Grok 4.6
Language Available Full comparison ↗ xai/grok-4.6 · by xAI
· decoder-only-transformer
Pricing — 1 offering(s)
Input tokens
- $2.00 / 1M tokens (input)
Current
2026-08-12 → present
Standard-variant rate, confirmed via the launch announcement's own body text (raw HTML, not an AI-summarized…
Standard-variant rate, confirmed via the launch announcement's own body text (raw HTML, not an AI-summarized fetch — this repo's CLAUDE.md flags AI summaries as unreliable for pricing tables). xAI also serves a "fast" variant at 2x the price ($4/1M) with 2x the output speed — same convention as this registry's other multi-tier entries (e.g. gpt-5.6-luna's long-context rate): the base/standard rate is tracked here, the variant is noted, not tracked as a separate tier.
Output tokens
- $6.00 / 1M tokens (output)
Current
2026-08-12 → present
Standard-variant rate (see input tier's note for the fast-variant caveat: $12/1M at 2x output speed, not…
Standard-variant rate (see input tier's note for the fast-variant caveat: $12/1M at 2x output speed, not tracked as a separate tier).
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Grok 4.6 fits into a cost-aware routing setup
See how →Capability profile
How Grok 4.6 rates across core capability dimensions, with the task-level evidence behind each rating.
Headline focus per xAI's own launch framing: "long-running agents and more ambitious interactive and visual work" — sustained reasoning across many steps (research, codebase navigation, app-building), with improved self-testing/verification behaviors along the way.
CursorBench 3.2: 69.9%. DeepSWE v1.1 (pass@1): 65.9%. FrontierCode v1.1 (Extended): 61.3%. Vendor-reported (xAI launch page, x.ai/news/grok-4-6, raw HTML-verified — not an AI-summarized fetch). No SWE-bench Verified/Pro score published for this model; these are xAI's own comparison benchmarks, not yet formal anchors in the companion modelglass-coding registry.
APEX-Agents: 57.5% (agentic task-completion benchmark). Vendor-reported, xAI launch page. Agentic tool use is this release's headline capability, not an incidental feature.
500K tokens reported — see architecture note above on verification confidence.
Standard variant. A "fast" variant exists at 2x price with 2x output speed (per xAI's own launch copy) — not separately rated here, see the registry pricing entry's notes.
Matches GPT-5.6 Sol's AA Intelligence Index score (61) at roughly a fifth of Sol's price ($2/$6 vs $4/$20 per 1M tokens) — see closest_competitors on the registry pricing entry.
Operator guidance
Matches GPT-5.6 Sol's general-intelligence score at a fraction of the price — a strong default for agentic/coding tasks where budget matters. Superseded by Grok 4.7 (six weeks later) for anything prioritising the latest ceiling over cost.
Use cases
- Long-running agentic workflows (research, codebase navigation, app-building)
- General-purpose coding and knowledge work
- Tasks needing sustained multi-step reasoning with self-verification
Limitations
- Superseded by Grok 4.7 within six weeks of its own release — evaluate whether the newer model's price/performance still favors 4.6 before routing to it.
- Capability ratings beyond the cited benchmark figures are qualitative, not single cited benchmark runs.