Grok 4.6: Release Date, Pricing, Benchmarks & Availability

Same specs, a longer training run

xAI released Grok 4.6 on August 12, 2026, four weeks after. The context window stayed at 500,000 tokens and pricing stayed at $2 per million input tokens and $6 per million output tokens, both unchanged from Grok 4.5. xAI describes the change as a longer supplemental training run on curated reasoning data with an improved optimizer, plus a new xhigh reasoning effort level alongside the existing low, medium and high settings. [1][2]

Independent tracking from Artificial Analysis put Grok 4.6's Intelligence Index score at 61, up from 56 for Grok 4.5, ranking it 6th of 187 models tracked and just behind Claude Opus 5 at 63. Output speed rose from 54.2 to 59.2 tokens per second, but time to first token rose from about 14.5 seconds to nearly 44, roughly tripling. On Anthropic's Arena human preference leaderboard, Grok 4.6 ranked 48th as of August 27, a wide gap from its benchmark placement. [2]

Atlas interpretation: A five point jump on the Artificial Analysis index is a real gain for a same-quarter point release, but it came bundled with a latency cost that a benchmark table does not show on its own. CoderSera's comparison of the two numbers, a tripled time to first token against a fairly modest throughput gain, is the more useful fact here than either score in isolation: it says the model got smarter by thinking longer before it starts answering, not by answering faster. [2]

Cheap at the list price, until a prompt gets long

The headline $2/$6 rate holds only under 200,000 tokens of prompt; CoderSera reports that crossing that threshold doubles the rate to $4 input and $12 output across the entire request rather than just the tokens over the line, and that cached input is billed separately at $0.50 per million tokens. Grok 4.6 shipped first through Cursor, Grok Build and the SpaceXAI API console, in the US regions only, and reached Amazon Bedrock, Google's Model Garden and Microsoft Foundry within two weeks. [2][1]

Atlas interpretation: No competing model at that list price matches Grok 4.6's Intelligence Index score, which is the strongest part of xAI's pitch. CoderSera's framing of the model as "frontier-adjacent, not frontier-leading", useful for batch and background agent work where the latency does not matter, less suited to interactive use or to prompts that regularly cross the pricing cliff, tracks with a release built around a longer training pass on the same architecture rather than a new frontier model. [2]