Models · Anthropic · Out sinceReleased 24 Jul 2026

Claude Opus 5#2 for research and analysis.#2 for research and analysis, best at max effort.

Available on OpenRouterOn OpenRouterAvailable nowGenerally availableClosed (can't be downloaded)Proprietary

Where you can use it:In apps:Claude (Team plan) · Opus 5

Claude Opus 5 is made by Anthropic. Among the models we track it ranks #2 for research and analysis, #6 for writing, #6 for coding. It's expensive to use.

API id claude-opus-5. Default effort high; thinking cannot be disabled at xhigh/max. Previous-generation Opus.

Writing & creativity: 71.1 / 100 · #6
Research & analysis: 84.7 / 100 · #2
Coding: 80.5 / 100 · #6
Price: Expensive$5 / $25
Price per 1M (blended)Blended / 1M: $10
MemoryContext: 1M
Longest answerMax output: 128K
Test resultsResults: 222 (153 independent153 indep.)
Out sinceReleased: 24 Jul 2026
Made byVendor: Anthropic
UnderstandsInputs: text, image

Scores are out of 100 for each area, compared with every model we track; “#” is its rank. Context is how much text it can read at once (1M is roughly 700,000 words). Blended price mixes the cost of what you send and what it writes back.

Thinking levels

Reasoning settings

LowMediumHighExtra highMax

Highlighted: where it did best for research and analysis (Maximum thinking).

Highlighted: dominant setting in its research and analysis composite.

Links

Using it through OpenRouter

On OpenRouter

OpenRouter is a service that gives access to many AI models in one place. This model is listed there as anthropic/claude-opus-5.

Route it as anthropic/claude-opus-5 at $5 in / $25 out per 1M tokens, 1M context. Listed since 24 Jul 2026.

include_reasoningmax_completion_tokensmax_tokensreasoningreasoning_effortresponse_formatstopstructured_outputstemperaturetool_choicetoolsverbosity

Thinking level

Reasoning effort

How long should Claude Opus 5Where Claude Opus 5think?peaks.

Each chart is one test. Left to right, the model thinks longer; higher is a better result. The yellow ring marks where it did best. Thinking level is how long the AI thinks before it answers. Higher is slower and costs more, and it isn't always better.

Score per reasoning setting on each benchmark with ≥2 settings; ring = peak. Canonical results (independent over vendor).

Artificial Analysis Coding Agent Index

Best at Extra high, worse abovePeaks at Extra high

Terminal-Bench 4.0

Best at MaxBest at Max

CursorBench

Best at MaxBest at Max

DeepSWE

Best at MaxBest at Max

FrontierCode

Best at MediumBest at Medium

Frontier-Bench v0.1

Best at Extra high, worse abovePeaks at Extra high

LMArena Code Arena (WebDev)

Best at MaxBest at Max

SWE Atlas - Codebase QnA

Best at Extra high, worse abovePeaks at Extra high

ProgramBench (fully resolved)

Best at Extra high, worse abovePeaks at Extra high

Terminal-Bench 3.0

Best at MaxBest at Max

Terminal-Bench 2.1

Best at MaxBest at Max

LMArena Text - Coding category

Best at High, worse abovePeaks at High

SciCode

Best at MaxBest at Max

Agents' Last Exam

Best at High, worse abovePeaks at High

Artificial Analysis Intelligence Index

Best at MaxBest at Max

AutomationBench

Best at MaxBest at Max

OSWorld 2.0

Best at MaxBest at Max

OSWorld 2.0 offline set

Best at MaxBest at Max

AA-Briefcase

Best at MaxBest at Max

ARC-AGI-2

Best at MaxBest at Max

GDPval-AA v2

Best at Extra high, worse abovePeaks at Extra high

GDPval-AA v2.1

Best at MaxBest at Max

GPQA Diamond

Best at High, worse abovePeaks at High

Humanity's Last Exam

Best at MaxBest at Max

LLM Creative Story-Writing Benchmark (Lech Mazur)

Best at Extra highBest at Extra high

LMArena Text - Creative Writing

Best at High, worse abovePeaks at High

LMArena Text - Expert

Best at High, worse abovePeaks at High

LMArena Text - Hard Prompts

Best at High, worse abovePeaks at High

LMArena Text - Instruction Following

Best at High, worse abovePeaks at High

LMArena Text - Longer Query

Best at High, worse abovePeaks at High

LMArena Text - Non-English

Best at High, worse abovePeaks at High

LMArena Text - Occupational: Legal & Government

Best at High, worse abovePeaks at High

LMArena Text - Occupational: Writing, Literature & Language

Best at High, worse abovePeaks at High

LMArena Text (overall)

Best at High, worse abovePeaks at High

WANDR

Best at MaxBest at Max

Every test result

Every result

All the numbers,with where they came from.

Every result we've found for this model, with who measured it and a link to where we read it.

All raw rows (vendor and independent kept separate), with setting label, source, date, harness and notes.

