Frontier AI · August 2026 snapshot

Every frontier model, measured on what matters

Live rankings across reasoning, coding and agentic benchmarks — with real API pricing and context windows. Filter to your workload, sort by the metric you care about, and put up to four models head-to-head.

38
models tracked: 13
labs: 11
benchmarks: 16
open-weight

Open the board Compare models

Sort

Showing 38 of 38 models · sorted by Overall composite (high → low, no data sinks last)

N/A

Kimi K3Recent

MoonshotJul 22, 2026

Open

SWE-V: 93.4/100
GPQA: —
OSWorld: —

DeepSeek-V4-Pro (0813)New

DeepSeekAug 13, 2026

Open

SWE-V: 80.6/100
GPQA: 90.1/100
OSWorld: —

N/A

Kimi K2.6

MoonshotApr 8, 2026

Open

SWE-V: 80.2/100
GPQA: —
OSWorld: —

N/A

Claude Fable 5Recent

AnthropicJun 9, 2026

SWE-Pro: 80/100

N/A

DeepSeek-V4 Flash

DeepSeekApr 24, 2026

Open

SWE-V: 79/100
GPQA: —
OSWorld: —

GPT-5.6 SolRecent

OpenAIJul 9, 2026

SWE-V: —
GPQA: 93.6/100
OSWorld: 62.6/100

N/A

Gemini 3 Flash

GoogleDec 17, 2025

SWE-V: 78/100
GPQA: —
OSWorld: —

N/A

Gemini 3.1 Pro

GoogleFeb 19, 2026

ARC-AGI-2: 77.1/100

N/A

Gemini 3 Pro

GoogleNov 18, 2025

SWE-V: 76.2/100
GPQA: —
OSWorld: —

N/A

Grok 4.5Recent

xAIJul 16, 2026

SWE-Pro: 64.7/100
Term-Bench: 83.3/100

N/A

Claude Haiku 4.5

AnthropicOct 15, 2025

SWE-V: 73.3/100
GPQA: —
OSWorld: —

Kimi K2 Thinking

MoonshotNov 6, 2025

Open

SWE-V: 71.3/100
GPQA: —
OSWorld: —

N/A

GLM-5.2Recent

Z.aiJun 13, 2026

Open

SWE-Pro: 62.1/100

N/A

Grok 4.6New

xAIAug 14, 2026

SWE-V: —
GPQA: —
OSWorld: —

N/A

GLM-5.3New

Z.aiAug 14, 2026

OpenSub first

SWE-V: —
GPQA: —
OSWorld: —

N/A

Muse Glimmer 30BNew

MetaAug 10, 2026

Open

SWE-V: —
GPQA: —
OSWorld: —

N/A

Qwen3.8 MaxNew

QwenAug 5, 2026

Open

SWE-V: —
GPQA: —
OSWorld: —

N/A

Claude Opus 5Recent

AnthropicJul 24, 2026

SWE-V: —
GPQA: —
OSWorld: —

N/A

GPT-5.6 TerraRecent

OpenAIJul 9, 2026

SWE-V: —
GPQA: —
OSWorld: —

N/A

GPT-5.6 LunaRecent

OpenAIJul 9, 2026

SWE-V: —
GPQA: —
OSWorld: —

N/A

Muse Spark 1.1Recent

MetaJul 9, 2026

SWE-V: —
GPQA: —
OSWorld: —

N/A

Claude Sonnet 5Recent

AnthropicJun 30, 2026

SWE-V: —
GPQA: —
OSWorld: —

N/A

MAI-Thinking-1Recent

MicrosoftJun 2, 2026

SWE-V: —
GPQA: —
OSWorld: —

N/A

Claude Opus 4.8Recent

AnthropicMay 28, 2026

SWE-V: —
GPQA: —
OSWorld: —

N/A

GPT-5.5 Pro

OpenAIApr 24, 2026

SWE-V: —
GPQA: —
OSWorld: —

N/A

GPT-5.5

OpenAIApr 24, 2026

SWE-V: —
GPQA: —
OSWorld: —

N/A

GLM-5

Z.aiFeb 11, 2026

Open

SWE-V: —
GPQA: —
OSWorld: —

N/A

Nova 2 Pro

AmazonDec 2, 2025

Preview

SWE-V: —
GPQA: —
OSWorld: —

N/A

Nova 2 Lite

AmazonDec 2, 2025

SWE-V: —
GPQA: —
OSWorld: —

N/A

Nova 2 Omni

AmazonDec 2, 2025

Preview

SWE-V: —
GPQA: —
OSWorld: —

N/A

Grok 4 Fast Reasoning

