HealthBench Professional

Microsoft logoMAI-Thinking-1 on HealthBench Professional

Rank 42 of 43, updated November 2, 2026.

MAI-Thinking-1 scores 0.350 on HealthBench Professional, rank 42 of 43 models on the board. Microsoft AI's first in-house frontier reasoning model, a sparse mixture-of-experts design in public preview on Microsoft Foundry. HealthBench Professional scores models on 525 tasks drawn from real clinician conversations, graded criterion by criterion against physician-written rubrics on a 0 to 1 scale.

42
rank of 43
0.350
length-adjusted score
–
score before the length adjustment, not printed in the source
0.353
gap to the leader, GPT-6 Astra (Anthropic run)
256K
context window, tokens
$2.00 / $8.00
input / output price per 1M tokens

Where it sits on the board

The models ranked around MAI-Thinking-1, by length-adjusted score. Differences are against MAI-Thinking-1.

RankModelLength-adjustedBefore the adjustmentDifference
37OpenAI logoGPT-5.20.4590.500+0.109
38Anthropic logoClaude Sonnet 4.60.442–+0.092
39OpenAI logoGPT-5.6 Luna (August)0.4410.468+0.091
40OpenAI logoGPT-5.10.3960.480+0.046
41OpenAI logoGPT-5.5 Instant0.3840.407+0.034
42Microsoft logoMAI-Thinking-10.350––
43Anthropic logoClaude Haiku 4.50.322–-0.028

Result and API facts

Rank42 of 43
Length-adjusted score0.350
Score before the length adjustmentnot printed in the source
LabMicrosoft
Context window256K tokens
API price per 1M tokens$2.00 in / $8.00 out
Licenseproprietary
SourceMAI-Thinking-1: Building a Hill-Climbing Machine (model card)
Parameters1.0T
Released2026-08-12

Source of this score

Read from MAI-Thinking-1: Building a Hill-Climbing Machine (model card, Microsoft AI, 2026-08-12). Vendor-reported. Confidence: verified. Configuration: length-adjusted (HealthBench Professional length penalty), standard GPT-5.4 grader and OpenAI rubrics, Microsoft AI run; printed at integer precision..

p. 54, Table 12 'Post-trained model evaluation results on various public benchmarks', Health group, column HealthBench Prof.; protocol Appendix K.6 p. 106: HealthBench Professional introduces a length penalty for the primary metric, to correct for a well-observed correlation between lengthy responses and artificially increased LLM-grader scores. For all reported scores, we use the standard GPT-5.4 grader and rubrics provided by OpenAI.full entry on the sources page
Model AIR-Bench CyberSec Instruct CyberSec Auto Long Fact Truthful QA HealthBench Prof. MedXpert QA MAI-Thinking-1 88 63 63 98 88 35 43 Sonnet 4.6 88 62 56 98 88 38 49

Position in the field

The gap to the leader, GPT-6 Astra (Anthropic run) at 0.703, is 0.353. Directly above sits GPT-5.5 Instant at 0.384. Directly below sits Claude Haiku 4.5 at 0.322. The scores on this page are compiled from published documents rather than from one controlled run, so a small gap between 2 models can reflect a difference in grader version, reasoning effort, or deployment setting as well as a difference in capability.

What does MAI-Thinking-1 score on HealthBench Professional?

MAI-Thinking-1 scores 0.350 on HealthBench Professional, which places it at rank 42 of 43 models on the board as of November 2, 2026. The number was read from MAI-Thinking-1: Building a Hill-Climbing Machine, listed on the sources page.

How much does MAI-Thinking-1 cost to run?

MAI-Thinking-1 is priced at $2.00 per million input tokens and $8.00 per million output tokens through Microsoft's API.

Pairings with a dedicated comparison page are linked; every other difference is in the score-difference matrix.