All results for this model
						NotesNotes
AA Analyst Agent	53.8%	MaxClaude Opus 5 (Adaptive Reasoning, Max Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run; scores move in 1.25-pt steps (small task set)
AA-Briefcase	1606	Highadaptive thinking, effort=high	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	AA-Briefcase (pre-v1.1), run by Artificial Analysis; xhigh/high use 15%/33% fewer output tokens than max
AA-Briefcase	1693	Extra highadaptive thinking, effort=xhigh	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	AA-Briefcase (pre-v1.1), run by Artificial Analysis; xhigh/high use 15%/33% fewer output tokens than max
AA-Briefcase	1720	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	AA-Briefcase (pre-v1.1), run by Artificial Analysis; xhigh/high use 15%/33% fewer output tokens than max
AA-Briefcase v1.1	1673	MaxClaude Opus 5 (Adaptive Reasoning, Max Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—
AA-Briefcase v1.1	1673	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—
Agents' Last Exam	51.9%	Loweffort=low	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	Agents' Last Exam V1 (long-horizon professional workflows, 55 sub-industries); vendor-estimated API cost/task $4.0619; competitor number as plotted by OpenAI
Agents' Last Exam	53.0%	Mediumeffort=medium	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	Agents' Last Exam V1 (long-horizon professional workflows, 55 sub-industries); vendor-estimated API cost/task $5.2907; competitor number as plotted by OpenAI
Agents' Last Exam	55.9%	Higheffort=high	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	Agents' Last Exam V1 (long-horizon professional workflows, 55 sub-industries); vendor-estimated API cost/task $7.2917; competitor number as plotted by OpenAI
Agents' Last Exam	55.5%	Extra higheffort=xhigh	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	Agents' Last Exam V1 (long-horizon professional workflows, 55 sub-industries); vendor-estimated API cost/task $10.0226; competitor number as plotted by OpenAI
Agents' Last Exam	52.7%	Maxeffort=max	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	Agents' Last Exam V1 (long-horizon professional workflows, 55 sub-industries); vendor-estimated API cost/task $9.7583; competitor number as plotted by OpenAI
ARC-AGI-1	97.5%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—
ARC-AGI-2	88.3%	HighClaude Opus 5 (High)	Independent testIndependentARC Prize ↗	30 Sep 2026	—	semi-private set; $1.45/task; data https://arcprize.org/media/data/leaderboard/v2.json
ARC-AGI-2	90.4%	MaxClaude Opus 5 (Max)	Independent testIndependentARC Prize ↗	30 Sep 2026	—	semi-private set; $2.06/task; data https://arcprize.org/media/data/leaderboard/v2.json
ARC-AGI-2	90.4%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	semi-private set
ARC-AGI-3	30.2%	HighClaude Opus 5 (High)	Independent testIndependentARC Prize ↗	30 Sep 2026	—	semi-private set; total cost $20,657; data https://arcprize.org/media/data/leaderboard/v3.json
ARC-AGI-3	30.2%	Highadaptive thinking, effort=high	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	ARC Prize Foundation verified, semi-private set (RHAE 30.16%); max-effort result not available at release
Artificial Analysis Coding Agent Index	59.2	No reasoningeffort=none	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	3 Sep 2026	—	Artificial Analysis Coding Agent Index v1.4 (index score 0-100); vendor-estimated API cost/task $3.53; competitor number as plotted by OpenAI
Artificial Analysis Coding Agent Index	59.4	Loweffort=low	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	3 Sep 2026	—	Artificial Analysis Coding Agent Index v1.4 (index score 0-100); vendor-estimated API cost/task $2.3; competitor number as plotted by OpenAI
Artificial Analysis Coding Agent Index	64.1	Mediumeffort=medium	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	3 Sep 2026	—	Artificial Analysis Coding Agent Index v1.4 (index score 0-100); vendor-estimated API cost/task $3.17; competitor number as plotted by OpenAI
Artificial Analysis Coding Agent Index	65.6	Higheffort=high	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	3 Sep 2026	—	Artificial Analysis Coding Agent Index v1.4 (index score 0-100); vendor-estimated API cost/task $3.92; competitor number as plotted by OpenAI
Artificial Analysis Coding Agent Index	68.1	Extra higheffort=xhigh	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	3 Sep 2026	—	Artificial Analysis Coding Agent Index v1.4 (index score 0-100); vendor-estimated API cost/task $8.17; competitor number as plotted by OpenAI
Artificial Analysis Coding Agent Index	67	Maxeffort=max	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	3 Sep 2026	—	Artificial Analysis Coding Agent Index v1.4 (index score 0-100); vendor-estimated API cost/task $8.94; competitor number as plotted by OpenAI
Artificial Analysis Coding Agent Index	59.7	MaxOpus 5 (max)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	Claude Code	agent Claude Code; components: DeepSWE v1.1 62.5, SWE-Atlas-QnA 62.1, Terminal-Bench v4 54.5; avg cost $10.79/task; avg wall time 42 min/task
Artificial Analysis Intelligence Index	39.4	LowClaude Opus 5 (Adaptive Reasoning, Low Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug claude-opus-5-low; list price $5/25 per 1M in/out; cost to run AA Intelligence Index $1.10/task (AA marks this variant deprecated)
Artificial Analysis Intelligence Index	44.8	MediumClaude Opus 5 (Adaptive Reasoning, Medium Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug claude-opus-5-medium; list price $5/25 per 1M in/out; cost to run AA Intelligence Index $2.19/task (AA marks this variant deprecated)
Artificial Analysis Intelligence Index	48.1	HighClaude Opus 5 (Adaptive Reasoning, High Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug claude-opus-5-high; list price $5/25 per 1M in/out; cost to run AA Intelligence Index $3.61/task (AA marks this variant deprecated)
Artificial Analysis Intelligence Index	49.7	Extra highClaude Opus 5 (Adaptive Reasoning, Xhigh Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug claude-opus-5-xhigh; list price $5/25 per 1M in/out; cost to run AA Intelligence Index $4.88/task (AA marks this variant deprecated)
Artificial Analysis Intelligence Index	50.8	MaxClaude Opus 5 (Adaptive Reasoning, Max Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA slug claude-opus-5; list price $5/25 per 1M in/out; cost to run AA Intelligence Index $5.86/task (AA marks this variant deprecated)
