xAI Launches Grok 4.6 at $2 per Million Tokens - Half the Price of Rival Frontier Models

Grok 4.6 is out. xAI released the model on August 12, built specifically for long-running agents, codebase work, and interactive and visual projects - pricing it at $2 per million input tokens and $6 per million output tokens, identical to Grok 4.5's price, while claiming this version handles materially harder tasks across longer chains. SpaceX acquired Cursor for $60 billion earlier this year, restructuring xAI's developer distribution; the model launched across the combined company's full reach on the same day.

Same Price as Grok 4.5 - Half the Cost of Rival Frontier Models

Half the price. At $2 per million input tokens, xAI undercuts Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol at comparable capability tiers by roughly half - a pricing gap the company is leading with explicitly in its launch materials. To pull developers onto the model before the promo window closes, xAI doubled the included usage in Cursor and Grok Build for the first week. Grok 4.5 benchmarks already showed xAI competitive on coding; this release extends that sequence after a longer training run, improved supervised fine-tuning, and additional reinforcement learning across agent harnesses, STEM, and software engineering domains.

1753 ELO on Knowledge Work - Coding Benchmarks Still Trail GPT-5.6 Sol Max

Mixed results across the full benchmark table. On GDPVal-AA v2, a knowledge-work ELO leaderboard, Grok 4.6 leads all models at 1753 - the same figure Musk cited on X at launch. On the Artificial Analysis Intelligence Index, which composites nine benchmarks, the model matches GPT-5.6 Sol at a composite score of 61, trailing Anthropic's Fable 5 Max at 62.

Coding benchmarks show a clear gap. On DeepSWE v1.1, xAI's model posted 65.9%, up sharply from Grok 4.5's 54%, but GPT-5.6 Sol Max scored 73% on the same test; Terminal-Bench v3.0 follows the same pattern, with xAI at 26% against roughly 34% for Sol Max and Fable 5 Max. All figures come from xAI's launch table, with third-party scores drawn from self-reported results; independent evaluation on public arenas had not yet confirmed the numbers at launch. A discount priced at half the competition makes more sense when the model leads on knowledge work than when it trails on the coding benchmarks developers use daily to choose between frontier models.

Available in Cursor, Grok Build, and the API - Plus OpenRouter, Vercel, and Cloudflare

Four surfaces, all live August 12. Grok 4.6 ships in the xAI API, Grok Build, Cursor, and Grok Bot on X; third-party access arrived simultaneously via OpenRouter, Vercel, and Cloudflare, giving developers routing through model marketplaces a path to the new model without switching providers. xAI's Grok Bot - AI agents that sign into tools and run overnight jobs - now runs on Grok 4.6 as its underlying inference layer. Grok 4.7, a 2.1-trillion-parameter successor on a larger architecture than the current 1.5T V9 base, is in development; xAI has indicated it expects to ship within weeks of this release.