RankingEffort benchmarks

Step 5 Preview · Reasoning

Compare models

19 measured results · 2 effort settings · 19 benchmark/harness combinations

Model evidence and sources →

Change model

Model

Current configurationSources

Compare with

Choose a second model and its effort to see both profiles and benchmark differences below.

Compare withModel and configuration

Capability profile

One configuration, seven capabilities

Comparison score · 0–100

AStep 5 Preview · Reasoning2 of 7 capabilities supported

Gaps are unknown, not zero. Hollow points are preliminary.

Agentic
A53.8Supported
Hard reasoning
A54.9Supported
Coding
A74.1Preliminary
Human pref
Unknown
Knowledge
A34.8Preliminary
Multimodal
A27.6Preliminary
Long context
A60.0Preliminary

Each profile belongs to the named configuration. Capability scores use a shared panel for all axes, distinct from the overall ranking panel; they are not accuracy percentages or direct head-to-head wins.

Capability score evidence

Step 5 Preview · Reasoning: 44.3 overall comparison score. Meets the overall evidence requirements.

6 areas measured. The available task mix affects this estimate; missing areas remain unknown.

Which sources carry the weight?

Artificial Analysis · 100.0% direct comparison weight · independent of the model provider

Which benchmark families carry the weight?

Humanity’s Last Exam · 8.5% direct comparison weight
Critpt · 8.4% direct comparison weight
omniscience · 8.4% direct comparison weight
Mmmu Pro · 8.4% direct comparison weight
GDPval · 7.7% direct comparison weight
Other families (10) · 58.6% direct comparison weight

These shares describe direct comparison weight, not fractions of the final score or confidence. Source balancing does not give every family equal weight or remove correlations between overlapping sources.

How sensitive is this score?

The audit removes one source at a time and refits against the same overall reference panel. These published results use the matching dataset audited 2026-10-20.

Both ranks use the same surviving configurations for each removal. The cohort starts with 444 configurations supported in the original audit, including separate effort settings. These are diagnostic positions, not the default leaderboard ranks. Eligibility is not rechecked after removal.

What changes when a source is removed?
Removed source	Rank before → after	Rank movement	Cohort remaining / lost support	Score after (change)
Artificial Analysis	Comparison support lost; no rank		203 remaining / 241 lost	Unavailable
ARC Prize	79 → 73	Up 6	444 remaining / 0 lost	50.2 (+5.9 points)
Datacurve	79 → 78	Up 1	444 remaining / 0 lost	39.9 (-4.4 points)
LiveCodeBench	79 → 79	No movement	444 remaining / 0 lost	44.2 (-0.1 points)
Scale	79 → 78	Up 1	444 remaining / 0 lost	46.4 (+2.1 points)
Terminal-Bench	79 → 76	Up 3	444 remaining / 0 lost	44.4 (+0.1 points)
Cognition	79 → 80	Down 1	444 remaining / 0 lost	42.8 (-1.6 points)
LiveBench	79 → 78	Up 1	443 remaining / 1 lost	45.2 (+0.8 points)
SWE-rebench	79 → 76	Up 3	444 remaining / 0 lost	44.6 (+0.2 points)
Epoch AI	78 → 79	Down 1	431 remaining / 13 lost	44.9 (+0.5 points)
Vals AI	79 → 78	Up 1	443 remaining / 1 lost	43.5 (-0.8 points)
Provider reports · Anthropic	79 → 79	No movement	444 remaining / 0 lost	44.3 (-0.0 points)
Provider reports · Moonshot AI	79 → 80	Down 1	444 remaining / 0 lost	43.7 (-0.6 points)
Provider reports · Meta	79 → 79	No movement	444 remaining / 0 lost	44.3 (-0.1 points)
Provider reports · Mistral	79 → 79	No movement	444 remaining / 0 lost	44.3 (+0.0 points)
Provider reports · OpenAI	79 → 78	Up 1	444 remaining / 0 lost	44.5 (+0.2 points)
All provider reports	79 → 80	Down 1	444 remaining / 0 lost	43.8 (-0.5 points)

Lost support means a configuration no longer has a comparison path to the complete reference panel. Its rank is unavailable. A surviving path alone does not establish leaderboard eligibility.