Artificial Analysis output speed	47 tok/s	LowClaude Opus 5 (Adaptive Reasoning, Low Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	tokens/sec (median output speed, first-party API); TTFT 3.0s; list price $5/25 per 1M in/out (AA marks this variant deprecated)
Artificial Analysis output speed	50 tok/s	MediumClaude Opus 5 (Adaptive Reasoning, Medium Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	tokens/sec (median output speed, first-party API); TTFT 10.9s; list price $5/25 per 1M in/out (AA marks this variant deprecated)
Artificial Analysis output speed	53 tok/s	HighClaude Opus 5 (Adaptive Reasoning, High Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	tokens/sec (median output speed, first-party API); TTFT 22.9s; list price $5/25 per 1M in/out (AA marks this variant deprecated)
Artificial Analysis output speed	51 tok/s	Extra highClaude Opus 5 (Adaptive Reasoning, Xhigh Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	tokens/sec (median output speed, first-party API); TTFT 48.4s; list price $5/25 per 1M in/out (AA marks this variant deprecated)
Artificial Analysis output speed	51 tok/s	MaxClaude Opus 5 (Adaptive Reasoning, Max Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	tokens/sec (median output speed, first-party API); TTFT 68.3s; list price $5/25 per 1M in/out (AA marks this variant deprecated)
ArXivMath (no tools)	90.8%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	ArXivMath June 2026, no tools
ArXivMath (with tools)	91.3%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	June 2026, with tools
AutomationBench	20.4%	Lowadaptive thinking, effort=low	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	AutomationBench (Zapier; end-to-end business workflows across connected apps); Zapier public leaderboard; vendor-reported cost/task $0.75
AutomationBench	24.0%	Mediumadaptive thinking, effort=medium	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	$0.89/task
AutomationBench	23.9%	Mediumadaptive thinking, effort=med	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	AutomationBench (Zapier; end-to-end business workflows across connected apps); Zapier public leaderboard; vendor-reported cost/task $0.89
AutomationBench	20.5%	Higheffort=high	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	AutomationBench 1.0.6 (Zapier); vendor-estimated API cost/task $2.27; competitor number as plotted by OpenAI
AutomationBench	20.6%	Highadaptive thinking, effort=high	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	AutomationBench (Zapier; end-to-end business workflows across connected apps); Zapier public leaderboard; vendor-reported cost/task $1.03
AutomationBench	25.3%	Extra highadaptive thinking, effort=xhigh	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	AutomationBench (Zapier; end-to-end business workflows across connected apps); Zapier public leaderboard; vendor-reported cost/task $1.15
AutomationBench	26.9%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	AutomationBench (Zapier; end-to-end business workflows across connected apps); Zapier public leaderboard; vendor-reported cost/task $1.27
BrowseComp	90.8%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	read from launch-post benchmark table image (values printed in table)
Chartography (with tools)	83.4%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—
CursorBench	40.7%	LowOpus 5 Low	Independent testIndependentCursor (CursorBench) ↗	1 Oct 2026	Cursor agent	CursorBench 4.0 (tasks from real Cursor sessions; v4.0 introduced 2026-09-10); rank 26; $4.87/task; 31,995 tokens/task; 57 steps/task
CursorBench	40.7%	Lowadaptive thinking, effort=low	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Cursor agent	CursorBench 4.0, run end-to-end by Cursor in its production agent harness; vendor-reported cost/task $4.87
CursorBench	43.3%	MediumOpus 5 Medium	Independent testIndependentCursor (CursorBench) ↗	1 Oct 2026	Cursor agent	CursorBench 4.0 (tasks from real Cursor sessions; v4.0 introduced 2026-09-10); rank 19; $6.94/task; 45,272 tokens/task; 72 steps/task
CursorBench	43.3%	Mediumadaptive thinking, effort=med	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Cursor agent	CursorBench 4.0, run end-to-end by Cursor in its production agent harness; vendor-reported cost/task $6.94
CursorBench	44.7%	HighOpus 5 High	Independent testIndependentCursor (CursorBench) ↗	1 Oct 2026	Cursor agent	CursorBench 4.0 (tasks from real Cursor sessions; v4.0 introduced 2026-09-10); rank 16; $9.00/task; 61,405 tokens/task; 86 steps/task
CursorBench	44.7%	Highadaptive thinking, effort=high	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Cursor agent	CursorBench 4.0, run end-to-end by Cursor in its production agent harness; vendor-reported cost/task $9.0
CursorBench	46.1%	Extra highOpus 5 Extra High	Independent testIndependentCursor (CursorBench) ↗	1 Oct 2026	Cursor agent	CursorBench 4.0 (tasks from real Cursor sessions; v4.0 introduced 2026-09-10); rank 14; $11.43/task; 80,094 tokens/task; 103 steps/task
CursorBench	46.1%	Extra highadaptive thinking, effort=xhigh	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Cursor agent	CursorBench 4.0, run end-to-end by Cursor in its production agent harness; vendor-reported cost/task $11.43
CursorBench	46.6%	MaxOpus 5 Max	Independent testIndependentCursor (CursorBench) ↗	1 Oct 2026	Cursor agent	CursorBench 4.0 (tasks from real Cursor sessions; v4.0 introduced 2026-09-10); rank 12; $11.95/task; 85,384 tokens/task; 106 steps/task
CursorBench	46.6%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Cursor agent	CursorBench 4.0, run end-to-end by Cursor in its production agent harness; vendor-reported cost/task $11.95
CursorBench 3.2.0	70.0%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	1 Sep 2026	Cursor agent	CursorBench 3.2.0 (Cursor production harness); not comparable to CursorBench 4.0
DeepSWE	58.1%	Loweffort=low	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	DeepSWE v1.1 (113 long-horizon SWE tasks); GPT models run in Codex; vendor-estimated API cost/task $1.66; competitor number as plotted by OpenAI
DeepSWE	58.1%	Lowclaude-opus-5_low	Independent testIndependentDeepSWE (Datacurve) via Epoch AI ↗	1 Oct 2026	mini-SWE-agent	read from Epoch AI benchmark_data.zip (deepswe_external.csv); original leaderboard https://deepswe.datacurve.ai/; harness mini-swe-agent; pass@4 85.0%; ±2.3; 4 runs; $1.66/task
