Grok 4.7 Launches at $2 per Million Tokens

Published September 23, 2026 - San Francisco, California. xAI released Grok 4.7 on September 21, 2026 at $2 per million input tokens and $6 per million output tokens - the same headline price and serving speed as Grok 4.6.

The release is a quality upgrade rather than a pricing change. Grok 4.7 ships with a 500K token context window, multimodal text-and-image input, text-only output, and stronger agentic-coding performance on CursorBench 4.0 (x.ai news, September 21, 2026).

At a glance

• Released September 21, 2026; API model ID `grok-4.7`.
• Standard direct API pricing: $2 input / $6 output per million tokens; cache $0.50 per million.
• Long-context above 200K: $4 input / $12 output / $1 cached per million.
• 500K token context; multimodal text-and-image input; text-only output.
• Knowledge cutoff May 2026.
• At frontier of CursorBench 4.0 (long-running coding tasks) on price-performance.
• Available in Cursor, Grok Build, Grok API, GitHub Copilot from launch.
• Same headline price and serving speed as Grok 4.6.

Data last verified September 23, 2026 from x.ai news, the xAI pricing page, llm-stats.com, aisignalhq.com, and the GitHub Copilot changelog.

Quick Answer

xAI released Grok 4.7 on September 21, 2026 at $2 input and $6 output per million tokens - the same headline price as Grok 4.6.

The release ships a 500K token context window, multimodal input, and stronger agentic-coding performance on CursorBench 4.0. Reasoning effort runs from low to xhigh. The model is live in Cursor, Grok Build, the Grok API, and GitHub Copilot from launch (x.ai news, September 21, 2026).

For most workloads, Grok 4.7 is a drop-in replacement for Grok 4.6 with a quality bump rather than a pricing or speed change. The headline win is on long-running coding tasks inside tools like Cursor and the GitHub Copilot cloud agent.

How Grok 4.7 differs from Grok 4.6

Grok 4.6 shipped in August 2026 as a quality refresh of Grok 4.5 with the same API prices. Grok 4.7 keeps that pattern. The xAI launch post highlights four improvements:

• Works longer on difficult tasks before asking for clarification, including multi-hour coding sessions in Cursor.
• Stronger self-check behaviour: the model re-examines its own output before delivering.
• Best-calibrated safeguards xAI has published; safer on dual-use cyber requests, web-search hallucination, and self-harm refusal than the Grok 4.6 release (xAI model card, September 21, 2026).
• Frontier-of-price-performance on CursorBench 4.0, a long-horizon coding benchmark (x.ai news, September 21, 2026).

The price tier - $2 input / $6 output / $0.50 cached input per million tokens - matches Grok 4.6 exactly. xAI calls the launch 'twice as fast, at half the price of comparable models' (x.ai news, September 21, 2026).

Pricing tiers

xAI publishes two tiers for Grok 4.7, separated by prompt size:

Long-context pricing doubles in for prompts bigger than 200K tokens. The tier boundary is sharp; code-agent workflows that average near 200K should budget for the higher tier (xAI pricing, September 21, 2026).

What Grok 4.7 can and cannot do

Grok 4.7 is positioned for coding, agentic tasks, and knowledge work. xAI highlights three capabilities at launch:

• Coding: Function calling, structured outputs, code execution, prompt caching, multi-turn tool use. Targets long-running coding sessions inside Cursor and GitHub Copilot.
• Agentic workflows: Server-side tool routing via third-party harnesses; web search and X search tools via the Grok API.
• Multimodal input: Text and image input with text-only output. The model understands screenshots and diagrams but does not generate images.

There is no Batch API for Grok 4.7 - xAI has not shipped batch support for any Grok model as of September 21 (aisignalhq.com, September 22, 2026). The model is also not licensed for redistribution; the proprietary license applies to all Grok 4.x releases.

