Grok 4.20 just set the honesty record — but ranks 8th on smarts

xAI released Grok 4.20 on March 10, 2026, and it immediately set a record that no other AI had achieved: the lowest hallucination rate of any model tested. Hallucination (when an AI confidently states something false as if it were true) is one of the most persistent and costly problems in AI deployment. Grok 4.20 achieves a 78% non-hallucination rate on the Artificial Analysis Omniscience test — meaning it gets the answer right, or admits it doesn't know, 78% of the time on hard factual questions where other models fabricate answers.

The twist: despite this honesty record, Grok 4.20 ranks only 8th globally on the Artificial Analysis Intelligence Index with a score of 48. Gemini 3 Pro, Claude Opus 4.6, and GPT-5.2 all score higher on raw reasoning. xAI deliberately optimized for factual caution at the cost of peak reasoning performance — a tradeoff that matters enormously depending on what you're building.

What Hallucination Rate Actually Means (And Why 78% Is Remarkable)

Most AI models, when asked a question they don't actually know the answer to, will generate a plausible-sounding response anyway. This is hallucination — the model doesn't "know" it's wrong, it just pattern-matches to what a confident, knowledgeable answer would look like. For casual chatbot use, this is annoying. For legal, medical, financial, or research applications, it's dangerous.

The IFBench score (82.9%, #1 globally) measures something different from intelligence: does the model do what you actually asked? This is distinct from whether the answer is brilliant — it's about reliability. Grok 4.20 won't reinterpret your instructions, add unrequested caveats, or veer off into related topics you didn't ask about. It does the task as specified.

Three Variants — Which One to Use

xAI released Grok 4.20 in three API variants, each optimized for different use cases:

• grok-4.20-0309-reasoning — Activates a chain-of-thought reasoning mode (the model shows its work before answering). Best for complex multi-step problems, mathematical analysis, or situations where you want to verify the AI's logic. Slower, but more accurate on hard problems.
• grok-4.20-0309-non-reasoning — Fast, direct responses without extended reasoning. Best for factual Q&A, summarization, classification, or high-volume production workloads where speed matters. This is what runs at 265 tokens per second.
• grok-4.20-multi-agent-0309 — Coordinates multiple Grok instances to tackle complex tasks in parallel. Best for research workflows, multi-step data processing, or tasks that benefit from multiple independent AI passes.

The Rapid Release Cycle

On March 17, Elon Musk confirmed that Grok 4.20.1 had shipped the day before, with point releases rolling out every 3–4 days containing "significant capability improvements." This is an unusual cadence for a frontier AI model — most major labs ship once every few months. xAI appears to be running a near-continuous deployment model, where the API model improves faster than users can test it.

Grok 4.20 is now the default model on grok.com, the X app, and the xAI API, replacing Grok 3 entirely for consumer chat.

Who Should Choose Grok 4.20 Over GPT or Claude

The case for Grok 4.20: If your application has a high cost of hallucination — legal document analysis, medical information retrieval, financial data processing, customer-facing factual Q&A — the 78% honesty rate is a material advantage. Getting a confident wrong answer is often worse than getting no answer at all.

The case against: If you need best-in-class reasoning, creative synthesis, or complex problem-solving, the #8 ranking on the Intelligence Index means models like Claude Opus 4.6 or GPT-5.2 will outperform it on difficult tasks. The tradeoff is real.

The cost argument: 60% cheaper than Grok 3, with the highest output speed (265 tokens/sec) in its class, makes Grok 4.20 competitive on pure economics for high-volume workloads where accuracy requirements are high but reasoning difficulty is moderate.

How to Try It

pip install openai

from openai import OpenAI
client = OpenAI(
    api_key="YOUR_XAI_API_KEY",
    base_url="https://api.x.ai/v1"
)
response = client.chat.completions.create(
    model="grok-4.20-0309-non-reasoning",  # fastest variant
    messages=[{"role": "user", "content": "Summarize Q4 2025 earnings for Apple"}]
)
print(response.choices[0].message.content)

Get an API key at console.x.ai. The API uses the same format as OpenAI, so switching from existing code is straightforward. Full release notes are at docs.x.ai.

Related Content — Get Started with Easy Claude Code | Free Learning Guides | More AI News