MODEL CAPABILITIES / OPEN WEIGHTSOriginal Epoch AI chart ↗
FrontierMath Tier 4: open vs. closed models 62 evaluation results, colored by open or closed weights. Labels identify record-setting models over release date and leading open-weight models. Error bars are one standard error. FrontierMath Tier 4: open vs. closed Tier 4 (v2) accuracy by model release date 62 evaluations · 56 models Closed weights Open weights Overall frontier Open-weight frontier Accuracy 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Apr 2025 Jul Oct Jan 2026 Apr Jul Sep Model release date Gemini 2.5 Pro 2025-06-17 · 0.000% ± 0.000 pp (1 SE) Closed weights · Epoch: API access Claude Opus 4.1 (32k thinking) 2025-08-05 · 2.439% ± 2.439 pp (1 SE) Closed weights · Epoch: API access GPT-5 mini (high) 2025-08-07 · 12.195% ± 5.174 pp (1 SE) Closed weights · Epoch: API access GPT-5 nano (high) 2025-08-07 · 2.439% ± 2.439 pp (1 SE) Closed weights · Epoch: API access Claude Sonnet 4.5 (32k thinking) 2025-09-29 · 2.439% ± 2.439 pp (1 SE) Closed weights · Epoch: API access GPT-5 Pro (high) 2025-10-07 · 19.512% ± 6.266 pp (1 SE) Closed weights · Epoch: API access Claude Opus 4.5 (32k thinking) 2025-11-24 · 4.878% ± 3.406 pp (1 SE) Closed weights · Epoch: API access GPT-5.2 (xhigh) 2025-12-11 · 31.700% ± 7.300 pp (1 SE) Closed weights · Epoch: API access Gemini 3 Flash Preview 2025-12-17 · 17.073% ± 5.949 pp (1 SE) Closed weights · Epoch: API access Claude Opus 4.6 (max) 2026-02-05 · 26.829% ± 7.006 pp (1 SE) Closed weights · Epoch: API access Grok 4.20 2026-02-17 · 17.073% ± 5.949 pp (1 SE) Closed weights · Epoch: API access Gemini 3.1 Pro Preview 2026-02-19 · 26.829% ± 7.006 pp (1 SE) Closed weights · Epoch: API access GPT-5.4 (xhigh) 2026-03-05 · 49.000% ± 8.000 pp (1 SE) Closed weights · Epoch: API access GPT-5.4 nano (high) 2026-03-17 · 12.195% ± 5.174 pp (1 SE) Closed weights · Epoch: API access GPT-5.4 mini (xhigh) 2026-03-17 · 9.756% ± 4.692 pp (1 SE) Closed weights · Epoch: API access Claude Opus 4.7 (max) 2026-04-16 · 31.707% ± 7.358 pp (1 SE) Closed weights · Epoch: API access Grok 4.3 Beta 2026-04-17 · 14.634% ± 5.589 pp (1 SE) Closed weights · Epoch: Hosted access (no API) GPT-5.5 (xhigh) 2026-04-23 · 72.500% ± 7.150 pp (1 SE) Closed weights · Epoch: API access GPT-5.5 Instant 2026-05-05 · 2.439% ± 2.439 pp (1 SE) Closed weights · Epoch: API access AI co-mathematician 2026-05-08 · 75.600% ± 6.700 pp (1 SE) Closed weights · Epoch: Unreleased Qwen3.7 Max 2026-05-19 · 34.146% ± 7.498 pp (1 SE) Closed weights · Epoch: API access Gemini 3.5 Flash (high) 2026-05-19 · 26.829% ± 7.006 pp (1 SE) Closed weights · Epoch: API access Claude Opus 4.8 (max) 2026-05-28 · 56.098% ± 7.847 pp (1 SE) Closed weights · Epoch: API access Claude Sonnet 5 (max) 2026-06-30 · 29.268% ± 7.194 pp (1 SE) Closed weights · Epoch: API access Grok 4.5 (high) 2026-07-08 · 24.390% ± 6.790 pp (1 SE) Closed weights · Epoch: API access GPT-5.6 Sol (max) 2026-07-09 · 82.927% ± 5.949 pp (1 SE) Closed weights · Epoch: API access GPT-5.6 Sol (pro, max) 2026-07-09 · 80.488% ± 6.266 pp (1 SE) Closed weights · Epoch: API access GPT-5.6 Terra (max) 2026-07-09 · 70.732% ± 7.194 pp (1 SE) Closed weights · Epoch: API access GPT-5.6 Luna (max) 2026-07-09 · 60.976% ± 7.713 pp (1 SE) Closed weights · Epoch: API access Gemini 3.6 Flash (high) 2026-07-21 · 21.951% ± 6.545 pp (1 SE) Closed weights · Epoch: API access Gemini 3.5 Flash-Lite (high) 2026-07-21 · 0.000% ± 0.000 pp (1 SE) Closed weights · Epoch: API access Claude Opus 5 (max) 2026-07-24 · 73.171% ± 7.006 pp (1 SE) Closed weights · Epoch: API access Qwen3.8 Max (xhigh) 2026-08-02 · 46.341% ± 7.885 pp (1 SE) Closed weights · Epoch: API access Grok 4.6 (xhigh) 2026-08-12 · 31.707% ± 7.358 pp (1 SE) Closed weights · Epoch: API access Gemini 3.7 Flash (high) 2026-08-13 · 36.585% ± 7.616 pp (1 SE) Closed weights · Epoch: API access Claude Fable 5.1 (max) 2026-09-01 · 87.805% ± 5.174 pp (1 SE) Closed weights · Epoch: API access Qwen3.8 Max (0902) (xhigh) 2026-09-01 · 34.146% ± 7.498 pp (1 SE) Closed weights · Epoch: API access Gemini 3.8 Flash (high) 2026-09-02 · 21.951% ± 6.545 pp (1 SE) Closed weights · Epoch: API access GPT-6 Astra (medium) 2026-09-03 · 97.561% ± 2.439 pp (1 SE) Closed weights · Epoch: API access GPT-6 Astra (low) 2026-09-03 · 87.805% ± 5.174 pp (1 SE) Closed weights · Epoch: API access GPT-6 Astra (none) 2026-09-03 · 82.927% ± 5.949 pp (1 SE) Closed weights · Epoch: API access o3-mini (high) 2025-01-31 · 0.000% ± 0.000 pp (1 SE) Closed weights · Epoch: API access o4-mini (high) 2025-04-16 · 4.878% ± 3.406 pp (1 SE) Closed weights · Epoch: API access GPT-5 (high) 2025-08-07 · 21.951% ± 6.545 pp (1 SE) Closed weights · Epoch: API access GPT-5.2 Pro (xhigh) 2025-12-11 · 46.000% ± 7.790 pp (1 SE) Closed weights · Epoch: API access GPT-5.4 Pro (xhigh) 2026-03-05 · 58.537% ± 7.790 pp (1 SE) Closed weights · Epoch: API access GPT-5.5 Pro (xhigh) 2026-04-23 · 78.049% ± 6.545 pp (1 SE) Closed weights · Epoch: API access Claude Fable 5 (max) 2026-06-09 · 90.200% ± 4.600 pp (1 SE) Closed weights · Epoch: API access GPT-6 Astra (high) 2026-09-03 · 97.600% ± 2.400 pp (1 SE) Closed weights · Epoch: API access GPT-6 Astra (max) 2026-09-03 · 97.600% ± 2.400 pp (1 SE) Closed weights · Epoch: API access GPT-6 Astra (xhigh) 2026-09-03 · 97.600% ± 2.400 pp (1 SE) Closed weights · Epoch: API access DeepSeek v4 (max) 2026-04-24 · 2.439% ± 2.439 pp (1 SE) Open weights · Epoch: Open weights (unrestricted) Kimi K2.7 Code 2026-06-12 · 12.195% ± 5.174 pp (1 SE) Open weights · Epoch: Open weights (unrestricted) Inkling Small (xhigh) 2026-07-15 · 17.073% ± 5.949 pp (1 SE) Open weights · Epoch: Open weights (unrestricted) Inkling (xhigh) 2026-07-15 · 4.878% ± 3.406 pp (1 SE) Open weights · Epoch: Open weights (unrestricted) DeepSeek V4 Flash 0731 (max) 2026-07-31 · 24.390% ± 6.790 pp (1 SE) Open weights · Epoch: Open weights (unrestricted) GLM-5.3-Flash (max) 2026-08-20 · 17.073% ± 5.949 pp (1 SE) Open weights · Epoch: Open weights (unrestricted) Kimi K2.6 2026-04-20 · 25.641% ± 7.083 pp (1 SE) Open weights · Epoch: Open weights (unrestricted) GLM-5.2 (max) 2026-06-16 · 29.268% ± 7.194 pp (1 SE) Open weights · Epoch: Open weights (unrestricted) Kimi K3 (Max) 2026-07-16 · 39.024% ± 7.713 pp (1 SE) Open weights · Epoch: Open weights (non-commercial) DeepSeek V4 Pro 0813 (max) 2026-08-13 · 26.829% ± 7.006 pp (1 SE) Open weights · Epoch: Open weights (unrestricted) GLM-5.3 (max) 2026-08-14 · 29.268% ± 7.194 pp (1 SE) Open weights · Epoch: API access Publicly available weights under the GLM-5.3 license; supersedes Epoch's launch-time API-only classification. o3-mini (high) · 0.0% o4-mini (high) · 4.9% GPT-5 (high) · 22.0% GPT-5.2 Pro · 46.0% GPT-5.4 Pro (xhigh) · 58.5% GPT-5.5 Pro (xhigh) · 78.0% Claude Fable 5 (max) · 90.2% Kimi K2.6 · 25.6% GPT-6 Astra · 97.6% high / xhigh / max LEADING OPEN MODELS Kimi K3 (max) 39.0% GLM-5.2 (max) 29.3% GLM-5.3 (max) 29.3% DeepSeek V4 Pro 0813 (max) · 26.8% Open = publicly available weights, including license-restricted releases. Closed = weights not publicly available. Lines: best reported accuracy by release date. Error bars: ±1 standard error. Coincident evaluations overlap. Source: Epoch AI · CC BY · 10 Sep 2026 · GLM-5.3 access corrected using its official model repository Adapted visualization · Tier 4 (v2)