DeepSWE	68.9%	Mediumclaude-opus-5_medium	Independent testIndependentDeepSWE (Datacurve) via Epoch AI ↗	1 Oct 2026	mini-SWE-agent	read from Epoch AI benchmark_data.zip (deepswe_external.csv); original leaderboard https://deepswe.datacurve.ai/; harness mini-swe-agent; pass@4 89.4%; ±1.2; 4 runs; $3.29/task
DeepSWE	68.9%	Mediumeffort=medium	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	DeepSWE v1.1 (113 long-horizon SWE tasks); GPT models run in Codex; vendor-estimated API cost/task $3.29; competitor number as plotted by OpenAI
DeepSWE	72.8%	Higheffort=high	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	DeepSWE v1.1 (113 long-horizon SWE tasks); GPT models run in Codex; vendor-estimated API cost/task $6.08; competitor number as plotted by OpenAI
DeepSWE	72.8%	Highclaude-opus-5_high	Independent testIndependentDeepSWE (Datacurve) via Epoch AI ↗	1 Oct 2026	mini-SWE-agent	read from Epoch AI benchmark_data.zip (deepswe_external.csv); original leaderboard https://deepswe.datacurve.ai/; harness mini-swe-agent; pass@4 87.6%; ±1.9; 4 runs; $6.08/task
DeepSWE	73.2%	Extra highclaude-opus-5_xhigh	Independent testIndependentDeepSWE (Datacurve) via Epoch AI ↗	1 Oct 2026	mini-SWE-agent	read from Epoch AI benchmark_data.zip (deepswe_external.csv); original leaderboard https://deepswe.datacurve.ai/; harness mini-swe-agent; pass@4 85.8%; ±3.1; 4 runs; $9.07/task
DeepSWE	73.2%	Extra higheffort=xhigh	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	DeepSWE v1.1 (113 long-horizon SWE tasks); GPT models run in Codex; vendor-estimated API cost/task $9.07; competitor number as plotted by OpenAI
DeepSWE	73.7%	Maxeffort=max	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	DeepSWE v1.1 (113 long-horizon SWE tasks); GPT models run in Codex; vendor-estimated API cost/task $11.84; competitor number as plotted by OpenAI
DeepSWE	73.6%	Maxclaude-opus-5_max	Independent testIndependentDeepSWE (Datacurve) via Epoch AI ↗	1 Oct 2026	mini-SWE-agent	read from Epoch AI benchmark_data.zip (deepswe_external.csv); original leaderboard https://deepswe.datacurve.ai/; harness mini-swe-agent; pass@4 88.5%; ±3.9; 4 runs; $11.84/task
DeepSWE	62.5%	MaxOpus 5 (max)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	Claude Code	AA Coding Agent Index component; pass@1 avg of 3 attempts
DeepSWE	68.8%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	DeepSWE v1.1, avg of 5 trials; read from launch-post benchmark table image (values printed in table)
Design Arena (all categories)	1329	Defaultclaude-opus-5	Independent testIndependentDesign Arena ↗	1 Oct 2026	—	Bradley-Terry Elo; ±3.0 SE; 15954 battles; win rate 57.9%
Design Arena (fullstack)	1304	Defaultclaude-opus-5	Independent testIndependentDesign Arena ↗	1 Oct 2026	—	Bradley-Terry Elo; ±10.7 SE; 1195 battles; win rate 57.3%
Epoch Capabilities Index	162.9	Default	Independent testIndependentEpoch AI ↗	1 Oct 2026	—	Epoch Capabilities Index; 90% CI 160.2-166.7; best of listed model versions
EQ-Bench Creative Writing v3 (Elo)	2133	Defaultclaude-opus-5	Independent testIndependentEQ-Bench ↗	1 Oct 2026	—	leaderboard rank 3; rubric score 17.07/20; slop 6.59; avg length 6003 chars; Elo judged by Claude Sonnet 4.6; effort not stated by EQ-Bench
Frontier-Bench v0.1	25.0%	Lowadaptive thinking, effort=low	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	mini-SWE-agent (GKE)	Frontier-Bench v0.1 internal run, 74 tasks, mean reward over 5 attempts; xhigh best, max "within noise, landing at 43%"; high uses 19% fewer output tokens than xhigh, low 64% fewer
Frontier-Bench v0.1	39.0%	Highadaptive thinking, effort=high	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	mini-SWE-agent (GKE)	Frontier-Bench v0.1 internal run, 74 tasks, mean reward over 5 attempts; xhigh best, max "within noise, landing at 43%"; high uses 19% fewer output tokens than xhigh, low 64% fewer
Frontier-Bench v0.1	44.4%	Extra highadaptive thinking, effort=xhigh	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	mini-SWE-agent (GKE)	Frontier-Bench v0.1 internal run, 74 tasks, mean reward over 5 attempts; xhigh best, max "within noise, landing at 43%"; high uses 19% fewer output tokens than xhigh, low 64% fewer
Frontier-Bench v0.1	43.3%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	Harbor	Frontier-Bench v0.1 (agentic terminal coding); table value from Harbor evaluation; read from launch-post benchmark table image (values printed in table)
Frontier-Bench v0.1	43.0%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	mini-SWE-agent (GKE)	Frontier-Bench v0.1 internal run, 74 tasks, mean reward over 5 attempts; xhigh best, max "within noise, landing at 43%"; high uses 19% fewer output tokens than xhigh, low 64% fewer
FrontierCode	41.9%	Loweffort=low	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	FrontierCode 1.1 Main (Cognition); GPT models run in Codex CLI; vendor-estimated API cost/task $2.68; competitor number as plotted by OpenAI
FrontierCode	42.0%	Lowadaptive thinking, effort=low	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Claude Code	FrontierCode v1.1 Main (150 tasks, run & scored by Cognition; composite of correctness + mergeability rubric; penalizes out-of-scope edits); Opus 5 peaks at medium; vendor-reported cost/task $2.64
FrontierCode	53.4%	Mediumadaptive thinking, effort=med	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Claude Code	FrontierCode v1.1 Main (150 tasks, run & scored by Cognition; composite of correctness + mergeability rubric; penalizes out-of-scope edits); Opus 5 peaks at medium; vendor-reported cost/task $4.61
FrontierCode	48.0%	Highadaptive thinking, effort=high	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Claude Code	FrontierCode v1.1 Main (150 tasks, run & scored by Cognition; composite of correctness + mergeability rubric; penalizes out-of-scope edits); Opus 5 peaks at medium; vendor-reported cost/task $7.62
FrontierCode	43.6%	Extra highadaptive thinking, effort=xhigh	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Claude Code	FrontierCode v1.1 Main (150 tasks, run & scored by Cognition; composite of correctness + mergeability rubric; penalizes out-of-scope edits); Opus 5 peaks at medium; vendor-reported cost/task $8.99
FrontierCode	53.4%	Maxclaude-opus-5_max	Independent testIndependentFrontierCode (Cognition) via Epoch AI ↗	1 Oct 2026	Claude Code	read from Epoch AI benchmark_data.zip (frontiercode_external.csv); original leaderboard https://cognition.com/frontiercode; harness claude-code; Mean@5
