Grok 4.20 Reasoning

What is Grok 4.20 Reasoning?#

Grok 4.20 Reasoning is the thinking model in xAI's Grok 4.20 line, a model built for speed, agentic tool calling, and strict prompt adherence, with a low hallucination rate as its headline claim. It went live on the API on March 10, 2026 as the 0309 snapshot and got refreshed weights in April 2026. It accepts text and image input, outputs text, and runs on Unifically as xai/grok-4.20-0309-reasoning. The context window is 1M tokens. Grok 4.20 Non-Reasoning and Grok 4.20 Multi-Agent are separate models on the same base, each with its own page.

Key features of Grok 4.20 Reasoning#

Fast for a thinking model

Artificial Analysis measures 98.9 output tokens per second, #46 of 195 models and well above the 70 median, with an Intelligence Index of 38 that beats the class median of 36. At launch its median speed on xAI's API was 267 tokens per second, so expect it to get faster again as capacity frees up.

Highest Arena rank in the Grok line

The 0309 reasoning build holds #35 on the Arena text leaderboard with an Elo of 1472 across 62,263 votes as of September 2, 2026, ahead of Grok 4.5 at #38 and Grok 4.6 high at #48. Human voters like its answers more than the raw benchmarks suggest.

Agentic tool calling with strict adherence

Function calling and structured outputs on the standard request, and a model xAI tuned to follow the prompt exactly. Grok 4.3 later pushed the same tool-calling line to 98% on Tau2-Bench Telecom.

A million tokens of context

Whole repositories, long transcripts, or document sets in one request, with image input for screenshots and charts.

Best for#

Use cases#

Use Grok 4.20 Reasoning as the general model behind assistants that need a little thought per turn: support agents that call tools, research helpers that digest long threads, and internal chat over a million tokens of documents. It is a sensible default for conversational products, where Arena voters rank it above every other Grok model, and for pipelines that mix extraction with judgment calls. For routine code it holds its own; when the task turns into a real engineering job, hand off to Grok 4.5 or Grok 4.6, which share the same API shape.

Limitations#

Thinking costs time. Artificial Analysis measures about 16 seconds to first token at the default effort, so it is the wrong pick for chat that has to feel instant; Grok 4.20 Non-Reasoning answers in 0.64 seconds.

Its Intelligence Index of 38 and 72.2% SWE-bench Verified put it a step below Grok 4.5 and Grok 4.6 on hard problems, and Grok 4.3 now delivers the same price point with better tool-calling scores.

The 0309 ID has been refreshed once already; re-test prompts after each xAI weight update.

Grok 4.20 Reasoning vs Non-Reasoning#

Same weights, same price, one difference. Reasoning scores 38 on the Intelligence Index against 22 for Non-Reasoning and wins any task with a hidden step in it. Non-Reasoning answers in 0.64 seconds to first token instead of about 16, at 95.8 tokens per second, and is the better fit for extraction, routing, and short replies. Run Reasoning by default and drop to Non-Reasoning where you can measure that thinking does not change the output.

When to use Grok 4.20 Reasoning#

Use it when you want a 1M-context model that thinks, calls tools, and ranks well with human voters, at a mid-range cost. Skip it when first-token latency is the product, or when the work is frontier coding, where the newer Grok models are clearly better.

API examples#

Call Grok 4.20 Reasoning from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/xai/grok-4.20-0309-reasoning.

curl -X POST https://api.unifically.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "xai/grok-4.20-0309-reasoning",
    "messages": [
      { "role": "user", "content": "Summarize the open risks in this incident thread and rank them by blast radius." }
    ]
  }'

The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.