Overview
Grok 4.7 is SpaceXAI's September 2026 model update for coding, agentic tasks, and professional knowledge work. SpaceXAI describes Grok 4.7 as its frontier and most capable model for coding and knowledge work; it keeps the same 500,000-token context window, standard serving speed, and base API rates.
The release focuses on harder, longer assignments rather than a larger context limit or lower token price. SpaceXAI says a larger base model, a longer reinforcement-learning run, stronger self-verification, and native understanding of the Grok Bot harness support the stated longer-horizon performance gains; separately, Grok 4.7 introduces an entirely new safeguard stack. Teams should evaluate the vendor-published gains against their own agent trajectories before changing production defaults.
What's New
A larger base model trained for multi-hour work
Grok 4.7 uses a new, larger base model than Grok 4.6. SpaceXAI also extended reinforcement learning with a harder task mix weighted toward work that can take many hours, including repository-scale coding, terminal work, professional analysis, and document production.
This changes the upgrade case from raw context capacity to sustained execution quality. Both versions retain a 500,000-token context window, but 4.7 is trained to use long trajectories more effectively and remain engaged with difficult tasks for longer.
Stronger verification and context management
SpaceXAI says Grok 4.7 checks its work more carefully and manages longer context better than 4.6. The practical value is lower risk of an agent drifting from requirements, skipping validation, or presenting an intermediate result as finished during a long task.
These are vendor claims rather than a guarantee for every workflow. Teams should compare completion rate, human corrections, tool-call count, total tokens, latency, and regression failures on a fixed internal evaluation set.
Native understanding of the Grok Bot harness
Grok 4.7 was trained to understand the Grok Bot harness natively. SpaceXAI links that training to better conversational behavior and general knowledge work, including creating documents and presentations for professional tasks.
API users do not need Grok Bot to call the model. The relevant migration decision is whether their own agent harness benefits from the same longer-horizon behavior when prompts, tools, permissions, and retry policies are held constant.
A new safeguard stack
SpaceXAI says Grok 4.7 uses an entirely new safeguard stack and is the strongest model it has tested on refusals and jailbreak resistance. The company says Grok 4.7 leads its cited dual-use cybersecurity and biological evaluations on utility for benign tasks and safe refusal on dangerous ones.
Safety-sensitive teams should still keep scoped credentials, sandboxing, approval gates, and audit logs. Vendor evaluations help with shortlisting, but they do not replace testing against an organization's own threat model and allowed-use policy.
Performance Benchmarks
SpaceXAI published the following comparisons. Grok 4.7 uses xhigh effort except where DeepSWE is marked high effort; Grok 4.6 uses high effort.
| Benchmark | Grok 4.7 | Grok 4.6 | Change |
|---|---|---|---|
| CursorBench 4.0 | 46.3% | 40.4% | +5.9 points |
| DeepSWE v1.1 | 71.0% | 65.2% | +5.8 points |
| AA Briefcase v1.1 | 1,657 | 1,546 | +111 |
| Terminal-Bench 4.0 | 38.0% | 20.3% | +17.7 points |
| Harvey Legal Agent Benchmark | 19.6% | 15.8% | +3.8 points |
| HealthBench Professional | 56.7% | 48.5% | +8.2 points |
| EEBench | 64.0% | 53.0% | +11.0 points |
The largest published relative improvement is on Terminal-Bench 4.0, while the broader table suggests gains across coding, office work, legal tasks, clinical reasoning, and electrical engineering. Because effort settings and harnesses affect results, the table should be treated as vendor evidence, not a universal production forecast.
Compared With Grok 4.6
| Decision area | Grok 4.7 | Grok 4.6 |
|---|---|---|
| Model and training | Larger base model; longer reinforcement learning on harder, multi-hour tasks | Earlier base model focused on long-running agents and visual or interactive work |
| Self-verification | Stronger vendor-stated checking and long-context management | Added more self-testing than Grok 4.5, but is the lower benchmark baseline here |
| Agent and professional work | Native Grok Bot harness understanding; stronger published coding and knowledge-work results | Supports the same broad coding, research, and knowledge-work categories |
| Safeguards | Entirely new safeguard stack with new jailbreak, cyber, and biosafety evaluations | Earlier safeguard implementation |
| Context and modalities | 500K context; text and image input; text output | 500K context; text and image input; text output |
| Standard API pricing | $2 input, $0.50 cached input, and $6 output per million tokens below 200K prompts | Same standard rates |
| Fast option | Same model at twice the output speed and twice the token price in Cursor and Grok Build | Grok 4.6 release material already documented a Fast variant at twice the price; its launch post did not specify the same Cursor and Grok Build-only restriction |
Migration & Compatibility
Change the model identifier to grok-4.7 and run an evaluation before routing production traffic. The context window, input and output modalities, base token rates, supported reasoning efforts, Responses API, and Chat Completions support remain broadly continuous with Grok 4.6.