FrontierCode	48.0%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Claude Code	FrontierCode v1.1 Main (150 tasks, run & scored by Cognition; composite of correctness + mergeability rubric; penalizes out-of-scope edits); Opus 5 peaks at medium; vendor-reported cost/task $12.28
FrontierCode v1.1 (Extended)	63.6%	Mediumadaptive thinking, effort=medium (best)	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	Claude Code	FrontierCode v1.1 Extended, best score at medium
FrontierMath (Tiers 1-3)	85.6%	Maxclaude-opus-5_max	Independent testIndependentEpoch AI ↗	24 Jul 2026	—	FrontierMath Tiers 1-3 (v2), Epoch-run; stderr 2.1pt; data https://epoch.ai/data/benchmark_data.zip (frontiermath_tiers_1_3_v2.csv)
FrontierMath Tier 4	73.2%	Maxclaude-opus-5_max	Independent testIndependentEpoch AI ↗	24 Jul 2026	—	FrontierMath Tier 4 (v2), Epoch-run; stderr 7.0pt; data https://epoch.ai/data/benchmark_data.zip (frontiermath_tier_4_v2.csv)
FrontierSWE v2	52.0%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	1 Sep 2026	Proximal harness	reported as 0.52
GDPval-AA v2	1827	Extra highadaptive thinking, effort=xhigh	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	xhigh uses 25% fewer output tokens than max (1861)
GDPval-AA v2	1824	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	1 Sep 2026	—	read from launch-post benchmark table image (values printed in table)
GDPval-AA v2.1	1294	Lowadaptive thinking, effort=low	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	GDPval-AA v2.1 (Artificial Analysis, Elo from blind pairwise comparisons, 44 occupations); vendor-reported cost/task $0.58
GDPval-AA v2.1	1476	Mediumadaptive thinking, effort=med	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	GDPval-AA v2.1 (Artificial Analysis, Elo from blind pairwise comparisons, 44 occupations); vendor-reported cost/task $1.36
GDPval-AA v2.1	1581	Highadaptive thinking, effort=high	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	GDPval-AA v2.1 (Artificial Analysis, Elo from blind pairwise comparisons, 44 occupations); vendor-reported cost/task $3.03
GDPval-AA v2.1	1676	Extra highadaptive thinking, effort=xhigh	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	GDPval-AA v2.1 (Artificial Analysis, Elo from blind pairwise comparisons, 44 occupations); vendor-reported cost/task $4.96
GDPval-AA v2.1	1708	MaxClaude Opus 5 (Adaptive Reasoning, Max Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	95% CI 1682.9-1732.95
GDPval-AA v2.1	1708	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	GDPval-AA v2.1 (Artificial Analysis, Elo from blind pairwise comparisons, 44 occupations); vendor-reported cost/task $6.76
GPQA Diamond	88.9%	LowClaude Opus 5 (Adaptive Reasoning, Low Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-low) (AA marks this variant deprecated)
GPQA Diamond	91.9%	MediumClaude Opus 5 (Adaptive Reasoning, Medium Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-medium) (AA marks this variant deprecated)
GPQA Diamond	93.7%	HighClaude Opus 5 (Adaptive Reasoning, High Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-high) (AA marks this variant deprecated)
GPQA Diamond	93.7%	Extra highClaude Opus 5 (Adaptive Reasoning, Xhigh Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-xhigh) (AA marks this variant deprecated)
GPQA Diamond	93.9%	Maxclaude-opus-5_max	Independent testIndependentEpoch AI ↗	24 Jul 2026	—	Epoch-run GPQA Diamond; stderr 1.5pt; data https://epoch.ai/data/benchmark_data.zip (gpqa_diamond.csv)
GPQA Diamond	93.2%	MaxClaude Opus 5 (Adaptive Reasoning, Max Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5) (AA marks this variant deprecated)
GPQA Diamond	93.4%	Default	Independent testIndependentVals.ai ↗	1 Sep 2026	—	±1.244 stderr; $0.095/test
GPQA Diamond	92.9%	Defaultclaude-opus-5	Independent testIndependentEpoch AI ↗	6 Aug 2026	—	Epoch-run GPQA Diamond; stderr 1.8pt; data https://epoch.ai/data/benchmark_data.zip (gpqa_diamond.csv)
Harvey LAB-AA	93.5%	MaxClaude Opus 5 (Adaptive Reasoning, Max Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	criteria pass rate, AA-run
HealthBench Professional	59.8%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	read from launch-post benchmark table image (values printed in table)
HLE Diamond	38.6%	Defaultclaude-opus-5	Independent testIndependentScale AI SEAL ↗	1 Oct 2026	—	rank 4 (Scale rank accounts for CI); ±3; entry added 2026-09-09
Humanity's Last Exam	43.4%	LowClaude Opus 5 (Adaptive Reasoning, Low Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-low) (AA marks this variant deprecated)
Humanity's Last Exam	51.3%	MediumClaude Opus 5 (Adaptive Reasoning, Medium Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-medium) (AA marks this variant deprecated)
Humanity's Last Exam	52.8%	HighClaude Opus 5 (Adaptive Reasoning, High Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-high) (AA marks this variant deprecated)
Humanity's Last Exam	54.4%	Extra highClaude Opus 5 (Adaptive Reasoning, Xhigh Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-xhigh) (AA marks this variant deprecated)
Humanity's Last Exam	54.9%	MaxClaude Opus 5 (Adaptive Reasoning, Max Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5) (AA marks this variant deprecated)
Humanity's Last Exam	56.6%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	1 Sep 2026	—	no tools
Humanity's Last Exam (with tools)	63.6%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	read from launch-post benchmark table image (values printed in table)
IOI (Vals)	84.3%	Maxeffort=max	Independent testIndependentVals.ai ↗	29 Sep 2026	—	±9.964 stderr; $16.477/test
LegalBench (Vals)	87.0%	Maxeffort=max	Independent testIndependentVals.ai ↗	29 Sep 2026	—	vals id anthropic/claude-opus-5; rank 8/149; ±0.417 stderr; $0.011976/test
LiveCodeBench	89.0%	Maxeffort=max	Independent testIndependentVals.ai ↗	1 Sep 2026	—	±0.913 stderr; $0.134/test
LLM Creative Story-Writing Benchmark (Lech Mazur)	3.5	HighClaude Opus 5 (high)	Independent testIndependentLech Mazur (LLM Creative Story-Writing Benchmark) ↗	28 Sep 2026	—	rank 5/56; Thurstone comparison score (centered at 0); est. win chance 91%; 95% bootstrap 3.344 to 3.557
LLM Creative Story-Writing Benchmark (Lech Mazur)	3.8	Extra highClaude Opus 5 (xhigh)	Independent testIndependentLech Mazur (LLM Creative Story-Writing Benchmark) ↗	28 Sep 2026	—	rank 2/56; Thurstone comparison score (centered at 0); est. win chance 93%; 95% bootstrap 3.726 to 3.836
LMArena Code Arena (WebDev)	1660	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	Code Arena \| WebDev overall (agentic web-dev, raw); rank 11 (CI rank 8-13); 95% CI 1654-1666; 21088 votes
