XNews

Grok 4.6 Model Launch and Early Impressions

Grok 4.6 Model Launch and Early Impressions

xAI released Grok 4.6 on August 13, 2026 at unchanged pricing and with availability across Grok Build, Cursor, Grok Bot, and the API. Price remained $2 per million input tokens and $6 per million output tokens. The model has a 500,000-token context window, text and image input, text output, and supports function calling, structured output, and reasoning. Grok 4.6 focused on long-range agent tasks and heavier interaction and visual work; longer trajectories showed more self-testing and verification, and first visual drafts were stronger than 4.5. The Artificial Analysis composite index placed Grok 4.6 at 61, equal to GPT-5.6 Sol Max, five points above Grok 4.5's 56, and one point below Fable 5 Max at 62. All 10 metrics rose versus 4.5; DeepSWE increased from 54% to 65.9%, APEX-Agents from 47.1% to 57.5%, Terminal-Bench from 15.7% to 26%, GDPVal-AA from 1526 to 1753, and AA-Briefcase from 1313 to 1577. APEX-SWE added 2.8 points. A Cursor team member used Grok 4.6 for browser navigation, functional and visual QA, inbox triage, drafting release posts, and Remotion video creation. Stress tests included a spreadsheet app, browser Age of Empires 2, MSN Messenger, Excalidraw presentation mode, a fictional quarterly board deck, and a 60-90 second X TypeScript SDK launch video. Reported strengths included dense summary information, quiet behavior on small changes, narration on multi-file changes, and faster and smarter behavior than 4.5. Prompt tests found wording such as "work very hard" added little, length helped through specificity, and acceptance criteria drove results. Self-verifiability was easiest for websites, harder for 3D, harder for video, and similar for physical work. Musk said Grok 4.7 was expected within about a month, called SpaceX training data unique, and praised Anthropic without mentioning OpenAI.

Key insights

Grok 4.6 launch and availability

Grok 4.6 was launched and made available in Grok Build, Cursor, Grok Bot, and the API. Pricing remained the same as Grok 4.5 at $2 per million input tokens and $6 per million output tokens. The model has a 500,000-token context window, accepts text and image inputs, produces text output, and supports function calling, structured output, and reasoning.

Benchmark positioning relative to GPT-5.6 and Fable 5

The Artificial Analysis composite index placed Grok 4.6 at 61, equal to GPT-5.6 Sol Max, five points above Grok 4.5's 56, and one point below Fable 5 Max at 62. All ten metrics measured by the team rose relative to Grok 4.5. DeepSWE increased from 54% to 65.9%, APEX-Agents from 47.1% to 57.5%, Terminal-Bench from 15.7% to 26%, GDPVal-AA from 1526 to 1753, and AA-Briefcase from 1313 to 1577.

Longer agentic and visual tasks strengthened

Grok 4.6 was built with two areas strengthened relative to 4.5: long-range agent behavior for research, analysis, codebase edits, and turning ideas into usable products; and heavier interaction and visual work such as establishing application structure and visual language in one pass before iteration. Team observations included more self-testing and verification on longer trajectories, and first drafts on visual and interaction projects were notably stronger than 4.5.

Practitioner report on information density and speed

A Cursor team member used Grok 4.6 for weeks as a daily driver. Reported use included browser navigation such as clicking API keys in supplier consoles, functional and visual QA on running applications, inbox triage, drafting release posts, and making Remotion videos. The user valued that summaries contained real information rather than task restatement, and that short updates were enough to decide whether to interrupt. The model was quiet during small changes and narrated once many files changed, though it still sometimes said unnecessary things. Grok 4.6 was reported as fast and noticeably smarter than 4.5, pulling the user back to synchronous work when the result mattered.

Prompt style and acceptance criteria

Tests comparing long and short prompts found that wording such as "work very hard" added almost nothing for Grok 4.6. Longer prompts helped when they added specificity about known requirements; short prompts let the model's taste make more decisions, and the model's taste had improved enough for that to work in many cases. Long specifications could still be followed from start to finish, but components could repeat unless explicitly separated. The strongest lever was an acceptance instruction that defined how to verify the work.

Self-verifiability as core constraint

The practitioner's most general observation was that the hardest tasks were those where the model's work was hard to verify. Websites were easiest because the DOM is text and pages can be read, screenshotted, and checked against intent. 3D work was harder because of a dimension that cannot be read directly. Video was harder because time requires reasoning across frames. Physical work was similar because a single screenshot cannot answer whether motion occurred correctly.

Grok 4.7 expectations and Musk statements

Musk said Grok 4.7 was expected within about a month. He stated that SpaceX's training data was exceptionally strong and unique, and that he would be surprised if any model had stronger real engineering capability than 4.7. He also praised Anthropic as a great company that would likely soon release a better model, while leaving OpenAI unmentioned.

Source posts