Frontier Board

Every benchmark people actually watch.

The leaderboards behind the headlines: who runs them, what they test, who's on top right now, and what each model costs. Click any card to go to the source.

Scores refreshed 15 Oct 2026

Artificial Analysis previewIndex

Artificial Analysis

by Artificial Analysis

Intelligence Index built from 10 hard evals, plus price, speed and latency for every API model.

  1. 1Claude Opus 5.558
  2. 2Claude Sonnet 5.556
  3. 3Muse Wonder Full56
  4. Artificial Analysis Intelligence Index · top 3 of 54 current models
BridgeBench previewCoding

BridgeBench

by BridgeMind

Rates frontier coding models on 9 axes: reasoning, front end, back end, one-shot, security, trust, efficiency, speed and cost.

  1. 1Claude Opus 5.5760
  2. 2GPT-6 Astra728
  3. 3Claude Fable 5.1689
  4. BridgeBench (overall) · top 3 of 29 current models
Epoch AI Benchmarking Hub previewIndex

Epoch AI Benchmarking Hub

by Epoch AI

89 benchmarks and the Epoch Capabilities Index (ECI), with open data (CC BY 4.0).

  1. 1Claude Opus 5.5166.8
  2. 2GPT-6 Astra166.4
  3. 3GPT-6.1 Sol166.3
  4. Epoch Capabilities Index · top 3 of 81 current models
Arena (formerly LMArena) previewHuman votes

Arena (formerly LMArena)

by Arena

Blind head-to-head votes from real people. Elo for text, WebDev, vision and more.

Open the source for the live table.

Scale SEAL Leaderboards previewExpert evals

Scale SEAL Leaderboards

by Scale AI

20+ private, expert-built evals: agentic coding, frontier reasoning, safety and more.

Open the source for the live table.

Humanity's Last Exam previewReasoning

Humanity's Last Exam

by CAIS & Scale AI

2,500 expert-written questions at the edge of human knowledge.

  1. 1GPT-6 Astra54.8%
  2. 2Claude Fable 5.146.5%
  3. 3Gemini 3.1 Pro46.4%
  4. Humanity's Last Exam · top 3 of 13 current models
ARC Prize previewReasoning

ARC Prize

by ARC Prize Foundation

ARC-AGI puzzles that are easy for people and hard for AI, scored against cost per task.

  1. 1GPT-6 Astra95.0%
  2. 2GPT-6.1 Sol94.2%
  3. 3Claude Opus 5.593.3%
  4. ARC-AGI-2 · top 3 of 44 current models
Terminal-Bench previewAgents

Terminal-Bench

by Stanford & Laude Institute

Real tasks in a terminal: compiling, debugging, setting up servers, data work.

  1. 1GPT-5.584.7%
  2. 2GPT-5.481.8%
  3. 3Gemini 3.1 Pro80.2%
  4. Terminal-Bench · top 3 of 14 current models
CodingSWE-bench
Coding

SWE-bench

by Princeton & Stanford

Fix real GitHub issues in real Python repos. Verified is the human-checked subset.

  1. 1Claude Opus 4.783.5%
  2. 2Gemini 3.5 Flash79.3%
  3. 3Claude Opus 4.678.7%
  4. SWE-bench Verified · top 3 of 16 current models
OpenRouter Rankings previewUsage

OpenRouter Rankings

by OpenRouter

Which models developers actually pay for, ranked by tokens and spend.

Open the source for the live table.

Vellum LLM Leaderboard previewAggregator

Vellum LLM Leaderboard

by Vellum

Provider-reported and independent results on non-saturated benchmarks.

Open the source for the live table.

Aider Leaderboards previewCoding

Aider Leaderboards

by Aider

Polyglot code-editing across C++, Go, Java, JS, Python and Rust.

Open the source for the live table.

★Real use

The tests that are hard to game

Blind human votes, fresh tasks, private question sets and creators' own benchmarks. In B.O.B. these count double.

SimpleBench

AI Explained (YouTube)

Trick questions a person gets right and a pattern-matching model gets wrong.

  1. 1Claude Fable 581.9%
  2. 2Claude Opus 580.6%
  3. 3Gemini 3.1 Pro79.6%
  4. top 3 of 30 current models

BridgeBench (overall)

BridgeMind (YouTube)

A vibe-coding creator's own test suite, run on real app-building tasks.

