How the release frontier moved
ECI index points · model release date · 227 attributed releases since 2023
Loading chartCurrent ECI estimates by release date · not archived historical scores
Compare AI models
Benchmark results and the cost of the same token workload.
Prices & metadata · 2026-10-25Benchmarks · 2026-10-23ECI · 2026-10-24
Which models lead? Top 10 on GPQA Diamond · higher is better
Loading chartWhat do you pay for that performance? 1M input + 250k output tokens
Loading chart1M input + 250k output tokens across requests. Dashed frontier: best observed trade-offs. Current list-price arithmetic, not evaluation cost. Price axis uses log(1 + USD); score axis is zoomed. Select a point for its source and evaluation details.
Different routes to stronger models
Loading chartWhat does a matching score cost today
GPQA Diamond · 1M input + 250k output tokens · 21 priced evaluationsLoading chartCurrent list prices · blank means unavailable · no caching or batch discounts
Prices we can trace back to the source
Selected dated offerings · USD per million tokens · uncached input
Loading chartSelected offerings, not a continuous price index or equal-quality comparison. Claude 3 Haiku’s price was announced before availability. DeepSeek V3 uses the announced post-promotion tariff from 8 February 2025.
Look beyond a single score
Compare ability, cost and progress — each with its own measure.
How do the latest models compare?
Latest published release date among models with an approved GPQA Diamond result.
Loading chartOne model per developer within this reviewed sample. Other developers may lack compatible public evaluations; this is not a complete newest-model directory.
What makes up the bill?
Same newest reviewed selection · 1M input + 250k output tokens.
Loading chartCurrent list-price arithmetic, not historical release pricing or measured evaluation cost. No caching or batch discounts assumed.
Who moved ahead — and when?
GPQA Diamond · evaluation dates · best included mean score per developer.
Loading chartSteps change only at recorded evaluations. This sample is not a complete historical frontier; an individual model is not assumed to improve between points.
Where is each model strongest?
Leading ten on GPQA Diamond, plus the newest reviewed model per developer. Columns are separate tests.
Loading chartBlank cells mean no approved result. Read each column independently: identical percentages across different tests do not imply equal difficulty. No composite score is calculated. Prices captured 2026-10-25.
Epoch AI · CC BY 4.0 · GPQA Diamond, gpqa-diamond, max reasoning setting. Evaluation dates vary; listed workload cost is not evaluation cost.Full analysis ↗Methodology ↗
Prices captured 2026-10-25 · OpenRouter list prices and Epoch AI evaluations.