xAIDec 1, 2025

Open

SWE-V: —
GPQA: —
OSWorld: —

N/A

MiniMax-M2

MiniMaxOct 27, 2025

Open

SWE-V: —
GPQA: —
OSWorld: —

N/A

Gemini 3 Deep Think

GoogleAug 1, 2025

Sub first

SWE-V: —
GPQA: —
OSWorld: —

N/A

Mistral Medium 3.1

MistralAug 1, 2025

SWE-V: —
GPQA: —
OSWorld: —

N/A

Qwen3 235B Instruct 2507

QwenJul 24, 2025

Open

SWE-V: —
GPQA: —
OSWorld: —

N/A

Qwen3 Coder Flash

QwenJul 23, 2025

Open

SWE-V: —
GPQA: —
OSWorld: —

N/A

Devstral Medium 2507

MistralJul 10, 2025

Open

SWE-V: —
GPQA: —
OSWorld: —

N/A

Llama 4 Maverick

MetaApr 5, 2025

Open

SWE-V: —
GPQA: —
OSWorld: —

Model	Overall	SWE-V	GPQA	OSWorld	$ & ctx	Released
N/A Kimi K3Moonshot	93.4	93.4	—	—	$61M	Jul 22, 2026
#1 DeepSeek-V4-Pro (0813)DeepSeek · new	87.9	80.6	90.1	—	$1.401M	Aug 13, 2026
N/A Kimi K2.6Moonshot	80.2	80.2	—	—	$1.07256K	Apr 8, 2026
N/A Claude Fable 5Anthropic	80	—	—	—	$161M	Jun 9, 2026
N/A DeepSeek-V4 FlashDeepSeek	79	79	—	—	$0.81512K	Apr 24, 2026
#2 GPT-5.6 SolOpenAI	78.3	—	93.6	62.6	$11.251.1M	Jul 9, 2026
N/A Gemini 3 FlashGoogle	78	78	—	—	$1.131M	Dec 17, 2025
N/A Gemini 3.1 ProGoogle	77.1	—	—	—	$4.501M	Feb 19, 2026
N/A Gemini 3 ProGoogle	76.2	76.2	—	—	$4.501M	Nov 18, 2025
N/A Grok 4.5xAI	74	—	—	—	$3500K	Jul 16, 2026
N/A Claude Haiku 4.5Anthropic	73.3	73.3	—	—	$2200K	Oct 15, 2025
#3 Kimi K2 ThinkingMoonshot	70.2	71.3	—	—	$1.07256K	Nov 6, 2025
N/A GLM-5.2Z.ai	62.1	—	—	—	$2.151M	Jun 13, 2026
N/A Grok 4.6xAI · new	—	—	—	—	$3500K	Aug 14, 2026
N/A GLM-5.3Z.ai · new	—	—	—	—	—1M	Aug 14, 2026
N/A Muse Glimmer 30BMeta · new	—	—	—	—	—256K	Aug 10, 2026
N/A Qwen3.8 MaxQwen · new	—	—	—	—	$31M	Aug 5, 2026
N/A Claude Opus 5Anthropic	—	—	—	—	$10200K	Jul 24, 2026
N/A GPT-5.6 TerraOpenAI	—	—	—	—	$4.501.1M	Jul 9, 2026
N/A GPT-5.6 LunaOpenAI	—	—	—	—	$0.451.1M	Jul 9, 2026
N/A Muse Spark 1.1Meta	—	—	—	—	—500K	Jul 9, 2026
N/A Claude Sonnet 5Anthropic	—	—	—	—	$61M	Jun 30, 2026
N/A MAI-Thinking-1Microsoft	—	—	—	—	—256K	Jun 2, 2026
N/A Claude Opus 4.8Anthropic	—	—	—	—	$10200K	May 28, 2026
N/A GPT-5.5 ProOpenAI	—	—	—	—	—400K	Apr 24, 2026
N/A GPT-5.5OpenAI	—	—	—	—	$3.44400K	Apr 24, 2026
N/A GLM-5Z.ai	—	—	—	—	$1.55200K	Feb 11, 2026
N/A Nova 2 ProAmazon	—	—	—	—	—300K	Dec 2, 2025
N/A Nova 2 LiteAmazon	—	—	—	—	—250K	Dec 2, 2025
N/A Nova 2 OmniAmazon	—	—	—	—	—300K	Dec 2, 2025
N/A Grok 4 Fast ReasoningxAI	—	—	—	—	$0.352M	Dec 1, 2025
N/A MiniMax-M2MiniMax	—	—	—	—	$0.52205K	Oct 27, 2025
N/A Gemini 3 Deep ThinkGoogle	—	—	—	—	—1M	Aug 1, 2025
N/A Mistral Medium 3.1Mistral	—	—	—	—	$0.80128K	Aug 1, 2025
N/A Qwen3 235B Instruct 2507Qwen	—	—	—	—	$0.39262K	Jul 24, 2025
N/A Qwen3 Coder FlashQwen	—	—	—	—	$0.52262K	Jul 23, 2025
N/A Devstral Medium 2507Mistral	—	—	—	—	$0.80128K	Jul 10, 2025
N/A Llama 4 MaverickMeta	—	—	—	—	$0.421M	Apr 5, 2025