How to read this chart

All 62 evaluation configurations are retained: 11 open-weight and 51 closed-weight results, representing 56 distinct model names. Scores and release dates come directly from Epoch’s underlying CSV, not estimated pixels.

A frontier result has no earlier-or-same-date result with an equal or higher mean score, except exact ties. Labels cover the overall frontier, the open-weight frontier, and the top three distinct open-weight score levels, including ties. Same-date, equal-score Astra configurations share a label.

Frontiers use reported means, not statistical significance. Error bars show ±1 standard error, not 95% confidence intervals. Results sharing coordinates overlap; inspect every configuration in the table.

What “open” means here

Open means publicly available model weights, using Epoch’s accessibility field with documented corrections for current availability. This is not a claim that training data, code, or licenses satisfy an open-source definition. Models without publicly available weights are classified as closed-weight.

GLM-5.3 is corrected to open weights. Its official model repository provides weights under the GLM-5.3 license. Epoch’s stale launch-time “API access” value is preserved separately in the table and downloads. The score remains 29.3%, tied with GLM-5.2; the earlier model retains the frontier position.

Kimi K3 is included. Epoch classifies its weights as non-commercial, but its note describes commercial-use thresholds requiring a separate agreement. It is better understood here as an open-weight model with license restrictions.

Dates describe model releases, not when evaluations were run. Accessibility reflects this snapshot and its documented corrections rather than a historical license reconstruction. Benchmark version: Tier 4 (v2), private set only.