Overlapping sources may remain correlated even with source balancing. Leave-one-source-out movement is a sensitivity diagnostic, not a spread of source-specific ranks or a confidence measure. Scoring methodology · Dated configuration audit

Every benchmark family

Removed evidence	Recomputed score	Change
arc-agi	49.1	+4.8
deepswe	39.9	-4.4
mmmu-pro	47.9	+3.5
terminal-bench-science	47.6	+3.3
itbench	41.9	-2.4
omniscience	46.2	+1.8
gdp-pdf	46.1	+1.7
frontiercode	42.8	-1.6
terminal-bench	43.3	-1.0
apex-agents	43.3	-1.0
enterprise-ops	43.4	-1.0
gdpval	43.4	-0.9
simpleqa	45.2	+0.9
briefcase	43.5	-0.9
epoch-game-puzzles	43.5	-0.9
ifbench	43.7	-0.7
analyst-agent	45.0	+0.7
aa-lcr	43.7	-0.6
gpqa	43.8	-0.6
hle	44.9	+0.5
mmlu-pro	44.8	+0.4
multi-swe-bench	43.9	-0.4
livebench-instructions	44.7	+0.4
livebench-language	44.6	+0.3
arc-agi-3	44.0	-0.3
harvey	44.0	-0.3
tau-bench	44.0	-0.3
swe-atlas-qna	44.1	-0.3
critpt	44.1	-0.2
swe-rebench	44.6	+0.2
automationbench	44.5	+0.2
livebench-coding	44.1	-0.2
enigma-eval	44.2	-0.2
vibe-code	44.2	-0.1
swe-atlas-test-writing	44.2	-0.1
livecodebench	44.2	-0.1
aime	44.2	-0.1
livebench-reasoning	44.2	-0.1
vals-legal-research	44.2	-0.1
vals-code-migration	44.3	-0.1
vals-tax-agent	44.3	-0.1
vals-finance-agent	44.3	-0.0
swe-atlas-refactoring	44.3	-0.0
gmmlu	44.3	+0.0
vals-excel-modeling	44.3	-0.0
frontiermath	44.3	+0.0
livebench-data	44.3	+0.0
swe-bench-pro	44.3	+0.0
math500	44.3	+0.0
livebench-math	44.3	-0.0

What evidence is missing from the fit?

Across all collected settings for this model: 24 observations, 16 contributing, 0 matched without graph weight and 8 excluded. Counts do not establish rank eligibility.

1 · Benchmark family or source protocol has not been reviewed
1 · Original SciCode grading defects are under review; retained as supporting evidence, excluded from aggregate and capability scores. Corrected SciCode-Verified requires a separate protocol review. Sources: https://arxiv.org/abs/2608.04975 and https://artificialanalysis.ai/evaluations/scicode
1 · Effort is not identified
3 · Supporting evidence outside the reviewed capability core
2 · No reviewed effort for this catalog observation; any separately collected effort measurement is counted separately

Inspect every observation and exclusion →

Benchmark results

These are published benchmark scores. Supported effort estimates also appear on the Capability leaderboard. Results from different harnesses occupy different rows. Missing cells stay unknown; no scores are copied between effort levels.

