Claude, Gemini, Grok and DeepSeek IQ Scores: What the Tests Show

No chatbot has an IQ, yet you will see a Gemini IQ, Claude IQ, Grok IQ and DeepSeek IQ quoted online. A public leaderboard converts each model's puzzle results into one. TrackingAI's log, read on 28 October 2026 with data to 22 October, puts the best Claude entry at 146, Gemini at 143, Grok at 142 and DeepSeek at 103 on its Mensa Norway line. On a private test of the same kind, those four scores fall by 16 to 19 points. The numbers move with the test, the model version and even whether the model sees pictures.

Below: the scores by model family, why they differ between tests, and what the number can and cannot mean. We use the public log, not a vendor claim. Our earlier report on the Mensa Norway test and the 151 headline explains how the scale works.

What are the Gemini IQ, Claude IQ, Grok IQ and DeepSeek IQ scores?

TrackingAI gives each AI entry the Mensa Norway test of 35 pattern puzzles, plus a second, unpublished test that has never been online. The chart uses the average of an entry's last seven sittings. The lists below give that average, the sitting count and the IQ that TrackingAI's line assigns. Entry names are TrackingAI's labels, with the model it says each one runs.

Claude first.

• Claude-5-Opus-Max (claude-opus-5.5, MAX thinking): 11 Mensa Norway sittings, last on 16 Oct 2026; average of its last 7 sittings 33.4 of 35 correct, IQ 146 on TrackingAI’s line; unpublished offline test IQ 130
• Claude-5-Opus (claude-opus-5.5): 12 Mensa Norway sittings, last on 15 Oct 2026; average of its last 7 sittings 32.9 of 35 correct, IQ 145 on TrackingAI’s line; unpublished offline test IQ 128
• Claude-5-Sonnet (claude-sonnet-5): 12 Mensa Norway sittings, last on 14 Oct 2026; average of its last 7 sittings 29.6 of 35 correct, IQ 135 on TrackingAI’s line; unpublished offline test IQ 116

Gemini and Grok follow. Gemini Pro is labelled legacy by TrackingAI and was last sat in September; the Grok 4 entry was last sat in July.

• Gemini Thinking (gemini-3.8-flash): 77 Mensa Norway sittings, last on 16 Oct 2026; average of its last 7 sittings 32.3 of 35 correct, IQ 143 on TrackingAI’s line; unpublished offline test IQ 127
• Gemini Pro (gemini-3.1-pro-preview): 87 Mensa Norway sittings, last on 24 Sep 2026; average of its last 7 sittings 32.9 of 35 correct, IQ 145 on TrackingAI’s line; unpublished offline test IQ 128
• Grok-4-5 (grok-4.7 high): 16 Mensa Norway sittings, last on 21 Oct 2026; average of its last 7 sittings 31.9 of 35 correct, IQ 142 on TrackingAI’s line; unpublished offline test IQ 125
• Grok-4 (grok-4.3): 46 Mensa Norway sittings, last on 22 Jul 2026; average of its last 7 sittings 33.0 of 35 correct, IQ 145 on TrackingAI’s line; unpublished offline test IQ 123

DeepSeek scores far lower. TrackingAI maps its deepseek-v3 entry to the deepseek-v4-pro model, last sat on 10 September 2026, and its DeepSeek-R1 entry was last sat in May.

• deepseek-v3 (deepseek-v4-pro): 55 Mensa Norway sittings, last on 10 Sep 2026; average of its last 7 sittings 19.1 of 35 correct, IQ 103 on TrackingAI’s line; unpublished offline test IQ 84
• DeepSeek-R1 (deepseek-r1-turbo): 48 Mensa Norway sittings, last on 20 May 2026; average of its last 7 sittings 21.1 of 35 correct, IQ 109 on TrackingAI’s line; unpublished offline test IQ 93

Why do the public and offline scores differ?

The Mensa Norway test is public, so a model may have met its puzzles in training. The offline test was written by a Mensa member and never posted online. For 12 of the 13 entries we compared, the public-test IQ is 10 to 22 points higher than the offline IQ. The exception, Claude-5-Sonnet-Vision, scores 109 on both. The median gap across all 13 is 17 points.

The offline test is also short. TrackingAI scores 14 items, so one item moves the result by about 6.6 points and the maximum is 136. The top offline entry is Claude-5-Opus-Max-Vision at IQ 135 on 5 sittings, which is close to that ceiling.

Does it matter if the model sees the pictures?

Yes, a lot for some models. TrackingAI runs a text-only form, where the puzzles are described in words, and a Vision form that reads the images. The Vision form scored lower on every pair we compared.

• Claude-5-Sonnet: IQ 135 as text, 109 with Vision (26 points lower)
• Gemini Pro: IQ 145 as text, 130 with Vision (14 points lower)
• Grok-4-5: IQ 142 as text, 127 with Vision (14 points lower)
• Claude-5-Opus: IQ 145 as text, 142 with Vision (3 points lower)

Is an AI IQ the same as a human IQ?

No. A human IQ is a position inside a norm group of people who took the same test. TrackingAI's figure is a straight-line conversion of the number of right answers: IQ = 63.5 + 3 x (correct answers – 5.833). Each right answer adds 3 points, and 35 of 35 gives 151.

That makes anything above Mensa Norway's own human limit of 145 an extension of the line, not a score a person can receive. A model that is excellent at pattern puzzles may be ordinary at planning or at long tasks. We argue the full case in AI scoring on an IQ-style test is not an IQ, and test the idea on a harder puzzle in ARC-AGI-3.

Which chatbot has the highest IQ?

It depends on the rule. Among entries with at least 7 Mensa Norway sittings, Claude-5-Opus-Max leads the four families here at 146, with Gemini Pro, Claude-5-Opus and Grok-4 at 145. Entries with fewer sittings can look higher, so a small count makes a weak ranking. Our ChatGPT IQ score guide covers the OpenAI side of the same log.

What should you check before sharing an AI IQ chart?

• How many sittings stand behind the figure, because an average of seven needs seven
• Whether the number is the public test or the offline test
• Whether the entry is a text form or a Vision form
• Whether the entry is a current model or a legacy one
• Whether the score is a rescaled count or a deviation from a human norm group

The short version: Claude, Gemini and Grok entries read about 142 to 146 on a public puzzle test and lower on a private one, DeepSeek reads about 103 to 109, and none of those numbers is an IQ in the human sense. For a place among people, take the IQ Metrics test or use our IQ percentile calculator.

Common questions