  1. 1Claude Opus 5.5760
  2. 2GPT-6 Astra728
  3. 3Claude Fable 5.1689
  4. top 3 of 29 current models

LMArena Text

LMArena (blind human votes)

Real people compare two anonymous answers and pick the better one. Nothing to memorise.

  1. 1Gemini 4 Argon1525
  2. 2Muse Wonder1513
  3. 3Claude Opus 5.51508
  4. top 3 of 80 current models

LMArena Agent Arena

LMArena (real agent sessions)

Scored from real users' agent sessions: did the task get done, did it hallucinate tools.

  1. 1Claude Opus 5.50.170
  2. 2GPT-6.1 Sol0.150
  3. 3Claude Fable 5.10.143
  4. top 3 of 50 current models

LMArena WebDev

LMArena (blind human votes)

People ask for a real web app, compare two builds side by side and vote for the one that works better.

  1. 1Claude Opus 5.51811
  2. 2GPT-6 Astra1780
  3. 3Claude Sonnet 5.51774
  4. top 3 of 82 current models

LiveBench 2.0 (real-work suites)

LiveBench / Abacus.AI

Contamination-free real work: design, research, browser tasks, coding, long-running apps. Q&A suites left out.

  1. 1Claude Fable 5.188.8%
  2. 2Claude Opus 5.586.6%
  3. 3GPT-6 Astra85.5%
  4. top 3 of 39 current models

Vals Index

Vals AI (private test sets)

Independent evaluator with private test sets of paid work: finance, legal, tax, cyber.

  1. 1Gemini 4 Argon68.9%
  2. 2Claude Sonnet 5.567.0%
  3. 3Claude Opus 5.567.0%
  4. top 3 of 42 current models

SWE-rebench

Nebius

Fresh GitHub issues collected after each model's training cutoff.

  1. 1Claude Fable 564.5%
  2. 2Grok 4.563.8%
  3. 3Claude Opus 563.4%
  4. top 3 of 34 current models

Kagi LLM Benchmark

Kagi

Private question set that is never published, so it can't leak into training.

  1. 1Claude Fable 591.4%
  2. 2Claude Opus 4.888.8%
  3. 3GPT-5.588.8%
  4. top 3 of 22 current models

EQ-Bench Creative Writing

Sam Paech (open source)

Independent, open-source writing test that also penalises AI slop and repetition.

  1. 1GPT-6 Astra2173
  2. 2Claude Fable 5.12162
  3. 3Claude Opus 52133
  4. top 3 of 46 current models

SkateBench

Theo (t3.gg, YouTube)

A creator's niche knowledge test: 390 trick definitions, no lab optimises for it.

  1. 1Gemini 3.1 Pro97.0%
  2. 2GPT-5.482.0%
  3. 3GPT-5.4 Pro79.0%
  4. top 3 of 8 current models
Cost

What the top models cost

List prices per 1M tokens on OpenRouter for the 20 highest-ranked models, plus what Artificial Analysis spent per task to run its Intelligence Index. A model that thinks longer costs more per task, even at the same price.

Price per 1M tokensInputOutput
Cost per task, AA Intelligence Index
  • MiMo-V2.6-Pro$0.13
  • Claude Haiku 5.5$0.21
  • GPT-6.1 Sol$0.72
  • Step 5 Preview$1.03
  • Gemini 3.8 Flash$1.24
  • GPT-5.6 Terra$1.40
  • Muse Spark 1.3$1.60
  • Gemini 4 Argon$1.99
  • Kimi K3$2.00
  • GLM-5.3$2.01
  • GPT-6 Astra$3.26
  • Grok 4.7$3.74
  • Qwen3.8 Max$5.41
  • Claude Sonnet 5.5$5.46
  • Claude Opus 5.5$5.98
  • Claude Fable 5.1$7.63

Source: Artificial Analysis, best setting per model.

Most score per dollar

Models with B.O.B. 60+, ranked by score ÷ blended price

  1. 1Claude Haiku 5.563 @ $0.20
  2. 2MiMo-V2.6-Pro76 @ $0.54
  3. 3Step 5 Preview62 @ $1.43
  4. 4Gemini 3.8 Flash64 @ $1.50
  5. 5Gemini 3.7 Flash62 @ $1.50
  6. 6Muse Spark 1.371 @ $2.00

Prices from OpenRouter list prices. Labs' direct prices, caching and batch discounts can be lower.