Grok 4.7 is xAI’s September 21, 2026 flagship for coding and knowledge work — a new, larger base model served at the same price and speed as Grok 4.6 ($2 / $6 per million tokens). xAI calls it “twice as fast, at half the price of comparable models”, and it is the default model in Grok Build and available on every Cursor plan.
It follows Grok 4.5 (July 8) and Grok 4.6 (August 12) on a cadence of roughly five to six weeks.
What’s new
- A new, larger base model. Compared with Grok 4.6, xAI pairs a bigger base with a longer reinforcement-learning run weighted toward problems that take many hours.
- Better self-checking and long context. xAI says the model verifies its own work more carefully and handles longer context better, and was trained to understand the Grok Bot harness natively, improving conversational and general knowledge-work tasks.
- Four reasoning levels. Low, medium, high (default) and xhigh. The headline numbers below are mostly xhigh.
- A new safeguard stack. xAI reports the best jailbreak resistance it has tested and gives select cybersecurity partners invite-only access to red-team capabilities.
Benchmarks
Vendor-reported at launch (xAI, September 21, 2026). Other models are shown at their maximum effort as xAI ran them; each lab uses its own harness, so Anthropic’s and OpenAI’s tables list different numbers for the same models.
| Benchmark | What it measures | Grok 4.7 (xhigh) | Grok 4.6 (high) | GPT-5.6 Sol (max) | Fable 5.1 (max) |
|---|---|---|---|---|---|
| CursorBench 4.0 | Long-running coding sessions | 46.3% | 40.4% | 41.7% | 51.8% |
| DeepSWE v1.1 | Real-codebase software engineering | 71.0%* | 65.2% | 72.7% | 70.0% |
| Terminal-Bench 4.0 | Multi-hour terminal work | 37.6% | 20.3% | 37.3% | 57.9% |
| EEBench | Electrical engineering | 64.0% | 53.0% | 39.4% | 56.4% |
| AA Briefcase v1.1 | Multi-hour office work (Elo) | 1,657 | 1,546 | 1,487 | 1,678 |
| Harvey Legal Agent Benchmark | Legal work | 19.6% | 15.8% | 2.5% | 6.7% |
| HealthBench Professional | Clinical reasoning | 56.7% | 48.5% | 60.5% | 62.1% |
*High effort. On GDPval (real professional tasks, Elo) xAI reports Fable 5.1 at 1735, Grok 4.7 at 1695, Grok 4.6 at 1605 and GPT-6 Astra at 1542.
xAI’s own framing is price-performance: on CursorBench 4.0 it plots Grok 4.7 on the frontier of average cost per task against Fable 5.1, Opus 5, GPT-5.6 Sol and Sonnet 5.
Pricing
| Per 1M tokens | Up to 200K prompt | Above 200K prompt |
|---|---|---|
| Input | $2.00 | $4.00 |
| Cached input | $0.50 | $1.00 |
| Output | $6.00 | $12.00 |
- Grok 4.7 Fast is the same model on faster infrastructure at 2x the standard token rates (1.5x for long context). It runs only in Cursor and Grok Build, billed through your plan there, and not in Grok Build’s free tier.
- The US regional endpoint (
us.api.x.ai) keeps inference in the United States at a 10% premium. - xAI recommends setting a
prompt_cache_keyso repeat requests hit a warm cache; otherwise you often pay full input price.
Specs
| Spec | Value |
|---|---|
| API model ID | grok-4.7 |
| Context window | 500,000 tokens |
| Output limit | No text output limit |
| Modalities | Text and image in, text out |
| Reasoning | Low, medium, high (default), xhigh; encrypted reasoning always returned on the Responses API |
| Tools | Function calling, web search, X search, code execution |
| Knowledge cutoff | May 2026 |
| Where it runs | xAI API, Grok Build, Cursor, OpenRouter, Vercel, Cloudflare |
How Grok 4.7 compares
- vs Grok 4.5: the same $2 / $6 list price, with the newer model far ahead on long tasks. Grok 4.6 sits between them.
