Grok 4.6 lands 27 days after Grok 4.5, at the same sticker price

SpaceXAI shipped a new flagship four weeks after the last one, left headline pricing at $2 in, $6 out and raised the cached-input rate by two-thirds. Its own comparison table shows a model that matches rivals on a composite index, leads on some benchmarks and trails on others.

Agent-drafted, human-editedPublished August 13, 2026

Reviewed by Trevor Brown

SpaceXAI released Grok 4.6 on Wednesday, 27 days after Grok 4.5 — the model it had called "our strongest model ever."1 2 The headline price did not move: $2 per million input tokens, $6 per million output.1 4 There is a fast variant at twice the price.1 One rate did move: cached input rose from $0.30 to $0.50 per million tokens below 200k prompt tokens, up two-thirds on the rate that matters most to the long-running-agent workloads the model is pitched at.5

Why it matters: four weeks is the fastest flagship turnaround in the Grok 4 line. Grok 4 came in July 2025, Grok 4.1 four months later, Grok 4.5 eight months after that.3 Holding headline pricing flat while compressing the cycle is a bid to win developers on cost and cadence rather than on a headline capability jump.

What the company says: Grok 4.6 is built for long-running agents and "more ambitious interactive and visual work," staying with multi-step tasks across research, codebases and app-building. SpaceXAI says it matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index at 61, up from 4.5's 56 and one point behind Claude Fable 5 Max at 62.1

What its own table shows: a mixed picture. Grok 4.6, tested as its High configuration against GPT-5.6 Sol Max and Claude Fable 5 Max, trails both rivals on two coding benchmarks (DeepSWE, at 65.9 percent against 73 and 70, and Terminal-Bench, where it trails by more than 8 points), beats GPT-5.6 Sol Max on three others, and posts the table's best score on three rows, including the GDPVal-AA work benchmark. All figures are the company's presentation; it notes competitor scores are "the best of self-reported or publicly available results."1

Availability: Cursor, Grok Build, the API, and partners including OpenRouter, Vercel and Cloudflare. Grok Build and Cursor users get twice their usual included usage for a week.1

The release caps a fast August for the company, which shipped the Grok Bot agent product in beta on August 11 and Imagine Image 2.0 on August 7.3

What we still don't know: SpaceXAI describes its "widest-ever suite of pre-deployment testing," plus what it calls extensive post-deployment and third-party testing, but the announcement names no outside testers and links no system card, so none of the safety claims can be checked from the post.

Provenance & disclosure

AI use

Agent-drafted, human-edited

Disclosure

Drafted by Claude from SpaceXAI's Grok 4.6 and Grok 4.5 announcements, its news index, its API page and its developer pricing table, all read August 13, then edited by Trevor Brown. Every benchmark figure is the company's own presentation and has not been independently verified. It was drafted by Claude Fable 5, a model made by a competitor of the company covered.

Reviewed by

Trevor Brown — Editor-in-chief

Sources

1.Official sourceThe August 12 release date, the long-running-agents and "more ambitious interactive and visual work" framing, the $2 / $6 per-million-token pricing and 2x-priced fast variant, the Artificial Analysis Intelligence Index scores (Grok 4.6 at 61, Grok 4.5 High at 56, GPT-5.6 Sol Max at 61, Fable 5 Max at 62), the benchmark table including DeepSWE v1.1 (65.9% vs 73% and 70%) and Terminal-Bench v3.0 (26% vs 34.6% and 34.1%), the "best of self-reported or publicly available results" caveat, the availability list and first-week doubled usage, and the "widest-ever suite of pre-deployment testing" safety language, from SpaceXAI, Introducing Grok 4.6.
2.Official sourceThe July 16 release date, the "our strongest model ever" description, and the identical $2 per million input / $6 per million output pricing, from SpaceXAI, Introducing Grok 4.5.