Grok 4.7: Price, Benchmarks, Context Window and How to Use It

SpaceXAI released Grok 4.7 on September 21, 2026, calling it its most capable model yet for coding and knowledge work. It is priced exactly like Grok 4.6 at $2 per million input tokens and $6 per million output tokens, has a 500,000-token context window, and is available now through the xAI API, Cursor and SpaceXAI’s own Grok Build coding agent.

What is Grok 4.7?

Grok 4.7 is the successor to Grok 4.6, which SpaceXAI shipped in August. In its launch post for Grok 4.7, the company describes it as “our most capable model for coding and knowledge work” and says it “works longer on difficult tasks, checks its own work more carefully, and comes with our best-calibrated safeguards to date.” The headline pitch is price-performance: the company claims it is “twice as fast, at half the price of comparable models.”

Under the hood, SpaceXAI says Grok 4.7 is built on a new, larger base model than Grok 4.6. It was trained with a longer reinforcement-learning run on a harder mix of tasks, weighted toward problems that take many hours to complete. The company highlights three practical improvements:

• Self-verification: the model is said to be better at checking its own work before it hands results back.
• Long context management: it is trained to handle longer working sessions without losing track of earlier material.
• Grok Bot awareness: it was trained to natively understand the Grok Bot agent harness, which SpaceXAI says makes it better at conversational tasks and general knowledge work.

The release lands in an unusually crowded week. OpenAI’s GPT-6 Sol and Luna and Anthropic’s Claude Opus 5.5 both arrived the following day, so the comparisons in SpaceXAI’s own launch material were already a generation behind within 24 hours. That matters when reading the benchmark table below.

How much does Grok 4.7 cost?

Standard API pricing is unchanged from Grok 4.6. The official xAI API pricing page lists two tiers based on prompt length, and the switch is not gradual: once a request’s prompt reaches 200,000 tokens, every token in that request is billed at the higher rate.

A few other billing details are worth knowing before you switch a production workload:

• US regional endpoint: requests sent to https://us.api.x.ai/v1 keep inference in the United States and cost 1.1 times the global rate. For now only grok-4.7 and grok-4.6 are served there.
• Priority Processing: setting service_tier: "priority" doubles the rate for all token types in exchange for higher scheduling priority. You are only billed at that rate if the response confirms the priority tier was used.
• No Batch API: the Grok 4.7 model page says Batch API is not supported for this model, so there is no asynchronous discount.
• Server-side tools cost extra: web search and code execution are billed at $5 per 1,000 calls on top of tokens, and X Search is billed per post ($5 per 1,000) and per profile ($10 per 1,000) returned.

Grok 4.7 specs and API details

The developer overview in the Grok 4.7 API documentation lists the following:

Two implementation notes stand out. First, on the Responses API Grok 4.7 always returns reasoning.encrypted_content, even if your request does not ask for it; SpaceXAI tells developers to pass those reasoning items back unchanged in the next request so multi-turn sessions keep the model’s prior reasoning. Second, the company “highly recommends” setting a prompt_cache_key (or the x-grok-conv-id header on Chat Completions). Without it, the docs warn, requests can land on a cache-cold server and you “often pay full input price”, which quietly erases the $0.50 cached-input discount.

Grok 4.7 benchmarks: what SpaceXAI claims

All figures below come from SpaceXAI’s own launch post and have not been independently reproduced here. Grok 4.7 was run at its xhigh reasoning setting and competitors at their highest settings. Note that the rival models listed are GPT-5.6 Sol and Anthropic’s Fable 5.1, not the GPT-6 Sol and Claude Opus 5.5 models that launched the next day.

Read plainly, SpaceXAI’s own numbers do not show Grok 4.7 beating the most expensive frontier models across the board. Fable 5.1 leads on CursorBench, Terminal-Bench, AA Briefcase and HealthBench, and GPT-5.6 Sol edges it on DeepSWE. Where Grok 4.7 does lead is electrical engineering and SpaceXAI’s legal-agent result, and it posts large jumps over Grok 4.6 on Terminal-Bench (20.3% to 37.6%) and EEBench (53.0% to 64.0%).

The real argument is cost. SpaceXAI’s chart plots CursorBench 4.0 score against average cost per task and places Grok 4.7 “at the frontier in price-performance.” With output tokens priced at $6 against $20 for GPT-5.6 Sol and $50 for Fable 5.1 in the company’s table, Grok 4.7 does not need to win outright to be the cheaper way to finish a long coding job.

