Grok 4.6 ships with 500K context and $2 input pricing

Grok 4.6, from SpaceXAI (formerly xAI), is a frontier model aimed at long agent runs and coding, with a 500K window, $2/$6 per million tokens, and a 61 on Artificial Analysis's Intelligence Index at launch.

Key points

• The API model string is grok-4.6, with a dot (some page URLs use a dashed grok-4-6). SpaceXAI's Grok 4.6 docs say it takes text and images, returns text, has a 500K-token window with no output-length cap and a knowledge cutoff of Feb 1, 2026, and lets you set reasoning effort to low, medium, high (the default) or xhigh.
• Prompts under 200K tokens cost $2 input, $0.50 cached and $6 output per million; once a prompt reaches 200K, the whole request is billed at $4/$1/$12. A fast variant costs double the standard rate.
• Function calling and structured outputs are supported, and the model page lists rate limits of 150 requests a second and 50M tokens a minute.
• Live on day one through SpaceXAI's API, Cursor, Grok Build, OpenRouter, Vercel and Cloudflare. No weights release is mentioned in the launch material.
• Since launch, GitHub Copilot began rolling it out on Aug 14 (GitHub's changelog lists the Pro, Pro+, Max, Business and Enterprise plans; on the last two an admin has to switch it on), Amazon Bedrock added it on Aug 19 (AWS), and it went into preview in Google Cloud's Model Garden on Aug 21 (SpaceXAI, Google) and in Microsoft Foundry on Aug 26 (Microsoft).
• Artificial Analysis (AA), which measures models independently, scored Grok 4.6 at 61 on its Intelligence Index at launch, 5 points up on Grok 4.5, level with GPT-5.6 Sol (max) and behind only Anthropic's Claude Opus 5 (63) and Fable 5 (62), according to its launch analysis. VentureBeat notes that puts it ahead of Moonshot's open-weights Kimi K3. AA has since revised the index, and on the current version (v4.3) its model page shows Grok 4.6 (high) at 44, three points behind GPT-5.6 Sol (max).
• SpaceXAI's own eval table adds CursorBench v3.2 69.9% (Grok 4.5: 66.7%), DeepSWE v1.1 65.9% (54%), Terminal-Bench v3.0 26% (15.7%) and APEX-Agents 57.5% (47.1%). Rival scores in it come from published system cards and leaderboards.
• On training, the same post says Grok 4.6 had a longer extra training phase than Grok 4.5's before fine-tuning, using curated synthetic reasoning data, engineering data and a revised optimizer; its fine-tuning examples were rebuilt with Grok 4.5 and screened by model-based checks; and its reinforcement learning covered agent tasks from general coding to kernel optimization and CAD.

“long-running agents and more ambitious interactive and visual work.” — SpaceXAI's announcement post, Aug 12, 2026

Why it matters

For a small studio, the draw is price: Grok 4.6's headline rates of $2/$6 are at least 60% below those of Claude Opus 5 and GPT-5.6 Sol, the rivals Artificial Analysis prices it against, though VentureBeat, comparing a wider field that includes open models, rates it mid-priced. Availability in Cursor and OpenRouter from day one, then in GitHub Copilot, Bedrock, Google Cloud and Foundry, means you can trial it without setting up a new vendor account. Watch the 200K prompt-token cliff, where the entire request flips to double rates. Shell-heavy agent work is the open question: SpaceXAI's table puts it at 26% on Terminal-Bench v3.0, up from Grok 4.5's 15.7% but short of the 34.1% and 34.6% it lists for Fable 5 Max and GPT-5.6 Sol Max, while Artificial Analysis's own run of Terminal-Bench v2.1 had it level with the top scorers. Apart from Artificial Analysis's results, every benchmark figure here comes from SpaceXAI's own table, so run your own evals before moving a production pipeline.

This is an AI-written summary of the reporting credited above and the other sources linked in the text, read and edited by Nicholas before publishing. The facts belong to those sources and the quote to SpaceXAI; the wording is ours. Read the original.