Grok 4.7 for builders: same $2/$6 as 4.6, pin the id you call
SpaceXAI published Grok 4.7 as its most capable model for coding and knowledge work. Same price and speed class as Grok 4.6: $2 per million input tokens and $6 per million output tokens. Available today in Cursor and Grok Build; also the Grok API, third-party coding harnesses, and model routers and cloud platforms. A fast variant doubles output speed at double the price.
This is not a Grok 4.5 recap — earlier Grok 4.5 Cursor-training notes covered a different release. Here the question is Grok 4.7 versus Grok 4.6 on the official card: larger base, longer reinforcement learning, stronger self-verification, native Grok Bot harness fit, same $2 / $6 list line. Copy the model id from the console you use, then measure the change on your own tasks.
What the launch page commits to
Grok 4.7 uses 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. It is better at verifying its own work and managing longer context. It was trained to natively understand the Grok Bot harness, which the page ties to better conversational tasks and general knowledge work.
Hold the gaps. The launch page does not publish an API model id string, a retirement date for Grok 4.6, millisecond latency, or cache prices. It does not publish the exact id for Cursor, Grok Build, the Grok API, a third-party harness, or a cloud router. Copy the id from the console or picker you actually call. A UI label that reads Grok 4.7 is not the pin — same failure mode as Gemini Arena label versus API pin.
The rate card versus Grok 4.6
Prices on the launch page are US dollars per million tokens.
| Line | Grok 4.7 | Grok 4.6 |
|---|---|---|
| Input | $2 | $2 |
| Output | $6 | $6 |
Same list price and speed class as Grok 4.6 for the standard model. The upgrade case is not a discount — it is whether the larger base, longer RL, and self-verification improve work you already run at $2 / $6.
The fast variant has twice the output speed at twice the price. The launch page does not name a distinct model id for it. Enable it only through the product control you use; do not bill that traffic at the $2 / $6 card by accident. Uncached check: 5M input + 1M output = $16 on the standard card; $32 if both lines double on the fast variant. That is a speed premium, not a migration discount from Grok 4.6.
Same component. Web and mobile. One codebase.
The free, open-source SDK gives you components that work the same on web and mobile — one codebase. github.com/otf-kit/sdk
Where Grok 4.7 is available
On the SpaceXAI launch page, Grok 4.7 is available today in Cursor and Grok Build. It is also available through the Grok API, third-party coding harnesses, and model routers and cloud platforms.
Separately, the GitHub Changelog for September 21, 2026 says Grok 4.7 is rolling out in GitHub Copilot for Pro, Pro+, Max, Business, and Enterprise. GitHub calls it xAI's latest reasoning model, building on Grok 4.6, for agentic coding and complex multistep workflows. Select it in the model picker in Visual Studio Code, Visual Studio, Copilot CLI, GitHub Copilot cloud agent, GitHub Copilot app, JetBrains, Xcode, and Eclipse. Rollout is gradual. Enterprise and Business admins manage access via Copilot model policy; under default enablement, new models are on unless the global default is off or this model is explicitly disabled. Billing is provider list pricing under usage-based billing.
Two surfaces: SpaceXAI's Cursor / Grok Build / Grok API list, and Copilot's SKU plus admin policy. Pin the id on the surface you bill against. Do not assume the Copilot picker string matches the Grok API string.
Harness fit: why Grok Bot training matters
The launch page says Grok 4.7 was trained to natively understand the Grok Bot harness. Test that if you already run delegated agent work. It is not a claim that every third-party harness gets the same lift. Grok Bot-shaped stacks — persistent agents, tool use, longer conversational loops — are the trial case. Thin completion calls with no harness conventions should treat the sentence as context, not a free upgrade. For Grok Bot product shape, see the earlier cross-app agents note.
Self-verification and longer context should shape evals: multi-file patches with tests, behavior-preserving refactors, longer office or terminal work if that is your product. Short autocomplete samples will not show the difference the page is selling.
Benchmark rows you can quote exactly
On CursorBench 4.0 (longer-running coding tasks per the page), Grok 4.7 scores 46.3% versus Grok 4.6 High at 40.4% — 5.9 points on the published numbers.
| Eval | Grok 4.7 | Grok 4.6 High |
|---|---|---|
| CursorBench 4.0 | 46.3% | 40.4% |
| DeepSWE v1.1 | 71.0%* | 65.2% |
| Terminal-Bench 4.0 | 38.0% | 20.3% |
| EEBench | 64.0% | 53.0% |
DeepSWE's asterisk for Grok 4.7 is marked high effort on the launch page. AA Briefcase v1.1 is 1,657 versus 1,546. The page also lists GPT-5.6 Sol Max and Fable 5.1 Max; this post does not invent a cross-vendor winner. Use the Grok 4.7 versus Grok 4.6 columns to decide a trial is worth running.
Public scores are not your repository. A higher CursorBench or Terminal-Bench number does not mean your failure modes moved. Use the table to justify a pin trial. Use your issues, your tests, and a sample of production traces to decide the pin stays.
Safety lines that are on the page
Quote only what the launch page states. Grok 4.7 has an entirely new safeguard stack; SpaceXAI calls it the strongest model they have tested on refusals and jailbreak resistance. In dual-use domains like cybersecurity and biological work, it leads on utility for benign tasks and safe refusal on dangerous ones, topping LatchBio's biosafety benchmark at 62.4%. On HackerBench v0.3 it allows only 3.3% of risky dual-use prompts through while rarely blocking legitimate security work. Select partners get invite-only red-team access for defense research. Vendor figures — not a substitute for your threat model or agent approval gates.
When to pin the model id
Pin Grok 4.7 when three conditions hold.
First, you can copy an exact model id from the console, provider docs, or picker for the surface you call. The launch page does not give that string — do not invent one or paste a marketing name into config.
Second, your workload matches the claims: longer-running coding, self-verification, longer context, or Grok Bot harness fit. If you only need $2 / $6 and Grok 4.6 evals are already green, there is no list-price reason to move.
Third, run a replay window on the same tasks against Grok 4.6 before flipping production. Tag fast-variant requests so the invoice separates $2 / $6 traffic from 2× priced traffic.
Skip the pin during Copilot's gradual rollout, when Business/Enterprise policy disables the model, or when you cannot see the id behind the label. Wait for the control plane you trust, then pin.
Builder checklist
- Confirm the $2 / $6 card and Cursor / Grok Build / API lines on https://x.ai/news/grok-4-7.
- Copy the model id from the console you will bill — not from this post.
- Keep Grok 4.6 as the control for one replay window on your own tasks.
- If you use Copilot, confirm SKU access and admin model policy from the GitHub Changelog.
- Only then change the pin, and only on the surface you measured.
Sources
- Introducing Grok 4.7 — SpaceXAI launch page: capability, training vs Grok 4.6, $2 / $6 pricing, fast variant, Cursor / Grok Build / API availability, CursorBench and other published rows, safety figures.
- Grok 4.7 is now available in GitHub Copilot — GitHub Changelog (September 21, 2026): Copilot SKUs, model picker surfaces, gradual rollout, Business/Enterprise model policy.
Buy once, own the code. Ship with the agent you already use.
- Free, open-source SDK — same component, web and mobile
- Paid kits include AI configs + 40+ tested prompts — your agent reads the whole project
- $99/kit or $149 for everything. No subscription, no sandbox limit.