CONFIRMEDDate confidence: CONFIRMEDStated as firmHow firmly the source words the date: confirmed, scheduled, estimated, delayed or unverified.Show all CONFIRMED pins84% CONFIDENCEOverall confidence: 84%How well the pin's source and references back up its dates.Weighted average of how firmly 6 references, the source included, support the pin's start and end times; a reference counts half as much for every 180 days older than the newestShow all pins at 75% confidence or better
SpaceXAI's announcement is dated 'Aug 12, 2026' and opens 'Today we are releasing Grok 4.6'; its API release notes list it under August 12 and VentureBeat[6] reported it at 10:26 a.m. PT the same day.- Start
- Aug 12, 202690% CONFIDENCEfrom the source
Grok 4.6 Released
SpaceXAI releases Grok 4.6, built on Grok 4.5 for long-running agents and interactive work at the same $2 and $6 per 1 million token price; it scores 61 on the Artificial Analysis Intelligence Index, level with GPT-5.6 Sol and one point behind Claude Fable 5.
- SpaceXAI released Grok 4.6 on 12 August 2026, five weeks after Grok 4.5, saying it 'builds on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive and visual work' and stays with complex tasks across many steps[1][2]
- On Artificial Analysis' Intelligence Index it scored 61, matching GPT-5.6 Sol (Max) and one point behind Claude Fable 5 (Max, 62), up from 56 for Grok 4.5 (High); VentureBeat has Claude Opus 5 first and Fable 5 second, with Grok 4.6 ahead of Moonshot's Kimi K3[1][6]
- SpaceXAI's evals table has it leading on knowledge work (GDPval-AA v2 1,753 against 1,741 for Fable 5 and 1,728 for GPT-5.6 Sol, AA-Briefcase 1,577, Harvey LAB 15.8%) but behind on agentic coding: DeepSWE v1.1 65.9% against 73% for GPT-5.6 Sol and 70% for Fable 5, Terminal-Bench v3.0 26% against 34.6% and 34.1%[1][6]
- The recipe was a longer supplemental training run than 4.5, with curated model-generated reasoning data and an improved optimizer, then SFT on trajectories that Grok 4.5 regenerated across reasoning efforts, harnesses and domains, and agentic RL covering knowledge work, coding and environments for kernel optimization, web development and CAD[1]
- Efficiency went the other way: Artificial Analysis puts it at $0.84 per index task, less economical than Grok 4.5, and its model page measures only about 56 output tokens per second; VentureBeat concluded the launch 'does not establish a clear performance lead'[3][6]
- Index revisions have since moved the number: AA's model page lists Grok 4.6 (high) at 44 on its current index and marks it deprecated, after Grok 4.7 arrived on 21 September, so 61 is a launch-era score[3]
Notable features
- Priced at $2 per 1 million input tokens, $0.50 per 1 million cached input and $6 per 1 million output tokens below 200K prompt tokens, and $4, $1 and $12 above, applied to the whole request; a faster variant costs twice as much[2][6]
- A 500K-token context window with text and image input and text-only output; reasoning effort low, medium, high (default) or xhigh[2]; knowledge cutoff 1 February 2026[5]
- Available the same day in Cursor and Grok Build (with 2x included usage for the first week), the API, OpenRouter, Vercel and Cloudflare[1][6]
- Rolled out in GitHub Copilot on 14 August for Pro, Pro+, Max, Business and Enterprise plans[4]
- SpaceXAI says its safeguards were recalibrated to the model's capabilities, with its 'widest-ever suite of pre-deployment testing' and third-party testing after release[1]
SpaceXAI evals table, high or max effort[1]
| Evaluation | Grok 4.6 (High) | Grok 4.5 (High) | GPT-5.6 Sol (Max) | Fable 5 (Max) |
|---|---|---|---|---|
| AA Intelligence Index | 61 | 56 | 61 | 62 |
| GDPval-AA v2 (Elo) | 1,753 | 1,526 | 1,728 | 1,741 |
| CursorBench v3.2 | 69.9% | 66.7% | 67.2% | 70.5% |
| DeepSWE v1.1 | 65.9% | 54% | 73% | 70% |
| FrontierCode v1.1 (Extended) | 61.3% | 56.6% | 60.6% | 63.6% |
| APEX-Agents | 57.5% | 47.1% | 56.7% | 59.2% |
| Terminal-Bench v3.0 | 26% | 15.7% | 34.6% | 34.1% |
| AA-Briefcase | 1,577 | 1,313 | 1,502 | 1,574 |