On the GDPval professional-task evaluation, the company reports Elo scores of 1,735 for Fable 5.1 (max), 1,695 for Grok 4.7 (xhigh), 1,605 for Grok 4.6 (high) and 1,542 for GPT-6 Astra (max).

Safety and cybersecurity changes

SpaceXAI says Grok 4.7 ships with “an entirely new safeguard stack” and calls it “the strongest model we’ve tested on refusals and jailbreak resistance.” The company’s specific claims:

• It tops LatchBio’s biosafety benchmark at 62.4%, according to SpaceXAI.
• On HackerBench v0.3, SpaceXAI’s internal benchmark for risky and malicious cyber tasks, it let through only 3.3% of risky dual-use prompts while, the company says, rarely blocking legitimate security work.
• Select cybersecurity partners are getting invite-only access to Grok 4.7’s red-team capabilities for defensive research.

HackerBench is SpaceXAI’s own benchmark, so treat that result as a vendor claim until outside evaluations appear.

How to use Grok 4.7 right now

Through the xAI API

Create a key in the xAI console and set the model to grok-4.7. The API accepts the official xAI SDK, the Vercel AI SDK, and the OpenAI SDK pointed at https://api.x.ai/v1, so most existing OpenAI-compatible code only needs a base URL and model name change. The xAI API release notes confirm the model went live on the public API at launch.

In Cursor

Grok 4.7 is available on all Cursor plans. Cursor’s models and pricing documentation places it in the “Cursor Models” usage pool alongside Grok 4.6, Grok 4.5 and Composer 2.5, which Cursor says comes with “significantly more included usage” than the pool for third-party models. That is not a coincidence: SpaceX closed its acquisition of Cursor’s parent company Anysphere on August 14, 2026, and Cursor now sits inside the SpaceXAI division. It is also the only third-party place, alongside Grok Build, to get the faster variant.

In Grok Build

Grok 4.7 is the default model in Grok Build, SpaceXAI’s terminal coding agent, which the company says can be tried for free. Grok 4.7 Fast is not included in Grok Build’s free tier. The launch post gives a one-line installer: curl -fsSL https://x.ai/cli/install.sh | bash.

Through model gateways

OpenRouter, Vercel and Cloudflare all carry the model, according to SpaceXAI’s docs. Gateway pricing and limits are set by each platform, so check their listings before routing traffic.

SpaceXAI’s announcement concentrates on developer channels. It does not set out a separate rollout timeline for the consumer Grok apps, so if you want to use Grok 4.7 without writing code, Cursor and Grok Build are the documented routes today.

What is Grok 4.7 Fast?

Grok 4.7 Fast is not a smaller or distilled model. According to SpaceXAI’s documentation it is “the same model served on faster infrastructure,” with roughly twice the output speed, billed at twice the standard token rates. There are three catches:

• It is available only in Cursor and Grok Build, not on the public xAI API.
• It is not part of Grok Build’s free tier.
• Cursor bills its own fast variant through your Cursor plan rather than at xAI’s list price, so the effective cost depends on which Cursor plan you are on and how much included usage you have left.

For interactive coding, where you sit waiting for each response, the faster variant may be worth the premium. For background agents that run unattended, the standard model at half the price is the more obvious choice.

Grok 4.7 vs Grok 4.6: should you switch?

For API users, the switch is close to free. Price, context window, reasoning settings and the list of supported tools are identical between the two models on xAI’s pricing and release-notes pages, and SpaceXAI says Grok 4.7 is served at the same speed. The differences are in capability, and by SpaceXAI’s measurements the gains are largest on long terminal-based work and engineering tasks.

• Switch now if you run agentic coding or multi-hour tasks where Grok 4.6 stalled; the reported Terminal-Bench jump is the biggest single improvement in the table.
• Test first if your prompts rely on specific refusal behaviour, because the safeguard stack is new.
• Watch the long-context threshold either way: prompts at or above 200,000 tokens double the bill for the whole request.

Bottom line

Grok 4.7 is not, on SpaceXAI’s own figures, the single strongest model on the market, and the comparisons it launched with were overtaken within a day by GPT-6 Sol and Claude Opus 5.5. What it offers is a strong coding and knowledge-work model at $2 in and $6 out, with a 500,000-token window and tight integration into Cursor, which SpaceX now owns. For teams running long agent loops where token bills add up quickly, that price point is the reason to run it through your own evaluations this week. Independent benchmark results should follow over the coming days and will be the better guide to how it stacks up against the newest rivals.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage