RankingMAI-Thinking-1

MAI-Thinking-1

Data updated 12 Sept 2026

18 published benchmark measures · 9 benchmark families contribute across 4 task areas. 2 capability estimates available in the full profile. See all results ↓ · Compare published benchmarks →

Model evidence summary

Choose a configuration

A model-wide score would mix effort settings. This page summarizes published evidence; only identified configurations receive leaderboard ranks.

Compare measured configurations →

The capability chart summarizes evidence across reported settings, not a runnable configuration or a leaderboard rank.

Performance profile

Capabilities

Adjusted comparison score · 0–100

Filled points are supported; hollow points are preliminary. Lines stop at unknown capabilities. Exact values and sources follow below.

AgenticNo comparable evidenceUnknown

0 core families; 0 direct opponents across 0 labs. No admitted core comparisons

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 11/11 scenarios unsupported. Without one publisher: No supported estimate; 15/15 unsupported. Smoothing check: No supported estimate; 3/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts improved on neutral predictions within evaluated coverage; source sensitivity still requires inspection. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Hard reasoning2 families · 0 with independent evidence · Supported9.0

2 core families; 5 direct opponents across 4 labs.

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: 6.7–12.8; 2/7 scenarios unsupported. Without one publisher: 5.7–21.9; 1/16 unsupported. Smoothing check: 3.2–21.0; 0/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts improved on neutral predictions within evaluated coverage. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Aime: 25.0 observed win share
Microsoft · Source 1 Source 2
GPQA: 0.0 observed win share
Microsoft · Source 1

Coding3 families · 0 with independent evidence · Supported1.3

3 core families; 5 direct opponents across 4 labs.

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: 0.9–9.3; 2/7 scenarios unsupported. Without one publisher: 0.7–7.9; 1/20 unsupported. Smoothing check: 0.4–5.1; 0/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts improved on neutral predictions within evaluated coverage. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

LiveCodeBench: 0.0 observed win share
Microsoft · Source 1
SWE-bench Pro: 0.0 observed win share
Microsoft · Source 1
SWE-bench Verified: 0.0 observed win share
Microsoft · Source 1

Human prefNo comparable evidenceUnknown

0 core families; 0 direct opponents across 0 labs. No admitted core comparisons

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 1/1 scenarios unsupported. Without one publisher: No supported estimate; 1/1 unsupported. Smoothing check: No supported estimate; 3/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: cross-family and cross-publisher predictions could not be evaluated from the available comparison network. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Knowledge2 families · 0 with independent evidence · PreliminaryUnknown

2 core families; 1 direct opponents across 1 labs. No connected comparison to the complete reference panel

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 4/4 scenarios unsupported. Without one publisher: No supported estimate; 5/5 unsupported. Smoothing check: No supported estimate; 3/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: almost no publisher-held-out comparisons remained estimable. Cross-source validity is unresolved. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

MMLU-Pro: 0.0 observed win share
Microsoft · Source 1
Simpleqa Verified: 100.0 observed win share
Microsoft · Source 1

MultimodalNo comparable evidenceUnknown

0 core families; 0 direct opponents across 0 labs. No admitted core comparisons

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 8/8 scenarios unsupported. Without one publisher: No supported estimate; 8/8 unsupported. Smoothing check: No supported estimate; 3/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: limited publisher-held-out coverage. Cross-source validity is unresolved. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Long context2 families · 0 with independent evidence · PreliminaryUnknown

2 core families; 1 direct opponents across 1 labs. No connected comparison to the complete reference panel

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 5/5 scenarios unsupported. Without one publisher: No supported estimate; 5/5 unsupported. Smoothing check: No supported estimate; 3/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts underperformed neutral predictions. Treat this as an exploratory estimate. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Graphwalks: 0.0 observed win share
Microsoft · Source 1
Longbench: 0.0 observed win share
Microsoft · Source 1

Scores estimate outcomes against a shared reference panel; they are not accuracy percentages. Sparse or disconnected evidence cannot qualify an overall profile. Open a capability to inspect its evidence.

Compare available effort levels across 0 benchmark/harness combinations →

Reported effort · Not specified

Settings reported in this model's published benchmark results, including results outside the aggregate. Effort names are provider-specific. These are not API defaults or equal compute budgets.

Not specified: 19 observations

Mixed settings means multiple settings occur in the evidence. Best across efforts means the source selected its best reported result across settings; it does not mean Max. Unspecified settings stay unknown. This model-summary chart combines reported settings. The leaderboard keeps identified configurations separate and excludes unknown effort.

Inspect each result and its source ↓ · Download effort evidence

Compare capability profiles →

Score contributions and missing evidence

9 contributing families across 4 capabilities. Fixed reference panels do not change when the catalog expands.

Capability is fitted jointly across families. Capability estimates below describe different task areas; their weighted sum is not the Capability score.

Results without reviewed compatibility or a reference match remain in the raw evidence below. Coverage counts only contributing results.

Model information & shareable badge

Lab: Microsoft
Catalog status: preview
Availability: Microsoft Foundry public preview
Family: MAI
Released: —
Context: —
License: proprietary
Model card: https://microsoft.ai/models/mai-thinking-1/
Default Capability family coverage: /badge/mai-thinking-1.svg

Benchmark scores & sources

Original results, evaluation harnesses, and evidence behind this model.

