xAI's flagship model (March 2026) with 2M token context window and deep reasoning. Best for complex multi-step tasks, analysis, and research with function calling and structured outputs.
Grok 4.20 Reasoning is xAI's reasoning-enabled Grok 4.20 API variant, pinned on TheRouter as xai/grok-4.20-reasoning and routed to the March 2026 grok-4.20-0309-reasoning snapshot. It keeps the Grok 4.20 family's 2-million-token context window and text+image input while enabling explicit chain-of-thought style reasoning for complex logic, math, coding, and long-horizon agent workflows.
In production routing, this model is the depth-first member of the Grok 4.20 line. Use it when you need the 2M context window plus stronger reasoning than the non-reasoning speed variant can provide. BenchLM's July 25, 2026 profile lists Grok 4.20 as a proprietary reasoning model with a 53.92/100 public score, 1474 Chatbot Arena Elo, and published evidence across agentic, coding, reasoning, knowledge, and multimodal categories.
The trade-off is latency and token volume. Reasoning mode improves hard-task accuracy but can spend extra hidden or visible reasoning tokens before the final answer, so it is not the cheapest choice for simple extraction, classification, or templated copy. For new xAI integrations that do not require 2M context, Grok 4.3 is often the newer default; for pinned long-context reasoning, Grok 4.20 Reasoning remains the safer stable target.
xAI exposes Grok 4.20 Reasoning as a native xAI API model with xAI's pricing tiers, reasoning-token accounting, server-side tools, and the documented caveat that logprobs/top_logprobs are silently ignored for Grok 4.20 and newer.
TheRouter serves the pinned xai/grok-4.20-reasoning route through its OpenAI-compatible chat endpoint with TheRouter pricing from standard-models.yaml; unsupported media-generation, audio, and embeddings endpoints are not exposed for this model.
| Type | Rate |
|---|---|
| Prompts under 200K tokens | |
| Input | $1.35 / 1M tokens |
| Output | $2.70 / 1M tokens |
| Cached input | $1.35 / 1M tokens |
| Prompts of 200K tokens or more | |
| Input | $2.70 / 1M tokens |
| Output | $5.40 / 1M tokens |
| Cached input | $0.432 / 1M tokens |
Tiered pricing: when a request's prompt (input, cached tokens included) is 200K tokens or more, every token in that request is billed at the long-context rates. This model publishes no separate cached-input price for the base group: cached tokens there are billed at the base input rate (that group's "Cached input" row shows that rate).
| Release date (public API) | Not publicly disclosed | unknown |
| API model slug (pinned) | grok-4.20-0309-reasoningdocs.x.ai ↗ | verified |
| Reasoning mode | Configurable reasoning is supported on current xAI reasoning models; xAI documents reasoning tokens and the reasoning.effort control, but does not publish a separate Grok 4.20 Reasoning model card.docs.x.ai ↗ | verified |
| Context window | 2,000,000 tokensdocs.x.ai ↗ | verified |
| Max completion tokens | 65,536 tokens on TheRouter | verified |
| Input modalities | Text, imagedocs.x.ai ↗ | verified |
| Output modalities | Text | verified |
| TheRouter pricing | $3.00 input / $9.00 output per 1M tokens on TheRouter | verified |
| Upstream list pricing | Short context: $1.25 input / $0.20 cached / $2.50 output per 1M tokens; long context (≥200k prompt tokens): $2.50 input / $0.40 cached / $5.00 output per 1M tokensdocs.x.ai ↗ | verified |
| Function calling / tools | Supported | verified |
| Structured outputs (response_format) | Supported (JSON mode) | verified |
| Knowledge cutoff | Not publicly disclosed | unknown |
| License | Proprietary | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
BenchLM public score BenchLM profile for Grok 4.20 Beta Reasoning; data verified July 25, 2026. BenchLM marks the global score provisional because coverage is 19 of 369 tracked benchmarks. | — | Not publicly disclosed | — |
Chatbot Arena Elo — Text Overall BenchLM reports ±4.7 confidence interval from 26,822 votes. | — | Not publicly disclosed | — |
SWE-bench Verified BenchLM secondary exact row for the Grok 4.20 reasoning family. | — | Not publicly disclosed | — |
LiveCodeBench Pro | — | Not publicly disclosed | — |
Terminal-Bench 2.0 | — | Not publicly disclosed | — |
DeepSearchQA | — | Not publicly disclosed | — |
ARC-AGI-2 | — | Not publicly disclosed | — |
GPQA Diamond | — | Not publicly disclosed | — |
HLE without tools | — | Not publicly disclosed | — |
MMMU-Pro | — | Not publicly disclosed | — |
CharXiv Reasoning | — | Not publicly disclosed | — |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "xai/grok-4.20-reasoning",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Use the standard OpenAI-compatible chat completions endpoint. Keep the model id pinned to xai/grok-4.20-reasoning when you need stable Grok 4.20 reasoning behavior instead of a moving latest alias.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "xai/grok-4.20-reasoning",
"messages": [
{"role": "system", "content": "Reason carefully, cite uncertainty, and give the shortest defensible answer."},
{"role": "user", "content": "Compare these two migration plans and identify the hidden failure modes."}
],
"reasoning": {"effort": "high"}
}'BenchLM's July 25, 2026 profile lists Grok 4.20 as a proprietary reasoning model with 2M context, 1474 Chatbot Arena Elo, and evidence across coding, agentic, reasoning, knowledge, and multimodal tasks. The page also marks the model as superseded by newer Grok generations for users who do not need the 2M context window.