LMArena Code Arena (WebDev)	1694	Maxclaude-opus-5-max	Independent testIndependentLMArena ↗	30 Sep 2026	—	Code Arena \| WebDev overall (agentic web-dev, raw); rank 6 (CI rank 5-8); 95% CI 1687-1700; 16747 votes
LMArena Text - Coding category	1534	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	text arena coding category, style control on; rank 15 (CI rank 7-31); 95% CI 1527-1540; 15273 votes
LMArena Text - Coding category	1530	Maxclaude-opus-5-max	Independent testIndependentLMArena ↗	30 Sep 2026	—	text arena coding category, style control on; rank 21 (CI rank 7-44); 95% CI 1522-1538; 7234 votes
LMArena Text - Creative Writing	1472	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 13 (rank range 7-33); 95% CI 1465.4-1479.3; 12964 votes; style-controlled
LMArena Text - Creative Writing	1470	Maxclaude-opus-5-max	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 15 (rank range 7-35); 95% CI 1460.9-1478.2; 6376 votes; style-controlled
LMArena Text - Expert	1542	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 3 (rank range 1-17); 95% CI 1533.8-1550.9; 6572 votes; style-controlled
LMArena Text - Expert	1530	Maxclaude-opus-5-max	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 12 (rank range 1-34); 95% CI 1518.1-1541.3; 3073 votes; style-controlled
LMArena Text - Hard Prompts	1516	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 11 (rank range 7-21); 95% CI 1511.2-1520.5; 38320 votes; style-controlled
LMArena Text - Hard Prompts	1515	Maxclaude-opus-5-max	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 13 (rank range 7-24); 95% CI 1508.7-1520.4; 18450 votes; style-controlled
LMArena Text - Instruction Following	1497	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 8 (rank range 4-16); 95% CI 1491.2-1502.8; 20700 votes; style-controlled
LMArena Text - Instruction Following	1493	Maxclaude-opus-5-max	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 10 (rank range 4-21); 95% CI 1485.6-1499.8; 9911 votes; style-controlled
LMArena Text - Longer Query	1505	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 10 (rank range 6-22); 95% CI 1499.3-1510.2; 27765 votes; style-controlled
LMArena Text - Longer Query	1500	Maxclaude-opus-5-max	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 16 (rank range 7-28); 95% CI 1493.0-1506.2; 13516 votes; style-controlled
LMArena Text - Non-English	1482	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 14 (rank range 5-23); 95% CI 1477.2-1486.7; 34543 votes; style-controlled
LMArena Text - Non-English	1481	Maxclaude-opus-5-max	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 15 (rank range 5-25); 95% CI 1475.0-1486.8; 16873 votes; style-controlled
LMArena Text - Occupational: Business, Management & Finance	1486	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 17 (rank range 6-37); 95% CI 1479.6-1493.3; 11158 votes; style-controlled
LMArena Text - Occupational: Legal & Government	1503	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 9 (rank range 1-35); 95% CI 1493.7-1512.6; 4946 votes; style-controlled
LMArena Text - Occupational: Legal & Government	1503	Maxclaude-opus-5-max	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 10 (rank range 1-43); 95% CI 1489.6-1515.3; 2454 votes; style-controlled
LMArena Text - Occupational: Writing, Literature & Language	1482	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 11 (rank range 7-26); 95% CI 1476.1-1488.6; 16518 votes; style-controlled
LMArena Text - Occupational: Writing, Literature & Language	1479	Maxclaude-opus-5-max	Independent testIndependentLMArena ↗	30 Sep 2026	—	rank 15 (rank range 7-33); 95% CI 1471.4-1486.9; 8098 votes; style-controlled
LMArena Text (overall)	1491	Highclaude-opus-5-high	Independent testIndependentLMArena ↗	30 Sep 2026	—	text overall, style control on; rank 13 (CI rank 5-22); 95% CI 1487-1495; 58380 votes
LMArena Text (overall)	1489	Maxclaude-opus-5-max	Independent testIndependentLMArena ↗	30 Sep 2026	—	text overall, style control on; rank 14 (CI rank 7-27); 95% CI 1484-1494; 28351 votes
MCP Atlas	85.8%	Extra highclaude-opus-5 (xhigh)	Independent testIndependentScale AI SEAL ↗	1 Oct 2026	—	rank 2 (Scale rank accounts for CI); ±2.1; entry added 2026-08-03
OfficeQA Pro	66.9%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—
OSWorld 2.0	22.3%	Lowclaude-opus-5_low	Independent testIndependentOSWorld 2.0 via Epoch AI ↗	1 Oct 2026	—	read from Epoch AI benchmark_data.zip (osworld_2_external.csv); original leaderboard https://osworld-v2.xlang.ai/; partial score 0.5489; tool setting batch tool; step budget 500
OSWorld 2.0	25.3%	Mediumclaude-opus-5_medium	Independent testIndependentOSWorld 2.0 via Epoch AI ↗	1 Oct 2026	—	read from Epoch AI benchmark_data.zip (osworld_2_external.csv); original leaderboard https://osworld-v2.xlang.ai/; partial score 0.6049; tool setting batch tool; step budget 500
OSWorld 2.0	29.0%	Highclaude-opus-5_high	Independent testIndependentOSWorld 2.0 via Epoch AI ↗	1 Oct 2026	—	read from Epoch AI benchmark_data.zip (osworld_2_external.csv); original leaderboard https://osworld-v2.xlang.ai/; partial score 0.6379; tool setting batch tool; step budget 500
OSWorld 2.0	30.2%	Extra highclaude-opus-5_xhigh	Independent testIndependentOSWorld 2.0 via Epoch AI ↗	1 Oct 2026	—	read from Epoch AI benchmark_data.zip (osworld_2_external.csv); original leaderboard https://osworld-v2.xlang.ai/; partial score 0.6767000000000001; tool setting batch tool; step budget 500
OSWorld 2.0	31.4%	Maxclaude-opus-5_max	Independent testIndependentOSWorld 2.0 via Epoch AI ↗	1 Oct 2026	—	read from Epoch AI benchmark_data.zip (osworld_2_external.csv); original leaderboard https://osworld-v2.xlang.ai/; partial score 0.6831; tool setting batch tool; step budget 500
OSWorld 2.0	75.4%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	1 Sep 2026	—	partial score; August 2026 task release; safeguard interventions scored 0
OSWorld 2.0 (strict pass rate)	39.6%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	1 Sep 2026	—	strict pass rate; August 2026 task release
OSWorld 2.0 offline set	55.2%	Loweffort=low	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	OSWorld 2.0 offline set, partial reward, v2026.08.08 release; vendor-estimated API cost/task $9.88; competitor number as plotted by OpenAI
OSWorld 2.0 offline set	60.3%	Mediumeffort=medium	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	OSWorld 2.0 offline set, partial reward, v2026.08.08 release; vendor-estimated API cost/task $12.67; competitor number as plotted by OpenAI
