Grok 4.20 Reasoning is a current pinned xAI API model with a million-token context window.
xAI's Grok 4.20 model card reports a 30% AgentHarm violation rate, 33% AgentDojo prompt-injection success rate, and 32% misuse violation rate under system-prompt override. These figures apply only to the documented tests.
ActionDo not use the model directly for autonomous high-risk decisions; minimize and isolate agent tool permissions.
Open source| Company | xAI |
|---|---|
| Release | 2026-03-10 |
| Release type | Model release |
| Parameters | Not disclosed |
| Architecture | Not disclosed |
| Context | 1,000,000 |
| Input modalities | text / image |
| Output modalities | text / tool_call |
| Tool use / function calling | Confirmed: Tool use, Function calling |
| Structured output / reasoning | Confirmed: Reasoning |
| Vision / audio / video | Confirmed: Vision |
| Open weights / License | No open weights currently / weight license not disclosed xai-grok-4-20-reasoning-current · License not applicable |
| Vendor API | Available Grok 4.20 Reasoning | xAI Docsxai-grok-4-20-reasoning-current · 2026-03-10 |
| OpenRouter | Not separately listed on OpenRouter |
| Pricing | Official · 2026-03-10: $1.25 / $2.5 input / output, per 1M tokens |
| Benchmark | No reliable public source |
| Public signals | No reliable public source |
| Use cases | Long-context analysis / coding and review / Vision understanding / Agent workflows |
| Claim evidence | Release event / Model specificationsGrok 4.20 Reasoning | xAI Docsgrok-4.20-0309-reasoning · 2026-03-10 · Grok 4.20 Reasoning · grok-4.20-0309-reasoning |
| OpenRouter | Not separately listed on OpenRouter |
|---|---|
| OpenRouter listing | Not separately listed in the current OpenRouter public list |
| Mapped version | grok-4.20-0309-reasoning xai-grok-4-20-reasoning-current |
| Pricing | Official · 2026-03-10: $1.25 / $2.5 input / output, per 1M tokens |
| Context | 1000000 |
| Speed/Latency | No reliable public source |
| Snapshot time | 2026-08-11T06:28:09.895Z |
| Source | View source |