Extracted data

62 results

Model / evaluation settingRelease dateAccuracySE (pp)Weight accessOriginal Epoch accessibilityLabels
GPT-6 Astra (high)2026-09-0397.600%±2.400Closed weightsAPI accessOverall frontier
GPT-6 Astra (max)2026-09-0397.600%±2.400Closed weightsAPI accessOverall frontier
GPT-6 Astra (xhigh)2026-09-0397.600%±2.400Closed weightsAPI accessOverall frontier
GPT-6 Astra (medium)2026-09-0397.561%±2.439Closed weightsAPI access
Claude Fable 5 (max)2026-06-0990.200%±4.600Closed weightsAPI accessOverall frontier
Claude Fable 5.1 (max)2026-09-0187.805%±5.174Closed weightsAPI access
GPT-6 Astra (low)2026-09-0387.805%±5.174Closed weightsAPI access
GPT-5.6 Sol (max)2026-07-0982.927%±5.949Closed weightsAPI access
GPT-6 Astra (none)2026-09-0382.927%±5.949Closed weightsAPI access
GPT-5.6 Sol (pro, max)2026-07-0980.488%±6.266Closed weightsAPI access
GPT-5.5 Pro (xhigh)2026-04-2378.049%±6.545Closed weightsAPI accessOverall frontier
AI co-mathematician2026-05-0875.600%±6.700Closed weightsUnreleased
Claude Opus 5 (max)2026-07-2473.171%±7.006Closed weightsAPI access
GPT-5.5 (xhigh)2026-04-2372.500%±7.150Closed weightsAPI access
GPT-5.6 Terra (max)2026-07-0970.732%±7.194Closed weightsAPI access
GPT-5.6 Luna (max)2026-07-0960.976%±7.713Closed weightsAPI access
GPT-5.4 Pro (xhigh)2026-03-0558.537%±7.790Closed weightsAPI accessOverall frontier
Claude Opus 4.8 (max)2026-05-2856.098%±7.847Closed weightsAPI access
GPT-5.4 (xhigh)2026-03-0549.000%±8.000Closed weightsAPI access
Qwen3.8 Max (xhigh)2026-08-0246.341%±7.885Closed weightsAPI access
GPT-5.2 Pro (xhigh)2025-12-1146.000%±7.790Closed weightsAPI accessOverall frontier
Kimi K3 (Max)2026-07-1639.024%±7.713Open weightsOpen weights (non-commercial)Open frontier; Top 3 open score levels (ties included)
Gemini 3.7 Flash (high)2026-08-1336.585%±7.616Closed weightsAPI access
Qwen3.7 Max2026-05-1934.146%±7.498Closed weightsAPI access
Qwen3.8 Max (0902) (xhigh)2026-09-0134.146%±7.498Closed weightsAPI access
Claude Opus 4.7 (max)2026-04-1631.707%±7.358Closed weightsAPI access
Grok 4.6 (xhigh)2026-08-1231.707%±7.358Closed weightsAPI access
GPT-5.2 (xhigh)2025-12-1131.700%±7.300Closed weightsAPI access
GLM-5.2 (max)2026-06-1629.268%±7.194Open weightsOpen weights (unrestricted)Open frontier; Top 3 open score levels (ties included)
Claude Sonnet 5 (max)2026-06-3029.268%±7.194Closed weightsAPI access
GLM-5.3 (max)2026-08-1429.268%±7.194Open weights (corrected)API accessTop 3 open score levels (ties included)
Claude Opus 4.6 (max)2026-02-0526.829%±7.006Closed weightsAPI access
Gemini 3.1 Pro Preview2026-02-1926.829%±7.006Closed weightsAPI access
Gemini 3.5 Flash (high)2026-05-1926.829%±7.006Closed weightsAPI access
DeepSeek V4 Pro 0813 (max)2026-08-1326.829%±7.006Open weightsOpen weights (unrestricted)Top 3 open score levels (ties included)
Kimi K2.62026-04-2025.641%±7.083Open weightsOpen weights (unrestricted)Open frontier
Grok 4.5 (high)2026-07-0824.390%±6.790Closed weightsAPI access
DeepSeek V4 Flash 0731 (max)2026-07-3124.390%±6.790Open weightsOpen weights (unrestricted)
GPT-5 (high)2025-08-0721.951%±6.545Closed weightsAPI accessOverall frontier
Gemini 3.6 Flash (high)2026-07-2121.951%±6.545Closed weightsAPI access
Gemini 3.8 Flash (high)2026-09-0221.951%±6.545Closed weightsAPI access
GPT-5 Pro (high)2025-10-0719.512%±6.266Closed weightsAPI access
Gemini 3 Flash Preview2025-12-1717.073%±5.949Closed weightsAPI access
Grok 4.202026-02-1717.073%±5.949Closed weightsAPI access
Inkling Small (xhigh)2026-07-1517.073%±5.949Open weightsOpen weights (unrestricted)
GLM-5.3-Flash (max)2026-08-2017.073%±5.949Open weightsOpen weights (unrestricted)
Grok 4.3 Beta2026-04-1714.634%±5.589Closed weightsHosted access (no API)
GPT-5 mini (high)2025-08-0712.195%±5.174Closed weightsAPI access
GPT-5.4 nano (high)2026-03-1712.195%±5.174Closed weightsAPI access
Kimi K2.7 Code2026-06-1212.195%±5.174Open weightsOpen weights (unrestricted)
GPT-5.4 mini (xhigh)2026-03-179.756%±4.692Closed weightsAPI access
o4-mini (high)2025-04-164.878%±3.406Closed weightsAPI accessOverall frontier
Claude Opus 4.5 (32k thinking)2025-11-244.878%±3.406Closed weightsAPI access
Inkling (xhigh)2026-07-154.878%±3.406Open weightsOpen weights (unrestricted)
Claude Opus 4.1 (32k thinking)2025-08-052.439%±2.439Closed weightsAPI access
GPT-5 nano (high)2025-08-072.439%±2.439Closed weightsAPI access
Claude Sonnet 4.5 (32k thinking)2025-09-292.439%±2.439Closed weightsAPI access
DeepSeek v4 (max)2026-04-242.439%±2.439Open weightsOpen weights (unrestricted)
GPT-5.5 Instant2026-05-052.439%±2.439Closed weightsAPI access
o3-mini (high)2025-01-310.000%±0.000Closed weightsAPI accessOverall frontier
Gemini 2.5 Pro2025-06-170.000%±0.000Closed weightsAPI access
Gemini 3.5 Flash-Lite (high)2026-07-210.000%±0.000Closed weightsAPI access

Source: Epoch AI, CC BY as marked in the supplied SVG. Adapted visualization. Extracted 10 September 2026. Full raw precision is preserved in the downloadable CSV.