xAI's models and pricing pages list grok-4.20-0309-reasoning with 1M context, short-context prices of $1.25 input / $2.50 output per 1M tokens, long-context prices of $2.50 / $5.00, and the caveat that logprobs/top_logprobs are ignored for Grok 4.20 and newer.
Choose the reasoning variant when the task needs multi-step planning, mathematical or scientific reasoning, code repair, tool-call decisions, or cross-document synthesis. Choose the non-reasoning variant when the task is simple, latency-sensitive, or high-volume — for example extraction, classification, summarisation, or templated transformations where reasoning tokens do not buy much accuracy.
The listed per-token price on TheRouter is the same family rate as Grok 4.20 Non-Reasoning ($3.00 input / $9.00 output per 1M tokens on TheRouter). The effective cost per completed task can still be higher because reasoning mode may generate additional reasoning tokens and take longer before producing the final answer.
Yes. Through TheRouter, call https://api.therouter.ai/v1/chat/completions with the OpenAI SDK, set baseURL to TheRouter, and use model: "xai/grok-4.20-reasoning". The model supports chat, streaming, tool calls, structured outputs, image input, and TheRouter's standard reasoning option; it is not used for embeddings, audio, image generation, or video generation.
Use Grok 4.3 when you want the newer xAI flagship and 1M context is enough. Keep Grok 4.20 Reasoning when the 2M context window, a pinned March 2026 snapshot, or compatibility with an existing Grok 4.20 integration is the decisive requirement.
| source | URL | retrieved | |
|---|---|---|---|
| Release date (public API) | — | — | unknown |
| API model slug (pinned) | docs.x.ai ↗ | 2026-07-31 | verified |
| Reasoning mode | docs.x.ai ↗ | 2026-07-31 | verified |
| Context window | docs.x.ai ↗ | 2026-07-31 | verified |
| Max completion tokens | — | — | verified |
| Input modalities | docs.x.ai ↗ | 2026-07-31 | verified |
| Output modalities | — | — | verified |
| TheRouter pricing | — | — | verified |
| Upstream list pricing | docs.x.ai ↗ | 2026-07-31 | verified |
| Function calling / tools | — | — | verified |
| Structured outputs (response_format) | — | — | verified |
| Knowledge cutoff | — | — | unknown |
| License | — | — | verified |
| BenchLM public score | benchlm.ai ↗ | 2026-07-26 | unknown |
| Chatbot Arena Elo — Text Overall | benchlm.ai ↗ | 2026-07-26 | unknown |
| SWE-bench Verified | benchlm.ai ↗ | 2026-07-26 | unknown |
| LiveCodeBench Pro | benchlm.ai ↗ | 2026-07-26 | unknown |
| Terminal-Bench 2.0 | benchlm.ai ↗ | 2026-07-26 | unknown |
| DeepSearchQA | benchlm.ai ↗ | 2026-07-26 | unknown |
| ARC-AGI-2 | benchlm.ai ↗ | 2026-07-26 | unknown |
| GPQA Diamond | benchlm.ai ↗ | 2026-07-26 | unknown |
| HLE without tools | benchlm.ai ↗ | 2026-07-26 | unknown |
| MMMU-Pro | benchlm.ai ↗ | 2026-07-26 | unknown |
| CharXiv Reasoning | benchlm.ai ↗ | 2026-07-26 | unknown |
| BenchLM verifies Grok 4.20 Reasoning profile with 19 public benchmark rows | benchlm.ai ↗ | 2026-07-26 | verified |
| xAI docs list Grok 4.20 pricing and compatibility caveats | docs.x.ai ↗ | 2026-07-31 | verified |
| When should I choose xai/grok-4.20-reasoning instead of xai/grok-4.20-non-reasoning? | benchlm.ai ↗ | 2026-07-26 | to verify |
| Should new xAI workloads use Grok 4.20 Reasoning or Grok 4.3? | benchlm.ai ↗ | 2026-07-26 | to verify |