makes self-verification more important
The repeated emphasis on checking its own work is strategically important. As an AI task becomes longer, an early mistake can propagate through dozens of later actions. A coding agent that misunderstands an interface, a research agent that adopts a false premise or a document agent that misreads a requirement can spend substantial compute producing a polished but incorrect result.
Self-verification therefore becomes more valuable as models are asked to operate for longer periods. It is not enough for an agent to continue working; it has to notice when the work is drifting and revisit intermediate assumptions.
SpaceXAI’s claim that Grok 4.7 is better at verification should ultimately be judged in real deployments, but it aligns with the problem the company is targeting: reliable execution over longer task horizons.
Safety has been rebuilt alongside capability
SpaceXAI says Grok 4.7 uses an entirely new safeguard stack and describes it as the strongest model the company has tested for refusals and resistance to jailbreaks.
In biological safety, the company says Grok 4.7 reaches 62.4% on LatchBio’s biosafety benchmark. On SpaceXAI’s HackerBench v0.3, which targets risky and malicious cyber tasks, the company reports that only 3.3% of risky dual-use prompts were allowed through while legitimate security work was rarely blocked.
Those claims are particularly relevant because a model optimized for long-running agentic work can potentially take more consequential actions than a conventional chatbot. SpaceXAI says it has begun giving selected cybersecurity partners invite-only access to Grok 4.7 red-team capabilities for defensive research.
Grok 4.7 is available across coding and model infrastructure
The launch is not a preview. SpaceXAI says Grok 4.7 is available from September 21 in Cursor and Grok Build, as well as through the Grok API, third-party coding harnesses, model routers and cloud platforms.
The company is also offering access to Grok 4.7 inside Grok Build for free. That gives developers a direct environment for testing the model on agentic coding work without beginning with a standalone API integration.
The broad launch footprint matters because agent models increasingly compete on ecosystem availability as much as on raw evaluation scores. Developers want models that can enter the tools and harnesses where their work already happens.
The release confirms a shift from answer quality to work endurance
The most useful way to read Grok 4.7 is not as a benchmark refresh but as another step toward AI systems evaluated by how long they can remain useful on a task. SpaceXAI repeatedly emphasizes multi-hour coding, terminal work, professional assignments, longer context and self-checking.
That changes what “better model” increasingly means. A model that produces an excellent first response but fails after ten tool calls may be less valuable for agentic work than one that can plan, execute, inspect and revise over a much longer horizon.
The competitive frontier is consequently moving toward persistence: how much useful work can the system complete before human intervention becomes necessary?
Grok 4.7 arrives after a period of anticipation
The model had been discussed publicly before the September 21 release, but official specifications, benchmark results, pricing and availability were not established until SpaceXAI published the launch materials. The new announcement therefore resolves much of the uncertainty around what would actually ship.
The confirmed product is a larger model with a longer reinforcement-learning run, stronger reported performance across several long-horizon evaluations, unchanged standard API pricing and a new safeguard stack. Claims circulated before launch that do not appear in the official model announcement should still be treated separately from the shipped specification.
That distinction is particularly important in fast-moving AI coverage, where founder comments and pre-release expectations can quickly be repeated as if they were product documentation.
The next AI battleground is sustained professional work
Grok 4.7 makes SpaceXAI’s direction unusually explicit. Coding remains a flagship workload, but the company is expanding the same long-horizon model toward documents, presentations, legal work, engineering and other professional tasks.
The commercial proposition combines stronger persistence with aggressive token pricing. Whether Grok 4.7 becomes the preferred model for a particular workflow will depend on real-world reliability, latency, tool integration and the kinds of errors it makes outside curated evaluations.
But the release makes the strategic target clear: frontier AI is being asked to stop behaving like a chatbot that answers and start behaving like a worker that stays with the problem. Grok 4.7 is SpaceXAI’s latest attempt to make that transition practical for coding and knowledge work.