SpaceXAI released the Grok 4.6 model on August 12, 2026. The 1.5-trillion-parameter model showed a significant jump in agentic tasks — the DeepSWE score rose from 54% to 65.9% — while maintaining a cost of $2 per million input tokens and $6 per million output tokens.
What happened
Grok 4.6 is built on the pre-training of Grok 4.5 with 1.5 trillion parameters. The main improvements come from post-training: SFT trajectories were regenerated through the Grok 4.5 model itself, extended reinforcement learning was applied for agentic tasks — kernel optimization, web development, CAD. The optimizer was also updated and training data filtering was applied. The model is available via API, Cursor, Grok Build, OpenRouter, Vercel, and Cloudflare. For the first week, Cursor and Grok Build received doubled usage limits.
Context
SpaceXAI was previously known as xAI. The company maintains an aggressive update schedule: after the release of Grok 4.6, Grok 4.7 with 2.1 trillion parameters and improvements across all metrics with slightly slower inference is promised within a few weeks. The method of regenerating SFT trajectories through the model itself is a form of bootstrapped distillation, which allows improving instruction following without retraining from scratch. In benchmarks, Grok 4.6 achieved an AA Intelligence Index of 61 (on par with GPT-5.6 Sol Max, 1 point below Anthropic Fable 5 Max) and GDPVal-AA v2 — 1753, which is a leading score.
Why this matters for the industry
The price of $2/$6 per million tokens — roughly half that of the closest competitors GPT-5.6 and Fable 5 Max — sets a new price/performance standard in the frontier segment. The rapid release pace, one release every few weeks, puts pressure on OpenAI and Anthropic, forcing them to accelerate their own update cycles and reconsider strategies tied to a single provider. For startups and agentic products, the Grok model lineup could become a de facto infrastructure layer.
Why this matters for users
Developers in Cursor can already test the improved agentic capabilities — the model handles autonomous code checking and multi-step tasks better — without changing their plan. An API key is generated at console.x.ai. Integration via OpenRouter, Vercel, and Cloudflare Workers AI happens without infrastructure changes. The doubled weekly limits in Cursor and Grok Build provide a window for practical testing.
What is still unknown / limitations
The lack of a technical paper and independent verification of benchmarks limits confidence in architectural novelty assessments. The claim of "85% savings" that appeared in a NextBigFuture publication is exaggerated: real data indicates a price reduction of roughly half compared to competitors, not 85%.
Sources
- Introducing Grok 4.6 — SpaceXAI
- SpaceXAI Launches Grok 4.6 — Unite.AI
- NextBigFuture — Grok 4.6 Analysis
Author
Look at AI, editorial team