  • MAI-Thinking-1 vs GPT-6 Astra (Anthropic run)
    0.350 vs 0.703, GPT-6 Astra (Anthropic run) by 0.353
  • MAI-Thinking-1 vs Claude Sonnet 5.5
    0.350 vs 0.692, Claude Sonnet 5.5 by 0.342
  • MAI-Thinking-1 vs Claude Fable 5
    0.350 vs 0.660, Claude Fable 5 by 0.310
  • MAI-Thinking-1 vs Claude Opus 5.5
    0.350 vs 0.656, Claude Opus 5.5 by 0.306
  • MAI-Thinking-1 vs GPT-6 Luna (Anthropic run)
    0.350 vs 0.655, GPT-6 Luna (Anthropic run) by 0.305
  • MAI-Thinking-1 vs Claude Haiku 5.5
    0.350 vs 0.648, Claude Haiku 5.5 by 0.298
  • MAI-Thinking-1 vs GPT-6 Astra
    0.350 vs 0.647, GPT-6 Astra by 0.297
  • MAI-Thinking-1 vs GPT-6.1 Sol
    0.350 vs 0.642, GPT-6.1 Sol by 0.292
  • MAI-Thinking-1 vs Claude Fable 5 (September card)
    0.350 vs 0.633, Claude Fable 5 (September card) by 0.283
  • MAI-Thinking-1 vs GPT-6.1 Terra
    0.350 vs 0.627, GPT-6.1 Terra by 0.277
  • MAI-Thinking-1 vs Muse Wonder Full
    0.350 vs 0.623, Muse Wonder Full by 0.273
  • MAI-Thinking-1 vs Claude Fable 5.1
    0.350 vs 0.621, Claude Fable 5.1 by 0.271
  • MAI-Thinking-1 vs GPT-6 Luna
    0.350 vs 0.608, GPT-6 Luna by 0.258
  • MAI-Thinking-1 vs GPT-6 Sol
    0.350 vs 0.608, GPT-6 Sol by 0.258
  • MAI-Thinking-1 vs GPT-5.6 Sol
    0.350 vs 0.605, GPT-5.6 Sol by 0.255
  • MAI-Thinking-1 vs Muse Wonder
    0.350 vs 0.605, Muse Wonder by 0.255
  • MAI-Thinking-1 vs Claude Opus 5
    0.350 vs 0.598, Claude Opus 5 by 0.248
  • MAI-Thinking-1 vs Muse Spark 1.1
    0.350 vs 0.593, Muse Spark 1.1 by 0.243
  • MAI-Thinking-1 vs Claude Sonnet 5
    0.350 vs 0.578, Claude Sonnet 5 by 0.228
  • MAI-Thinking-1 vs GPT-5.6 Terra
    0.350 vs 0.577, GPT-5.6 Terra by 0.227
  • MAI-Thinking-1 vs Claude Opus 4.8 (Opus 4.8 grader)
    0.350 vs 0.574, Claude Opus 4.8 (Opus 4.8 grader) by 0.224
  • MAI-Thinking-1 vs Grok 4.7
    0.350 vs 0.567, Grok 4.7 by 0.217
  • MAI-Thinking-1 vs Claude Opus 4.8
    0.350 vs 0.558, Claude Opus 4.8 by 0.208
  • MAI-Thinking-1 vs GPT-5.6 Luna
    0.350 vs 0.557, GPT-5.6 Luna by 0.207
  • MAI-Thinking-1 vs Gemini 4 Argon
    0.350 vs 0.551, Gemini 4 Argon by 0.201
  • MAI-Thinking-1 vs Muse Spark
    0.350 vs 0.541, Muse Spark by 0.191
  • MAI-Thinking-1 vs GPT-5.6 Sol (August)
    0.350 vs 0.540, GPT-5.6 Sol (August) by 0.190
  • MAI-Thinking-1 vs Kimi K3
    0.350 vs 0.532, Kimi K3 by 0.182
  • MAI-Thinking-1 vs GLM 5.3
    0.350 vs 0.528, GLM 5.3 by 0.178
  • MAI-Thinking-1 vs Claude Opus 4.7
    0.350 vs 0.519, Claude Opus 4.7 by 0.169
  • MAI-Thinking-1 vs GPT-5.5
    0.350 vs 0.518, GPT-5.5 by 0.168
  • MAI-Thinking-1 vs Gemini 3.8 Flash
    0.350 vs 0.517, Gemini 3.8 Flash by 0.167
  • MAI-Thinking-1 vs DeepSeek V4.1 Flash
    0.350 vs 0.507, DeepSeek V4.1 Flash by 0.157
  • MAI-Thinking-1 vs Grok 4.6
    0.350 vs 0.485, Grok 4.6 by 0.135
  • MAI-Thinking-1 vs GPT-5.4
    0.350 vs 0.481, GPT-5.4 by 0.131
  • MAI-Thinking-1 vs GPT-5
    0.350 vs 0.462, GPT-5 by 0.112
  • MAI-Thinking-1 vs GPT-5.2
    0.350 vs 0.459, GPT-5.2 by 0.109
  • MAI-Thinking-1 vs Claude Sonnet 4.6
    0.350 vs 0.442, Claude Sonnet 4.6 by 0.092
  • MAI-Thinking-1 vs GPT-5.6 Luna (August)
    0.350 vs 0.441, GPT-5.6 Luna (August) by 0.091
  • MAI-Thinking-1 vs GPT-5.1
    0.350 vs 0.396, GPT-5.1 by 0.046
  • MAI-Thinking-1 vs GPT-5.5 Instant
    0.350 vs 0.384, GPT-5.5 Instant by 0.034
  • MAI-Thinking-1 vs Claude Haiku 4.5
    0.350 vs 0.322, MAI-Thinking-1 by 0.028

Where the scores come from and how they are read is on the methodology page. The full ranking is on the leaderboard.