Published: August 3, 2026 · Updated: September 16, 2026
Grok 4.6 launched on August 12, 2026 — five days after Elon Musk’s initial “around August 7” target. The 1.5-trillion-parameter model scored 1753 ELO on the LMSYS Chatbot Arena, priced at $2 per million input tokens, and shipped same-day to Cursor and xAI’s Grok Build tool, giving AI coding-agent users immediate access.
Grok 4.6 is xAI’s 1.5-trillion-parameter AI model, launched August 12, 2026, at $2 per million input tokens — less than one-seventh the cost of Claude Opus 4.8 at $15/M. It scored 1753 ELO on the LMSYS Chatbot Arena and ships same-day to Cursor, Grok Build, OpenRouter, Vercel, and Cloudflare — giving enterprise coding-agent buyers automatic access to each xAI model upgrade.
What is Grok 4.6?
Grok 4.6 is xAI’s 1.5-trillion-parameter AI agent model, released August 12, 2026, with improved supervised fine-tuning and reinforcement learning over Grok 4.5. It scores 1753 ELO, costs $2/M input tokens via the xAI API, Cursor, and the Grok Build tool, and targets Claude Opus 4.8 and Kimi K3 in the enterprise coding-agent market. The model carries a 500,000-token context window — the same class as Grok 4.5 — and a knowledge cutoff of February 1, 2026.
| Model | ELO | Input $/M | Output $/M | Context |
|---|---|---|---|---|
| Grok 4.6 | 1753 | $2 | $6 | 500K |
| Grok 4.5 | ~1720 | $2 | $6 | 500K |
| Claude Opus 4.8 | ~1760 | $15 | $75 | 200K |
Why Post-Training Gains Matter for Business Users
Grok 4.6 reuses the 1.5-trillion-parameter architecture of Grok 4.5 and extracts its improvements entirely from post-training: supervised fine-tuning, which aligns the model to human preferences on real tasks, and reinforcement learning, which rewards responses that complete tasks correctly. The 1753 ELO score confirms the gains are measurable — Grok 4.5 scored around 1720 ELO on the same scale, according to xAI’s official announcement.
Post-training improvements produce three business-relevant gains: better instruction-following, fewer refusals on legitimate tasks, and more reliable tool use — the capability that determines how well a model runs multi-step agent workflows. Grok 4.5, the predecessor, was trained across tens of thousands of NVIDIA GB300 GPUs and fine-tuned jointly by xAI and Cursor using developer interaction data from real codebases. Grok 4.6 carries the same $2 per million input tokens and $6 per million output tokens pricing that made Grok 4.5 competitive against Opus-class models.
Cursor and Grok Build users got same-day access on August 12 — faster than Grok 4.5’s rollout, which took several days to reach third-party integrations. Beyond those two developer tools, xAI also made Grok 4.6 available via OpenRouter, Vercel, and Cloudflare on launch day. Pricing at 200K+ token prompts steps up to $4/$12 across the entire request, with cached input at $0.50/M. For a detailed comparison of how Grok 4.6 stacks up against Kimi K3 in open-weight coding benchmarks, see our Kimi K3 Review 2026.
Grok 4.7: 2.1T Model — Announced for September, Launch Pending
Grok 4.7, a 2.1-trillion-parameter model, was announced September 2, 2026, with a target release window of approximately September 11–12. As of September 16, 2026, xAI had not published an official launch announcement — this section will update when it ships. The model is a 40% parameter increase over Grok 4.6 and incorporates SpaceX engineering data on rockets, satellites, and manufacturing — a source of domain-specific training data no rival lab can replicate. Musk stated Grok 4.7 “will be better than 4.6 in every way, except slightly slower to serve, albeit with even better token efficiency.”
The sequence puts xAI on a roughly 5-week release cadence: Grok 4.5 shipped July 8, Grok 4.6 landed August 12, and Grok 4.7 is due around September 11–12. If the cadence holds, xAI is the fastest-shipping frontier lab by model-release frequency.
What the Cadence Means for Choosing an Agent Platform
A 5-week model cadence makes version-based procurement obsolete: a model a business licenses in June sits two generations behind by September. Businesses evaluating the best AI tools for business get more durable value from a platform’s distribution channels than from any single model version, because new models flow into existing channels within days. Grok 4.6 is now live through five channels: the xAI API, Grok Build, Cursor, OpenRouter, and Vercel/Cloudflare — the same infrastructure Grok 4.5 established, meaning enterprise buyers who adopted Grok 4.5 got the 4.6 upgrade without a procurement cycle. For context on what drove Grok’s channel expansion, see our earlier coverage: Grok 4.5 Is Here: SpaceXAI’s Opus-Class AI at a Fraction of the Price.
For Context
WithO2 covered xAI’s previous release in Grok 4.5 Is Here: SpaceXAI’s Opus-Class AI at a Fraction of the Price, which detailed the model’s $2/$6 pricing and its GitHub Copilot rollout — the same channels Grok 4.6 now ships through. Related coverage: best AI agents for enterprise.
Frequently Asked Questions
When did Grok 4.6 actually launch?
August 12, 2026 — five days after Musk’s “around August 7” target. xAI published the developer documentation and API model ID (grok-4.6) on that date.
What is Grok 4.6’s ELO score?
1753 ELO on the LMSYS Chatbot Arena scale, up from approximately 1720 for Grok 4.5.
How much does Grok 4.6 cost?
$2 per million input tokens, $6 per million output tokens. Cached input costs $0.50/M. Prompts of 200K tokens or more are billed at $4/$12 across the entire request.
Where is Grok 4.6 available?
Through the xAI API, Grok Build tool, Cursor, OpenRouter, Vercel, and Cloudflare — all available from launch day, August 12, 2026.
What is Grok 4.6’s context window?
500,000 tokens, the same class as Grok 4.5. Knowledge cutoff is February 1, 2026.
When does Grok 4.7 release?
Announced September 2, 2026, targeting September 11–12. It uses 2.1 trillion parameters — a 40% increase from Grok 4.6 — and incorporates SpaceX engineering data.
Last Updated: September 3, 2026