Labs covered

Official naming and lineups as published by each lab — from GPT‑5.6's Sol/Terra/Luna tiers to Claude's new Fable line and Meta's Muse family.

What the benchmarks actually measure

No single number tells you which model fits your workload. Score a coding agent on BrowseComp or a research assistant on Terminal-Bench and you learn nothing. These are the families we track and why each exists:

Reasoning & knowledge

Expert science QA, hardened knowledge exams, and abstract puzzle-solving.

GPQA Diamondleader: GPT-5.6 Sol 93.6%
Graduate-level, Google-proof science questions written by PhD experts. A proxy for deep scientific reasoning beyond memorization.
MMLU-Proleader: DeepSeek-V4-Pro (0813) 87.5%
Hardened successor to MMLU across 14 domains with ten answer options; measures broad knowledge plus robust reasoning.
Humanity's Last Examleader: Kimi K2 Thinking 44.9%
Extremely difficult multi-disciplinary exam designed at the edge of human knowledge; frontier scores remain far from saturation.
ARC-AGI-2leader: Gemini 3.1 Pro 77.1%
Abstract-reasoning puzzles requiring novel rule induction from few examples. Designed to resist memorization and pattern-matching shortcuts.

Coding

Resolving real GitHub issues, long-horizon software engineering, fresh contest problems.

SWE-bench Verifiedleader: Kimi K3 93.4%
Human-validated subset of real GitHub issues that must be resolved end-to-end in live repositories. The industry-standard agentic coding metric.
SWE-bench Proleader: Claude Fable 5 80%
Harder, contamination-resistant split of SWE-bench with longer-horizon enterprise-style tasks. Top frontier models score well below Verified.
LiveCodeBenchleader: DeepSeek-V4-Pro (0813) 93.5%
Continuously refreshed competitive-programming problems published after model cutoffs, neutralizing training-set contamination.

Agents & computer use

Terminal work, full-desktop computer control, and persistent web research.