Responses API integrations need one specific compatibility check: grok-4.7 always returns reasoning.encrypted_content, even when the request does not include it explicitly. Multi-turn applications should pass those reasoning items back unchanged in the next request. Chat Completions behavior is unchanged.
For repeated conversation context, SpaceXAI recommends setting prompt_cache_key with the Responses API or the x-grok-conv-id header with Chat Completions. Without stable routing, a request may land on a cache-cold server and incur the full input-token rate.
Pricing & Access
| Access path | Price or availability | Notes |
|---|---|---|
| xAI API below 200K prompt tokens | $2 input / $0.50 cached input / $6 output per 1M tokens | Same base rates as Grok 4.6 |
| xAI API at or above 200K prompt tokens | $4 input / $1 cached input / $12 output per 1M tokens | Higher-context rates apply to the request |
| US regional API endpoint | Standard token rates plus 10% | Keeps inference in the United States |
| Grok 4.7 Fast | Twice the standard token rates | Twice the output speed; only in Cursor and Grok Build |
| Cursor | Available on all plans | Grok 4.7 is documented as available on all Cursor plans; official 4.7 docs do not state plan-level Fast eligibility within Cursor |
| Grok Build | Grok 4.7 is the default coding-agent model | Fast is not included in the free tier |
| Model gateways | OpenRouter, Vercel, and Cloudflare | Provider pricing and limits may differ |
Grok 4.7 Fast is not available through the public xAI API. The standard model is available through the xAI API, the US regional endpoint, Grok Build, Cursor, and supported model gateways.
Safety & Cybersecurity
SpaceXAI reports that Grok 4.7 reached 62.4% on LatchBio's biosafety benchmark and allowed 3.3% of risky dual-use prompts through on its HackerBench v0.3 evaluation. The company also describes invite-only red-team access for selected cybersecurity partners; that restricted program is separate from ordinary Grok 4.7 availability.
These figures describe SpaceXAI's evaluation setup and should not be converted into a general safety percentage. Organizations handling code execution, biological information, or security tooling should validate refusal boundaries and legitimate-use false positives under their own policies.
Who Should Upgrade / Who Should Wait
Upgrade if:
- Your agents run multi-hour coding, terminal, research, document, or professional-analysis workflows where self-checking and task completion matter more than a context increase.
- You already use Grok 4.6 pricing and context limits and can switch model IDs without changing workload economics.
- You need the new safeguard behavior and can validate it against your own security and compliance scenarios.
- You use Cursor or Grok Build and want to test the twice-priced Fast option for latency-sensitive work.
Wait if:
- Your Grok 4.6 workloads are short and reliable, so the published long-horizon gains do not justify migration testing yet.
- You require more than 500,000 tokens of context; Grok 4.7 does not raise the limit.
- You need the Fast variant through the public xAI API or need Batch API support, neither of which is documented as available for Grok 4.7.
- Your deployment requires independently reproduced safety or benchmark results before adopting a new default model.
Known Limitations
- The context window remains 500,000 tokens, so 4.7 does not solve workloads that exceeded Grok 4.6's limit.
- Published performance comparisons come from SpaceXAI and use stated effort settings and evaluation harnesses that may not match production workloads.
- Grok 4.7 Fast is limited to Cursor and Grok Build and costs twice the standard token rates.
- The model page lists Batch API as unsupported.
- Safety benchmark results do not remove the need for application-level permissions, review, monitoring, and incident controls.
Sources
- Introducing Grok 4.7 — release identity, model changes, benchmark comparisons, safeguards, pricing, and availability.
- Grok 4.7 developer guide — API specifications, reasoning behavior, migration notes, Fast variant restrictions, and access paths.
- SpaceXAI release notes — release date, token rates, long-context pricing, and API behavior.