How to access Grok 4.7 today

Five access paths on launch day:

• Grok Build: Free trial at x.ai/build; ships the model to coding workflows.
• xAI Console: Pay-as-you-go at $2 / $6 / $0.50 per million tokens; API model ID `grok-4.7`.
• Cursor: Default for CursorBench-grade workflows inside the editor; Pro and Business tiers supported.
• GitHub Copilot: Rolling out to Pro, Pro+, Max, Business, and Enterprise SKUs through the model picker; default-on unless admin disables (GitHub Changelog, September 21, 2026).
• Third-party routers: OpenRouter and competitors list Grok 4.7 at $1.60 input and $4.80 output per million tokens - cheaper than xAI direct, with the same model behaviour (llmreference.com, September 21, 2026).

OpenRouter's lower price is structural; aggregators buy in bulk and pass savings through. For sensitive code that needs EU data residency, xAI direct is the safer pick.

How Grok 4.7 compares to the frontier

Grok 4.7 lands in a crowded early-September window alongside GPT-6 Sol and Luna (September 22) and Claude Opus 5.5 (September 22). On CursorBench 4.0, xAI positions Grok 4.7 at the frontier of price-performance - the position Anthropic claims for Claude Opus 5.5 and OpenAI claims for GPT-6 Sol.

The detailed comparison math depends on your workload.

Grok 4.7's $2 / $6 per million undercuts GPT-6 Sol's $5 / $30 by 2.5x on input and 5x on output; it matches GPT-6 Luna's $1 / $6 input tier at slightly higher output pricing.

Versus Claude Opus 5.5's $4 / $20, Grok 4.7 is half the input price and a third of the output price (xAI pricing, September 21, 2026; OpenAI API pricing, September 22, 2026).

On agentic coding - the workload xAI specifically targeted for Grok 4.7 - CursorBench 4.0 favours Grok for long-running tasks; FrontierCode 1.1 favours GPT-6 Sol; SWE-bench Multilingual favours Claude Opus 5.5. Pick the model that matches the test your workload is graded against.

What was not announced

Three items stayed off the launch post. First, no image or video generation - Grok 4.7 is text-only output, multimodal input; image gen remains a separate product path.

Second, no Batch API. Grok 4.7 does not support xAI's batch endpoint; customers running overnight jobs pay full price or route through third-party batch offerings.

Third, no formal context-window expansion. Grok 4.6 also shipped at 500K; Grok 4.7 keeps the same ceiling. Customers whose workloads need 1M+ tokens continue to use OpenAI's GPT-6 Astra or Anthropic's Claude Opus 5.5 (aisignalhq.com, September 22, 2026).

Frequently Asked Questions

Is Grok 4.7 cheaper than Grok 4.6?

No - Grok 4.7 ships at the same headline API price as Grok 4.6 ($2 input / $6 output / $0.50 cached per million tokens, under 200K prompt). It is a quality upgrade rather than a price change (x.ai news, September 21, 2026).

Does Grok 4.7 support image generation?

No - Grok 4.7 is text-only output. It accepts text and image input, so you can feed it screenshots and diagrams for analysis, but the response is text. For image generation, xAI continues to offer Aurorasa and other imaging models on the same API (xai.com, September 2026).

Can I use Grok 4.7 on GitHub Copilot free tier?

No - GitHub Copilot requires Pro, Pro+, Max, Business, or Enterprise for Grok 4.7 access. Free-tier Copilot subscribers do not get Grok 4.7 in the model picker (GitHub Changelog, September 21, 2026).

What is Grok 4.7's knowledge cutoff?

May 2026. Grok 4.7 was trained on data up to May 2026 and may not know about events after that month. The X search tool inside Grok Build adds live signal from the social network for breaking events (llm-stats.com, September 21, 2026).

Should I switch from Grok 4.6 to Grok 4.7?

For coding-agent workflows inside Cursor or the Copilot cloud agent, yes - the quality gain on long-running tasks is the headline xAI pitch. For Free-tier or chat workloads, the difference is smaller and Grok 4.6 remains a serviceable fallback at the same price tier.

Does Grok 4.7 have a faster variant?

Yes - xAI publishes a fast variant of Grok 4.7 served at 'twice the output speed at twice the price' on the API. The fast variant is targeted at low-latency code completion rather than long-running agentic coding (x.ai news, September 21, 2026).