Terminal-Bench 2leader: Grok 4.5 83.3%
Autonomous agent tasks executed inside a real terminal: build systems, debugging, sysadmin work across many turns of tool use.
OSWorld 2.0leader: GPT-5.6 Sol 62.6%
Computer-use agent benchmark driving full desktop environments — clicking, typing, cross-app workflows like a human operator.
BrowseCompleader: GPT-5.6 Sol 92.2%
Agentic web research: locating hard-to-find facts across the live internet through persistent search-and-read behavior.

Math

Exact competition mathematics under strict answer checking.

AIME 2026leader: Kimi K2 Thinking 94.5%
American Invitational Mathematics Examination problems — competition math demanding exact multi-step symbolic reasoning.

Read the fine print

Missing scores show as “—”. We never extrapolate numbers a lab hasn't published; sparse cells mean exactly that.
Many headline figures are vendor-reported on vendor-chosen harnesses. Where independent harnesses disagree materially (e.g. Kimi K3, DeepSeek V4 SWE-bench runs) the caveat is noted on the model page.
Benchmark columns aren't perfectly comparable across labs — SWE-bench Verified vs Pro vs Marathon measure different difficulty ceilings, so within-column comparison only.
Prices are standard-tier USD per million tokens from public rate cards; subscription-first releases (GLM-5.3, Deep Think) have none.
Release dates reflect best-documented announcements; some sources differ by days. Data compiled August 26, 2026.

