Grok 4.6 ships day one to Cursor, OpenRouter, Vercel and Cloudflare
xAI's 1.5T model matches GPT-5.6 Sol on the AA Index and adds an xhigh tier — but past 200K tokens the whole prompt reprices. Plus Claude Code 2.1.229.
Copy markdown
One model, live everywhere at once
Grok 4.6 (1.5T params, 500K context, Feb 1 2026 knowledge cutoff) shipped Aug 12 with same-day access on Cursor, Grok Build, the xAI API, OpenRouter, Vercel and Cloudflare. It reuses 4.5's V9 base with retuned fine-tuning, and xAI says it matches GPT-5.6 Sol on the nine-benchmark Artificial Analysis Intelligence Index.
The $2/$6 headline hides a 200K cliff
Standard rates are $2 per million tokens in and $6 out — but the moment a prompt crosses 200K tokens the entire request reprices to $4/$12, not just the overflow. Cached input also jumped 67% to $0.50/$1.00 versus 4.5's $0.30/$0.60, so budget your context windows before you wire it into an agent loop.
New xhigh tier, and double usage this week
4.6 adds a fourth reasoning-effort level, xhigh, that 4.5 never had — worth testing on hard agentic or coding runs. xAI is doubling included usage in Cursor and Grok Build for the first week only, making this weekend the cheap window to benchmark it head-to-head against your current model.
Claude Code 2.1.229 stops Bedrock/Vertex dropouts
The Aug 12 release adds SSE keepalive pings that prevent idle-timeout disconnects on Vertex and Bedrock, plus `claude remote-control --continue` to resume sessions and plugin-marketplace `command` sources for local commands. Update if long streaming calls have been dying mid-run.
Elsewhere: hidden reasoning traces were extractable
A new paper showed the encrypted chain-of-thought blocks OpenAI, Anthropic and Google return could be replayed into weaker sibling models to spill the stronger model's hidden reasoning verbatim, with Claude Haiku 4.5 most exposed. All three have since patched it, per the authors.