Grok 4.6: xAI's flagship for long agent runs
xAI shipped Grok 4.6 on August 12, 2026 — the new flagship of the Grok line. It builds on Grok 4.5 and isn’t a full generation change but a targeted reinforcement: long-running agents, more demanding interactive and visual tasks. Two things stand out in practice — a 500,000-token context window and staggered pricing that keeps long contexts predictable.
What's new
- Release: August 12, 2026, new flagship after Grok 4.5.
- Input/output: text and image input, text-only output. Context window: 500,000 tokens, no fixed output-length limit.
- Performance: Artificial Analysis Intelligence Index 61 — tied with GPT-5.6 Sol, one point behind Claude Fable 5.
- Pricing (staggered, per 1M tokens): under 200k prompt tokens $2 input / $0.50 cached input / $6 output; above that $4 / $1 / $12.
What happened
Grok 4.6 sharpens its predecessor’s profile rather than reinventing it. The emphasis is on agentic work over long stretches — tasks the model runs on its own across many steps, with intermediate evaluations and course corrections. Add better visual capabilities: Grok 4.6 takes text and image in and puts text out.
On the vendor-independent Artificial Analysis Intelligence Index the model lands at 61 — tied with GPT-5.6 Sol and one point behind Claude Fable 5. That’s frontier-adjacent, not frontier-leading; the real lever sits elsewhere.
Why it matters
For agent pipelines, the raw benchmark number rarely decides — the combination of context size and cost does. That’s exactly where Grok 4.6 positions itself: 500,000 tokens of context are enough to hold long histories, large documents or whole tool traces in one pass — and the staggered pricing keeps that affordable as long as the prompt stays under 200k tokens. Only above that does the rate double. Keep your prompts deliberately below that threshold and you run noticeably cheaper.
The second point is availability. Grok 4.6 runs not only through the xAI API (Responses and Chat Completions) but also in Cursor, Grok Build, OpenRouter, Vercel and Cloudflare — and across the major cloud marketplaces Microsoft Foundry, Google Enterprise Agent Platform and Amazon Bedrock. For a model from the xAI orbit, whose reach outside the X platform was long modest, that breadth of access is the real progress.
One caveat remains: the 500k context is a ceiling, not a quality guarantee across the full length — as with virtually all models, reliability drops toward the end of long contexts. Test that against your own use case rather than inferring it from the spec sheet.
What you can do now
If you build agents with long contexts: check Grok 4.6 on a real, multi-hour task and measure cost at realistic prompt sizes — the 200k threshold noticeably decides the rate.
If you source models through cloud marketplaces: Grok 4.6 is now broadly available and, for the first time, usable without an X tie-in. Details on benchmarks, the pricing tiers and access options are in the glossary entry → Grok 4.6, the read on the line as a whole in the hub → Grok.
Discover more
GPT-6 Astra Found Questions My First Security Review Missed
I used GPT-6 Astra and Fable 5.1 as independent reviewers for RLS, API and tenant-isolation checks. The useful part was the structured cross-review.
GlossaryGrok 4.6
Grok 4.6 (August 2026) is xAIs flagship model for long-running agentic work — 500k-token context, text and image input, tiered pricing from 2/6 USD per million tokens.
EncyclopediaLLM Model Families at a Glance
Claude, GPT, Gemini, Llama & co. — who builds what, where each family shines, and how to pick the right model for your own use case.