- vs Claude Fable 5.1: a fifth of the price ($2 / $6 vs $10 / $50) for most of the coding performance, but a clear gap on terminal work and clinical reasoning.
- vs GPT-6.1 Sol: the same $2 input price as Sol, with output $4 cheaper ($6 vs $10); OpenAI’s published table does not directly overlap xAI’s.
- vs Claude Sonnet 5.5: Sonnet 5.5 posts 70.6% on Terminal-Bench 4.0 on Anthropic’s table against Grok 4.7’s 37.6% on xAI’s, at the same $2 input price. Harnesses differ, so treat that as a warning to test, not a verdict.
Limitations
- Self-reported numbers. Every score here comes from xAI; the comparison rows use rivals’ maximum-effort settings under xAI’s harness.
- Effort costs tokens. xhigh results consume more reasoning tokens than the rate card suggests.
- No Fast on the public API. The 2x-speed variant is limited to Cursor and Grok Build.
- Smaller window. 500K tokens trails the 1M windows on Claude 5.5, GPT-6 and Gemini.
For the wider picture, see the 2026 LLM guide.
Best for
- Agentic coding in Grok Build and Cursor
- Long, multi-hour coding and terminal tasks where price-performance matters
- Professional knowledge work — legal, engineering and office deliverables
- Cost-conscious agent fleets that need a frontier-class model at $2 / $6
Pros & cons
Strengths
- Same $2 / $6 price and speed as Grok 4.6 while moving to a new, larger base model and a longer reinforcement-learning run weighted toward many-hour tasks
- Big jumps over Grok 4.6 on xAI's own table — Terminal-Bench 4.0 20.3% to 37.6%, EEBench 53.0% to 64.0%, CursorBench 4.0 40.4% to 46.3%
- Four reasoning levels (low, medium, high, xhigh) and a Fast variant at twice the output speed inside Cursor and Grok Build
- New safeguard stack — xAI reports only 3.3% of risky dual-use cyber prompts get through on HackerBench v0.3 while legitimate security work is rarely blocked
Limitations
- Still trails Claude Fable 5.1 on most rows of xAI's own table, including Terminal-Bench 4.0 (37.6% vs 57.9%) and CursorBench 4.0 (46.3% vs 51.8%)
- 500K context is half of the 1M windows now standard from Anthropic, OpenAI and Google
- Text and image input only, with no audio or video input
- All benchmark figures are xAI-reported and several are taken at xhigh effort, so cost per task can be well above the sticker price
Frequently asked questions
When was Grok 4.7 released?
xAI announced it on September 21, 2026 and made it available the same day on the xAI API as `grok-4.7`, in Cursor, in Grok Build, through third-party coding harnesses and via gateways such as OpenRouter, Vercel and Cloudflare.
How much does Grok 4.7 cost?
$2 per 1M input tokens, $0.50 per 1M cached input tokens and $6 per 1M output tokens below 200K prompt tokens. Above 200K prompt tokens the rates are $4 / $1 / $12. The Fast variant costs 2x the standard rates, or 1.5x for long-context requests.
How is Grok 4.7 different from Grok 4.6?
It uses a new, larger base model trained with a longer reinforcement-learning run on harder, longer tasks, and it is better at verifying its own work and managing long context. Price, speed and the 500K window stay the same. On xAI's table Terminal-Bench 4.0 rises from 20.3% to 37.6% and DeepSWE v1.1 from 65.2% to 71.0%.
Is Grok 4.7 better than Claude Fable 5.1 or GPT-5.6 Sol?
On xAI's own table it beats GPT-5.6 Sol on CursorBench 4.0 (46.3% vs 41.7%), EEBench, AA Briefcase and Harvey's legal benchmark and is level on Terminal-Bench 4.0, but it trails Fable 5.1 on most rows. It costs a fifth of Fable 5.1's list price. Labs run rivals under their own harnesses, so wait for independent testing.
What is the context window and API model name?
The API model is `grok-4.7` with a 500,000-token context window, no text output limit and a May 2026 knowledge cutoff. It is served through the Responses API with function calling, web search, X search and code execution.