Benchmark Almanac

How fast is the frontier moving?

The best capabilities index score rose about 20 points a year over the last two years, and the index set nine new records in 2026 so far, against six in 2024. Open models trail the best closed model by about seven months.

How far behind are open models?

The best open model today, Kimi K3.1 Pro, scores what the best closed model scored seven months ago. Since 2024 the median gap has been six months. We take the best score on Epoch AI’s Capabilities Index of any model released by the end of each month.

100 120 140 160 180 2023 2024 2025 2026 Capabilities index Closed weights Open weights 100 120 140 160 180 2023 2024 2025 2026 Capabilities index Closed weights Open weights

Epoch Capabilities Index, Epoch AI (CC BY 4.0)

Months behind

How long ago a closed model first reached the open models’ best, each month

0 5 10 15 2023 2024 2025 2026 0.6 7 0 5 10 15 2023 2024 2025 2026 0.6 7

Epoch Capabilities Index, Epoch AI (CC BY 4.0)

How long does Arena's number one last?

Seven models have held Arena's number one in 2026 so far, for a median of 28 days each. In 2025, nine held it for a median of 40 days. Gemini 4 Carbon is on top now. We read Arena's text leaderboard with style control from its daily snapshots since August 28th, 2024.

o1-preview o1 o3 Gemini 3 Pro Claude Fable 5 2025 2026 o1 o3 2025 2026

Source: Arena’s leaderboard dataset on Hugging Face (CC BY 4.0)

How long does the index's number one last?

Nine models have taken the top spot on Epoch AI’s Capabilities Index in 2026 so far, for a median of 39 days each. In 2025, seven held it for a median of 55 days, and in 2023, three for 115. GPT-6.2 Astra is on top now.

GPT-4 GPT-4 Turbo o1-mini o1 o3 GPT-5 2024 2025 2026 GPT-4 o1 2024 2025 2026

Epoch Capabilities Index, Epoch AI (CC BY 4.0)

Who held number one?

Anthropic has held number one on the most benchmark-days in every half year since the first half of 2026, 42% of them in the second half of 2026 so far.

Leaders by category
  • OpenAI
  • Anthropic
  • Google DeepMind
  • Qwen
  • Meta
  • Other labs

Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards

What does last year's best score cost now?

On the median board that prices each task, matching the best score from a year ago is 2.1 times cheaper now than it was then. It's at least ten times cheaper on 13 of the 47 boards, and 51 times cheaper on Vibe Code Bench.

  1. DTBench 11×
  2. LMCA 47×
  3. BullshitBench V2 9.5×
  4. TaxEval v2 9.1×
  5. Vibe Code Bench 51×
  6. CorpFin v2 10×
  7. WeirdML 2.1×
  8. BullshitBench 1×
  9. ARC-AGI-2 39×
  10. ARC-AGI 42×
  11. MedScribe 43×
  12. MedCode 2.3×
  13. MortgageTax 1.8×
  14. AIME 2024 1.5×
  15. MedQA Bias Evaluation 1×
  16. SAGE (Vals) 1.8×
  17. MGSM 2.0×
  18. ZeroBench 13×
  19. IOI (Vals) 1.3×
  20. ProofBench 41×
  21. CaseLaw v2 1.1×
  22. Pencil Puzzle Bench 1.1×
  23. AIME 2025 16×
  24. HMMT 2025 17×
  25. Vals Finance Agent 1.6×
  26. Vals Finance Agent v1 1×
  27. MathArena (Apex, IMO, USAMO) 5.6×
  28. BRUMO 2025 21×
  29. SMT 2025 1×
  30. SWE-bench Verified 6.3×
  31. MathArena Apex Shortlist 9.6×
  32. Vals Index v1 1.5×
  33. CMIMC 2025 19×
  34. LiveSQLBench 1.3×
  35. MCPMark 1.1×
  36. Vals Multimodal Index v1 4.7×
  37. ARCS 1×
  38. LevelBench-100 v4.1 1×
  39. Math Kangaroo 2025 (MathArena) 3.8×
  40. HMMT Nov 2025 1×
  41. SkateBench v2 2.8×
  42. LLM Physics Plotting Benchmark 1.3×
  43. Vibe Code Bench v1 1×
  44. BrowseSafe++ 1.5×
  45. AstaBench 1×
  46. VCBench 8.8×
  47. Agent-Diff 1.3×

Source: Official leaderboards

How this is counted

The capabilities index is Epoch AI’s. It puts models on one scale from many benchmarks. A line holds the best score of any model released by each month’s end, among the models whose weights we know are published or not. A model counts from its release date, so a model scored later still counts from the month it came out.

The months behind are counted back from each month to the day a closed model first scored as high as the best open model had by then. When the best open model improves, the gap shrinks. When a closed model sets a new best, the gap grows a month at a time.

The index's number one is the model with the best score among the models released by each day. A model takes the top spot on its release date, even if Epoch AI scored it later, and holds it until a model released after it scores higher.

Number one counts the same benchmarks as the Leaders page, from the source each benchmark’s page leads with. Contested time only. A leader’s clock runs for 45 days after it takes the top spot and 45 more after each score from another lab. It stops once the best score passes 85%, or at the last score once 45 days pass without another. Labs that never held a tenth of a half year are grouped as other labs.

The cost comparison uses the 47 leaderboards that list a cost per task for runs of at least five models released by November 21st, 2025 and three released since. On each one we take the best score among models out by then, the cheapest run that reached it then, and the cheapest run that reaches it now, from any model at any setting. A run counts from its model's release date, at the cost the board lists. Costs compare only within a board, since boards price runs their own way, so each row shows how many times cheaper the score got rather than one price.

More insights