Benchmark	Bucket	Score	Harness	Evidence	Source-recorded date
AdvancedIF rubric-level · source release snapshot; version not specified	Supporting evidence	85% Reported settings & source Microsoft own evaluation suite; Sonnet max reasoning; Tables 12/19. LongBenchV2 capped 256k,408 questions; CorpusQA GPT-5.4 high judge; 4 runs. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Tables 12 and 19 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
AIME 2025 · source release snapshot; version not specified	Hard reasoning	97% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
AIME 2026 · source release snapshot; version not specified	Hard reasoning	94.5% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
BFCL v3 · source release snapshot; version not specified	Supporting evidence	72% Reported settings & source Microsoft own evaluation suite; Sonnet max reasoning; Tables 12/19. LongBenchV2 capped 256k,408 questions; CorpusQA GPT-5.4 high judge; 4 runs. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Tables 12 and 19 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
CorpusQA · source release snapshot; version not specified	Supporting evidence	82% Reported settings & source Microsoft own evaluation suite; Sonnet max reasoning; Tables 12/19. LongBenchV2 capped 256k,408 questions; CorpusQA GPT-5.4 high judge; 4 runs. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Tables 12 and 19 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
GPQA Diamond · source release snapshot; version not specified	Hard reasoning	84.2% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
GraphWalks <=128k · source release snapshot; version not specified	Long context	90% Reported settings & source Microsoft own evaluation suite; Sonnet max reasoning; Tables 12/19. LongBenchV2 capped 256k,408 questions; CorpusQA GPT-5.4 high judge; 4 runs. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Tables 12 and 19 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
HealthBench Professional · source release snapshot; version not specified	Supporting evidence	35% Reported settings & source Microsoft own evaluation suite; Sonnet max reasoning; Tables 12/19. LongBenchV2 capped 256k,408 questions; CorpusQA GPT-5.4 high judge; 4 runs. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Tables 12 and 19 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
HMMT February 2026 · source release snapshot; version not specified	Supporting evidence	84.9% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
IF Bench · source release snapshot; version not specified	Supporting evidence	69% Reported settings & source Microsoft own evaluation suite; Sonnet max reasoning; Tables 12/19. LongBenchV2 capped 256k,408 questions; CorpusQA GPT-5.4 high judge; 4 runs. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Tables 12 and 19 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
LiveCodeBench v6 · source release snapshot; version not specified	Coding	87.7% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
LongBenchV2 · source release snapshot; version not specified	Long context	61% Reported settings & source Microsoft own evaluation suite; Sonnet max reasoning; Tables 12/19. LongBenchV2 capped 256k,408 questions; CorpusQA GPT-5.4 high judge; 4 runs. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Tables 12 and 19 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
MedXpertQA · source release snapshot; version not specified	Supporting evidence	43% Reported settings & source Microsoft own evaluation suite; Sonnet max reasoning; Tables 12/19. LongBenchV2 capped 256k,408 questions; CorpusQA GPT-5.4 high judge; 4 runs. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Tables 12 and 19 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
MMLU Pro · source release snapshot; version not specified	Knowledge	85% Reported settings & source Microsoft own evaluation suite; Sonnet max reasoning; Tables 12/19. LongBenchV2 capped 256k,408 questions; CorpusQA GPT-5.4 high judge; 4 runs. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Tables 12 and 19 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
MultiChallenge · source release snapshot; version not specified	Supporting evidence	53% Reported settings & source Microsoft own evaluation suite; Sonnet max reasoning; Tables 12/19. LongBenchV2 capped 256k,408 questions; CorpusQA GPT-5.4 high judge; 4 runs. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Tables 12 and 19 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
SimpleQA Verified · source release snapshot; version not specified	Knowledge	31% Reported settings & source Microsoft own evaluation suite; Sonnet max reasoning; Tables 12/19. LongBenchV2 capped 256k,408 questions; CorpusQA GPT-5.4 high judge; 4 runs. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Tables 12 and 19 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
SWE-bench Pro · source release snapshot; version not specified	Coding	52.8% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
SWE-bench Verified · source release snapshot; version not specified	Coding	73.5% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
Terminal-Bench 2.0 · 2.0	Coding	46% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Microsoft report Table 11 and section 4.1: MAI Terminal-Bench removes timeouts and uses a minimal ReAct harness; comparator values are cited from other model releases. A common evaluation protocol is not established.	Reviewed 2026-09-06
ARC-AGI-2	Hard reasoning	—	—	—	—
DeepSWE v1.1	Agentic	—	—	—	—
GDPval-AA	Agentic	—	—	—	—
GPQA Diamond	Hard reasoning	—	—	—	—
Humanity's Last Exam	Hard reasoning	—	—	—	—
LiveCodeBench	Coding	—	—	—	—
LMArena Text Arena	Human pref	—	—	—	—
MMLU-Pro	Knowledge	—	—	—	—
OSWorld-Verified	Agentic	—	—	—	—
SWE-bench Pro	Agentic	—	—	—	—
SWE-bench Verified	Agentic	—	—	—	—
Terminal-Bench 2.1	Agentic	—	—	—	—