Grok 4.6 released
On 12 August 2026 SpaceXAI, the company behind Grok (xAI in other records of the atlas), released Grok 4.6 in Cursor, Grok Build and its API. It builds on Grok 4.5 with a focus on long-running agent tasks; the company says it matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, 61 against 61, with Claude Fable 5 at 62. The price starts at $2 and $6 per million input and output tokens.
Why it matters
The maker places a cheaper model level with the two leading ones on a composite index run by an outside evaluator. Until now the atlas held only the January 2026 record of the blocks on Grok for this line, so this is its first record of the current generation.
API and prices. The API release notes of the same day ('August 12'): a 500K-token context, text and image input, text output, reasoning efforts low, medium, high (default) and xhigh; $2 / $0.50 / $6 per million input, cached input and output tokens below 200K prompt tokens and $4 / $1 / $12 above. The fast variant costs twice as much. In Cursor and Grok Build usage allowances were doubled for the first week. The weights are not open. The company's figures. AA Intelligence Index 61 (Grok 4.5 56, GPT-5.6 Sol 61, Fable 5 62). In the company's table Grok 4.6 at high effort leads on some measures (GDPVal-AA v2 1753 against 1728 and 1741, AA-Briefcase 1577, Harvey LAB 15.8%), is close on CursorBench v3.2 (69.9% against 67.2% and 70.5%) and trails GPT-5.6 Sol and Fable 5 on DeepSWE v1.1 (65.9% against 73% and 70%) and Terminal-Bench v3.0 (26% against 34.6% and 34.1%). The company takes competitors' figures from their system cards and leaderboards. Training. A longer supplemental training run than Grok 4.5, with curated model-generated data; supervised fine-tuning trajectories regenerated with Grok 4.5; reinforcement learning on agentic tasks including kernel optimisation, web development and computer-aided design. The company's name. Its site and documentation call it SpaceXAI; the atlas records the organisation as xAI. What the record does not claim. All figures are the company's own, with others' numbers from their reports; no independent measurements were read other than the Artificial Analysis index as quoted by the company.