OSWorld 2.0 offline set	65.9%	Higheffort=high	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	OSWorld 2.0 offline set, partial reward, v2026.08.08 release; vendor-estimated API cost/task $15.95; competitor number as plotted by OpenAI
OSWorld 2.0 offline set	70.1%	Extra higheffort=xhigh	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	OSWorld 2.0 offline set, partial reward, v2026.08.08 release; vendor-estimated API cost/task $23.91; competitor number as plotted by OpenAI
OSWorld 2.0 offline set	70.2%	Maxeffort=max	Independent testIndependentOpenAI (competitor result in OpenAI launch post) ↗	22 Sep 2026	—	OSWorld 2.0 offline set, partial reward, v2026.08.08 release; vendor-estimated API cost/task $24.11; competitor number as plotted by OpenAI
OSWorld 2.1 (partial score)	74.0%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	partial score; re-evaluated by Anthropic under Opus 5.5 conditions
OSWorld 2.1 (strict pass rate)	37.2%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—
ProgramBench (avg test pass rate)	74.7%	Extra high	Independent testIndependentProgramBench ↗	28 Sep 2026	mini-SWE-agent	avg behavioral-test pass rate ("Score" column); fully resolved 4.5%, almost (>=95% tests) 37.0%; avg cost $50.53/task
ProgramBench (fully resolved)	4.5%	Extra high	Independent testIndependentProgramBench ↗	28 Sep 2026	mini-SWE-agent	strict fully-resolved rate; almost (>=95% tests) 37.0%; avg cost $50.53/task
ProgramBench (fully resolved)	3.0%	Maxeffort=max	Independent testIndependentVals.ai ↗	29 Sep 2026	mini-SWE-agent	±1.209 stderr; $60.289/test; strict fully-resolved rate
ProgramBench (fully resolved)	85.4%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	mini-swe-agent
SciCode	49.2%	LowClaude Opus 5 (Adaptive Reasoning, Low Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-low) (AA marks this variant deprecated)
SciCode	51.5%	MediumClaude Opus 5 (Adaptive Reasoning, Medium Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-medium) (AA marks this variant deprecated)
SciCode	55.4%	HighClaude Opus 5 (Adaptive Reasoning, High Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-high) (AA marks this variant deprecated)
SciCode	55.7%	Extra highClaude Opus 5 (Adaptive Reasoning, Xhigh Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-xhigh) (AA marks this variant deprecated)
SciCode	56.4%	MaxClaude Opus 5 (Adaptive Reasoning, Max Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5) (AA marks this variant deprecated)
SimpleBench	80.6%	DefaultClaude Opus 5	Independent testIndependentSimpleBench ↗	24 Jul 2026	—	AVG@5, temp 0.7; rank 8th; human baseline 83.7%; data https://simple-bench.com/static/js/leaderboard-data.js
SimpleQA Verified	59.9%	Maxclaude-opus-5_max	Independent testIndependentEpoch AI ↗	10 Aug 2026	—	Epoch-run (no tools); ±1.55 stderr
SkillsBench	60.4%	Maxeffort=max	Independent testIndependentVals.ai ↗	27 Sep 2026	OpenHands	±4.578 stderr; $2.509/test
SWE Atlas - Codebase QnA	63.2%	Extra highOpus 5 (Claude Code) xHigh	Independent testIndependentScale AI SEAL ↗	1 Oct 2026	Claude Code	rank 1 (Scale rank accounts for CI); ±5.01; entry added 2026-08-04
SWE Atlas - Codebase QnA	62.1%	MaxOpus 5 (max)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	Claude Code	AA Coding Agent Index component; pass@1 avg of 3 attempts
SWE Atlas - Test Writing	62.2%	Extra highOpus 5 (Claude Code) xHigh	Independent testIndependentScale AI SEAL ↗	1 Oct 2026	Claude Code	rank 1 (Scale rank accounts for CI); ±5.58; entry added 2026-08-04
SWE-bench Multilingual	89.5%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—
SWE-bench Multimodal	59.4%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—
SWE-Bench Pro (public, v1)	79.2%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	avg of 5 trials
SWE-Bench Pro V2 (full)	99.4%	Extra highOpus 5 (Claude Code) xhigh	Independent testIndependentScale AI SEAL ↗	1 Oct 2026	Claude Code	rank 1 (Scale rank accounts for CI); ±0.4; entry added 2026-09-22; SWE-Bench Pro V2 (642 tasks, locked protocol, released 2026-09-22)
SWE-Bench Pro V2 (hard)	98.0%	Extra highOpus 5 (Claude Code) xhigh	Independent testIndependentScale AI SEAL ↗	1 Oct 2026	Claude Code	rank 1 (Scale rank accounts for CI); ±0; entry added 2026-09-22; SWE-Bench Pro V2 (642 tasks, locked protocol, released 2026-09-22)
SWE-bench Verified	96.0%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	24 Jul 2026	—	avg of 5 trials
SWE-bench Verified	97.0%	Default	Independent testIndependentVals.ai ↗	1 Sep 2026	bash-only (single bash tool) agent	±0.764 stderr; $1.291/test; Vals archived SWE-bench Verified on 2026-09-01 (saturated)
SWE-rebench	63.4%	HighOpus 5 [high]	Independent testIndependentSWE-rebench (Nebius) ↗	1 Oct 2026	SWE-rebench standard scaffold	time window 2026-05-15..2026-07-01 (111 problems, 65 repos); ±1.35; pass@5 74.8%; $3.47/problem
Terminal-Bench 2.1	76.4%	LowClaude Opus 5 (Adaptive Reasoning, Low Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-low) (AA marks this variant deprecated)
Terminal-Bench 2.1	86.1%	MediumClaude Opus 5 (Adaptive Reasoning, Medium Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-medium) (AA marks this variant deprecated)
Terminal-Bench 2.1	87.6%	HighClaude Opus 5 (Adaptive Reasoning, High Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-high) (AA marks this variant deprecated)
Terminal-Bench 2.1	84.6%	Higheffort=high	Independent testIndependentVals.ai ↗	27 Sep 2026	Terminus 2	±0.991 stderr; $0.887/test
Terminal-Bench 2.1	88.0%	Extra highClaude Opus 5 (Adaptive Reasoning, Xhigh Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-xhigh) (AA marks this variant deprecated)
Terminal-Bench 2.1	89.1%	MaxClaude Opus 5 (Adaptive Reasoning, Max Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5) (AA marks this variant deprecated)
Terminal-Bench 3.0	41.0%	Highshipped default effort (high)	Maker's own figureVendor-reportedAnthropic ↗	1 Oct 2026	Claude Managed Agents	74 tasks, two runs per model; $28 per solved task
Terminal-Bench 3.0	42.7%	Max	Independent testIndependentSnorkel AI / Terminal-Bench ↗	1 Oct 2026	mini-SWE-agent	TB 3.0 v0.1 (74 tasks); superseded by TB 4.0; tokens 7.3B, run cost $5.8k
Terminal-Bench 4.0	34.9%	Low	Independent testIndependentTerminal-Bench ↗	1 Oct 2026	Claude Code	TB 4.0.0 (66 tasks); ±3.94 95% CI; 330 trials; total run cost $2394; model release 2026-07-24