Model

Overall

SWE-V

GPQA

OSWorld

$ & ctx

Released

N/A

Kimi K3Moonshot

93.4

—

$61M

Jul 22, 2026

DeepSeek-V4-Pro (0813)DeepSeek · new

87.9

80.6

90.1

—

$1.401M

Aug 13, 2026

N/A

Kimi K2.6Moonshot

80.2

—

$1.07256K

Apr 8, 2026

N/A

Claude Fable 5Anthropic

—

$161M

Jun 9, 2026

N/A

DeepSeek-V4 FlashDeepSeek

—

$0.81512K

Apr 24, 2026

GPT-5.6 SolOpenAI

78.3

—

93.6

62.6

$11.251.1M

Jul 9, 2026

N/A

Gemini 3 FlashGoogle

—

$1.131M

Dec 17, 2025

N/A

Gemini 3.1 ProGoogle

77.1

—

$4.501M

Feb 19, 2026

N/A

Gemini 3 ProGoogle

76.2

—

$4.501M

Nov 18, 2025

N/A

Grok 4.5xAI

—

$3500K

Jul 16, 2026

N/A

Claude Haiku 4.5Anthropic

73.3

—

$2200K

Oct 15, 2025

Kimi K2 ThinkingMoonshot

70.2

71.3

—

$1.07256K

Nov 6, 2025

N/A

GLM-5.2Z.ai

62.1

—

$2.151M

Jun 13, 2026

N/A

Grok 4.6xAI · new

—

$3500K

Aug 14, 2026

N/A

GLM-5.3Z.ai · new

—

—1M

Aug 14, 2026

N/A

Muse Glimmer 30BMeta · new

—

—256K

Aug 10, 2026

N/A

Qwen3.8 MaxQwen · new

—

$31M

Aug 5, 2026

N/A

Claude Opus 5Anthropic

—

$10200K

Jul 24, 2026

N/A

GPT-5.6 TerraOpenAI

—

$4.501.1M

Jul 9, 2026

N/A

GPT-5.6 LunaOpenAI

—

$0.451.1M

Jul 9, 2026

N/A

Muse Spark 1.1Meta

—

—500K

Jul 9, 2026

N/A

Claude Sonnet 5Anthropic

—

$61M

Jun 30, 2026

N/A

MAI-Thinking-1Microsoft

—

—256K

Jun 2, 2026

N/A

Claude Opus 4.8Anthropic

—

$10200K

May 28, 2026

N/A

GPT-5.5 ProOpenAI

—

—400K

Apr 24, 2026

N/A

GPT-5.5OpenAI

—

$3.44400K

Apr 24, 2026

N/A

GLM-5Z.ai

—

$1.55200K

Feb 11, 2026

N/A

Nova 2 ProAmazon

—

—300K

Dec 2, 2025

N/A

Nova 2 LiteAmazon

—

—250K

Dec 2, 2025

N/A

Nova 2 OmniAmazon

—

—300K

Dec 2, 2025

N/A

Grok 4 Fast ReasoningxAI

—

$0.352M

Dec 1, 2025

N/A

MiniMax-M2MiniMax

—

$0.52205K

Oct 27, 2025

N/A

Gemini 3 Deep ThinkGoogle

—

—1M

Aug 1, 2025

N/A

Mistral Medium 3.1Mistral

—

$0.80128K

Aug 1, 2025

N/A

Qwen3 235B Instruct 2507Qwen

—

$0.39262K

Jul 24, 2025

N/A

Qwen3 Coder FlashQwen

—

$0.52262K

Jul 23, 2025

N/A

Devstral Medium 2507Mistral

—

$0.80128K

Jul 10, 2025

N/A

Llama 4 MaverickMeta

—

$0.421M

Apr 5, 2025

What the benchmarks actually measure

Reasoning & knowledge

Expert science QA, hardened knowledge exams, and abstract puzzle-solving.

GPQA Diamondleader: GPT-5.6 Sol 93.6%
Graduate-level, Google-proof science questions written by PhD experts. A proxy for deep scientific reasoning beyond memorization.
MMLU-Proleader: DeepSeek-V4-Pro (0813) 87.5%
Hardened successor to MMLU across 14 domains with ten answer options; measures broad knowledge plus robust reasoning.
Humanity's Last Examleader: Kimi K2 Thinking 44.9%
Extremely difficult multi-disciplinary exam designed at the edge of human knowledge; frontier scores remain far from saturation.
ARC-AGI-2leader: Gemini 3.1 Pro 77.1%
Abstract-reasoning puzzles requiring novel rule induction from few examples. Designed to resist memorization and pattern-matching shortcuts.

Coding

Resolving real GitHub issues, long-horizon software engineering, fresh contest problems.

SWE-bench Verifiedleader: Kimi K3 93.4%
Human-validated subset of real GitHub issues that must be resolved end-to-end in live repositories. The industry-standard agentic coding metric.
SWE-bench Proleader: Claude Fable 5 80%
Harder, contamination-resistant split of SWE-bench with longer-horizon enterprise-style tasks. Top frontier models score well below Verified.
LiveCodeBenchleader: DeepSeek-V4-Pro (0813) 93.5%
Continuously refreshed competitive-programming problems published after model cutoffs, neutralizing training-set contamination.

Agents & computer use

Terminal work, full-desktop computer control, and persistent web research.

Terminal-Bench 2leader: Grok 4.5 83.3%
Autonomous agent tasks executed inside a real terminal: build systems, debugging, sysadmin work across many turns of tool use.
OSWorld 2.0leader: GPT-5.6 Sol 62.6%
Computer-use agent benchmark driving full desktop environments — clicking, typing, cross-app workflows like a human operator.
BrowseCompleader: GPT-5.6 Sol 92.2%
Agentic web research: locating hard-to-find facts across the live internet through persistent search-and-read behavior.

Math

Exact competition mathematics under strict answer checking.

AIME 2026leader: Kimi K2 Thinking 94.5%
American Invitational Mathematics Examination problems — competition math demanding exact multi-step symbolic reasoning.

Read the fine print

Missing scores show as “—”. We never extrapolate numbers a lab hasn't published; sparse cells mean exactly that.
Many headline figures are vendor-reported on vendor-chosen harnesses. Where independent harnesses disagree materially (e.g. Kimi K3, DeepSeek V4 SWE-bench runs) the caveat is noted on the model page.
Benchmark columns aren't perfectly comparable across labs — SWE-bench Verified vs Pro vs Marathon measure different difficulty ceilings, so within-column comparison only.
Prices are standard-tier USD per million tokens from public rate cards; subscription-first releases (GLM-5.3, Deep Think) have none.
Release dates reflect best-documented announcements; some sources differ by days. Data compiled August 26, 2026.