Benchmark / source	Reasoning	Unspecified
AA-AnalystAgentArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Share of tasks solved on all 5 of 5 attempts · 80 tasks, 5 attempts per task · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	35% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
AA-Briefcase EloArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details AA-Briefcase v1.1 is an agentic knowledge work benchmark developed by Artificial Analysis. AA-Briefcase Elo is a combined metric that aggregates rubric pass rate, analytical quality Elo and presentation Elo · Higher is better · Evaluation results measured independently by Artificial Analysis Unit: elo. Source collected 2026-10-18.	1,424.32 Elo Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
APEX-Agents-AAArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	37.98% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
AA-LCR v1.1Artificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	88.33% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
AutomationBench-AAArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Share of task objectives completed with no guardrail violations · Higher is better · Benchmark developed by Zapier · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	50.96% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
CritPtArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Benchmark developed by Argonne and UIUC, with contributions from 60+ researchers globally. CritPt is under review by its authors following external feedback · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	20.86% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
EnterpriseOps-Gym-AAArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Overall task success rate across all eight enterprise domains (oracle tool mode) · Benchmark developed by ServiceNow Research · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	47.18% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
GDP.pdf: All-pass RateArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Share of attempts where every atomic criterion passed · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	14.8% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
GDPval-AA v2.1 LeaderboardArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Elo rating for performance on real-world work tasks · Anchored to DeepSeek V4.1 Flash (max) at 1600 · Higher is better · Evaluation results measured independently by Artificial Analysis Unit: elo. Source collected 2026-10-18.	1,586.63 Elo Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
Humanity's Last ExamArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	46.48% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
ITBench-AAArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Combined public and private Site Reliability Engineering (SRE) tasks · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	55.56% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
MLCR-AAArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Share of tasks judged accurate, complete, and concise · Expert + compound (hard) sets · Benchmark developed by Wisedocs · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	16.67% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
MMMU-ProArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	76.36% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
AA-Omniscience IndexArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details AA-Omniscience Index (higher is better) measures knowledge reliability and hallucination. It rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer. Scores range from -100 to 100, where 0 means as many correct as incorrect answers, and negative scores mean more incorrect than correct. · Evaluation results measured independently by Artificial Analysis Unit: index. Source collected 2026-10-18.	16.38 Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
AA-Omniscience AccuracyArtificial Analysis · 6000 questions; 1 attempt; correct-answer share; GPT-5.6 Luna medium grader Benchmark details AA-Omniscience Accuracy (higher is better) measures the proportion of correctly answered questions out of all questions, regardless of whether the model chooses to answer · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	41.53% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
SciCodeArtificial Analysis · Artificial Analysis published evaluation; see source for benchmark-specific settings Benchmark details Benchmark curated by scientists across 16 disciplines. SciCode is under review following an independent audit of the dataset · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	58.91% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
Terminal-Bench-Science 0.1Artificial Analysis · mini-swe-agent; Terminal-Bench-Science 0.1.0; 70 tasks; pass@1 averaged over 3 repeats Benchmark details Share of tasks passed (pass@1), averaged over three repeats · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	2.38% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
Terminal-Bench v4.0Artificial Analysis · mini-SWE-agent v2.4.6; 66 tasks; pass@1 averaged over 3 repeats; 500 steps; task timeout up to 8 hours Benchmark details Benchmark developed by the Laude Institute, Stanford, and open-source contributors · Independently benchmarked by Artificial Analysis · Evaluation results measured independently by Artificial Analysis Unit: percent. Source collected 2026-10-18.	33.33% Reported settings Step 5 Preview; Reported effort: reasoning; Model source: https://artificialanalysis.ai/models/step-5; Mode from evaluator isReasoning metadata; effort budget not specified	Not reported
Artificial Analysis Cyber IndexArtificial Analysis · Artificial Analysis Stirrup; defensive code audit, patch and end-to-end vulnerability tasks Benchmark details v1.0. Equal mean of CWE-Bench-AA, DeepsecBench-AA and CyberGym-E2E-AA. Display only. This composite measures defensive source-code tasks. Successful capability scores depend on safety refusals, and refusal rates are reported separately. It does not enter weighted Overall. https://artificialanalysis.ai/methodology/cyber-index Unit: index. Source collected 2026-10-18.	Not reported	47.71 Reported settings Artificial Analysis Stirrup; v1.0; Step 5 Preview; source label Step 5 Preview; mitigation standard; public model; source field cyberIndex; index 0–100; CWE-Bench-AA pass@1 0.525; refusal 0; DeepsecBench-AA median F2, 3 repeats, GPT-5.6 Sol High judge 0.158112846668792; refusal 0; CyberGym-E2E-AA pass@1 0.7480916030534351; refusal 0; https://artificialanalysis.ai/models/step-5; page sha256 344ca66f5aa3f3abd1901b12bd29db73950d90e553fcb3ae900dcf590d7a6db2; encrypted payload sha256 b5e6ae259181eaed471e288c4bbdbd1088438e0a81caf19ec0fa60f7092d48f7; Display only. This composite measures defensive source-code tasks. Successful capability scores depend on safety refusals, and refusal rates are reported separately. It does not enter weighted Overall.

What these comparisons mean

Each source reports its own effort labels. Matching labels do not establish equal compute budgets or identical fallback behavior. Open Reported settings to inspect the exact model name, agent and task coverage. Multiple results at one level remain visible rather than selecting the best. Source-reported None or Non-reasoning labels are not treated as evidence that an API supports those settings.

These measurements use their own dated source collection. The leaderboard fits identified configurations jointly; it contains no pooled model ratings. Explicit thinking budgets are separate configurations. Read the methodology.

Download all effort benchmark results · 8,915 results across 259 models · Collected 2026-10-18

Step 5 Preview · Reasoning

Compare with

One configuration, seven capabilities

Agentic

Hard reasoning

Coding

Human pref

Knowledge

Multimodal

Long context

Capability score evidence

Benchmark results