Terminal-Bench 4.0	26.3%	LowClaude Opus 5 (Adaptive Reasoning, Low Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-low) (AA marks this variant deprecated)
Terminal-Bench 4.0	28.5%	Lowadaptive thinking, effort=low	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Claude Code (--bare)	Terminal-Bench 4.0 (66 tasks); Anthropic internal runs, 5-15 trials/task; public leaderboard reports Opus 5 at 51.8%; vendor-reported cost/task $4.25
Terminal-Bench 4.0	44.9%	Medium	Independent testIndependentTerminal-Bench ↗	1 Oct 2026	Claude Code	TB 4.0.0 (66 tasks); ±3.83 95% CI; 330 trials; total run cost $3192; model release 2026-07-24
Terminal-Bench 4.0	34.3%	MediumClaude Opus 5 (Adaptive Reasoning, Medium Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-medium) (AA marks this variant deprecated)
Terminal-Bench 4.0	41.2%	Mediumadaptive thinking, effort=med	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Claude Code (--bare)	Terminal-Bench 4.0 (66 tasks); Anthropic internal runs, 5-15 trials/task; public leaderboard reports Opus 5 at 51.8%; vendor-reported cost/task $7.0
Terminal-Bench 4.0	50.3%	High	Independent testIndependentTerminal-Bench ↗	1 Oct 2026	Claude Code	TB 4.0.0 (66 tasks); ±3.73 95% CI; 330 trials; total run cost $4662; model release 2026-07-24
Terminal-Bench 4.0	46.0%	HighClaude Opus 5 (Adaptive Reasoning, High Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-high) (AA marks this variant deprecated)
Terminal-Bench 4.0	47.0%	Highadaptive thinking, effort=high	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Claude Code (--bare)	Terminal-Bench 4.0 (66 tasks); Anthropic internal runs, 5-15 trials/task; public leaderboard reports Opus 5 at 51.8%; vendor-reported cost/task $10.64
Terminal-Bench 4.0	53.9%	Extra high	Independent testIndependentTerminal-Bench ↗	1 Oct 2026	Claude Code	TB 4.0.0 (66 tasks); ±3.17 95% CI; 330 trials; total run cost $6086; model release 2026-07-24
Terminal-Bench 4.0	46.5%	Extra highClaude Opus 5 (Adaptive Reasoning, Xhigh Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5-xhigh) (AA marks this variant deprecated)
Terminal-Bench 4.0	50.6%	Extra highadaptive thinking, effort=xhigh	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Claude Code (--bare)	Terminal-Bench 4.0 (66 tasks); Anthropic internal runs, 5-15 trials/task; public leaderboard reports Opus 5 at 51.8%; vendor-reported cost/task $13.48
Terminal-Bench 4.0	54.5%	MaxOpus 5 (max)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	Claude Code	AA Coding Agent Index component run in the vendor agent harness; pass@1 avg of 3 attempts
Terminal-Bench 4.0	53.5%	Maxeffort=max	Independent testIndependentVals.ai ↗	29 Sep 2026	mini-SWE-agent	±1.336 stderr; $18.597/test
Terminal-Bench 4.0	51.8%	Max	Independent testIndependentTerminal-Bench ↗	1 Oct 2026	Claude Code	TB 4.0.0 (66 tasks); ±3.39 95% CI; 330 trials; total run cost $5969; model release 2026-07-24
Terminal-Bench 4.0	49.0%	MaxClaude Opus 5 (Adaptive Reasoning, Max Effort)	Independent testIndependentArtificial Analysis ↗	1 Oct 2026	—	AA-run evaluation (AA slug claude-opus-5) (AA marks this variant deprecated)
Terminal-Bench 4.0	52.3%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Claude Code (--bare)	Terminal-Bench 4.0 (66 tasks); Anthropic internal runs, 5-15 trials/task; public leaderboard reports Opus 5 at 51.8%; vendor-reported cost/task $15.83
Terminal-Bench-Science 0.1	29.0%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	Claude Code (--bare)	Terminal-Bench-Science 0.1 (70 tasks); public leaderboard: 30.0%
Toolathlon	80.6%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	internal harness, Pass@1
Vals Code Migration	57.5%	Maxeffort=max	Independent testIndependentVals.ai ↗	29 Sep 2026	—	±4.371 stderr; $60.510/test
Vals CorpFin v2	73.2%	Defaultvals id anthropic/claude-opus-5	Independent testIndependentVals.ai ↗	12 Aug 2026	—	vals id anthropic/claude-opus-5; rank 1/134; ±0.871 stderr; $0.860452/test
Vals Index	63.7%	Maxeffort=max	Independent testIndependentVals.ai ↗	30 Sep 2026	—	±0.957 stderr; $19.315/test
Vals Legal Research Bench	55.3%	Maxeffort=max	Independent testIndependentVals.ai ↗	29 Sep 2026	—	vals id anthropic/claude-opus-5; rank 2/72; ±3.456 stderr; $6.762498/test
Vals Public Benefits Bench v1.1	76.9%	Maxeffort=max	Independent testIndependentVals.ai ↗	29 Sep 2026	—	vals id anthropic/claude-opus-5; rank 1/45; ±1.096 stderr; $3.949099/test
Vals SRE Bench	12.2%	Maxeffort=max	Independent testIndependentVals.ai ↗	29 Sep 2026	—	±2.027 stderr; $23.633/test
Vals TaxEval v2	75.1%	Defaultvals id anthropic/claude-opus-5	Independent testIndependentVals.ai ↗	1 Sep 2026	—	vals id anthropic/claude-opus-5; rank 23/145; ±0.832 stderr; $0.114274/test
Vals Vibe Code Bench	88.4%	Default	Independent testIndependentVals.ai ↗	29 Sep 2026	OpenHands	±3.001 stderr; $33.876/test
Vending-Bench 2	$11,182	Default	Independent testIndependentAndon Labs ↗	1 Oct 2026	—	final money balance after simulated year, arithmetic mean across runs; ±$2,094; rank 4; only top 10 rendered server-side (57 more behind 'Show more')
WANDR	50.5%	Lowadaptive thinking, effort=low	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	WANDR (Perplexity wide-search benchmark), soft F1; Anthropic-modified setup (offline search/fetch, 980k token budget), not comparable to Perplexity numbers; vendor-reported cost/task $10.65
WANDR	58.1%	Mediumadaptive thinking, effort=med	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	WANDR (Perplexity wide-search benchmark), soft F1; Anthropic-modified setup (offline search/fetch, 980k token budget), not comparable to Perplexity numbers; vendor-reported cost/task $23.97
WANDR	64.6%	Highadaptive thinking, effort=high	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	WANDR (Perplexity wide-search benchmark), soft F1; Anthropic-modified setup (offline search/fetch, 980k token budget), not comparable to Perplexity numbers; vendor-reported cost/task $43.49
WANDR	67.0%	Extra highadaptive thinking, effort=xhigh	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	WANDR (Perplexity wide-search benchmark), soft F1; Anthropic-modified setup (offline search/fetch, 980k token budget), not comparable to Perplexity numbers; vendor-reported cost/task $53.84
WANDR	67.2%	Maxadaptive thinking, effort=max	Maker's own figureVendor-reportedAnthropic ↗	22 Sep 2026	—	WANDR (Perplexity wide-search benchmark), soft F1; Anthropic-modified setup (offline search/fetch, 980k token budget), not comparable to Perplexity numbers; vendor-reported cost/task $61.62