Security AI model leaderboards

dfbench by depthfirst

• depthfirst’s dfbench tests 12 models on finding real vulnerabilities in code they have not seen.
• Mythos 5 (run with Claude Security) leads with 69.0%, 3.3 points ahead of GPT-5.6 Sol.
• depthfirst’s specialized dfs-large1 is third at 62.2%, ahead of 9 general-purpose models.

PWNBench v0.1 by Novee

• Novee’s PWNBench v0.1 tests 14 models on greybox penetration testing of running web applications.
• Grok 4.6 leads with 63.3%, 3.3 points ahead of Claude Opus 5.
• The best open model, DeepSeek V4.1 Flash, is fifth at 49.2%, 14.1 points behind first.

CyberGym

• CyberGym tests eight models on reproducing 1,507 real C/C++ vulnerabilities from their advisories.
• DeepSeek V4.1 Flash (1017) (95.4%) and DeepSeek V4.1 Flash (94.8%) are nearly tied.
• Open models take the top five places; GPT-5.6 Sol and Claude Opus 5 rejected the tasks and have no score.

Legal AI model leaderboards

LAB by Harvey

• Harvey’s LAB tests 12 models on the legal work a first-year associate would get.
• Grok 4.6 leads with 22.1%, 7.4 points ahead of GLM 5.3.
• GLM 5.3 and Kimi K3, both open models, rank second and third, ahead of every OpenAI and Anthropic model.

LAB: Contracts by Harvey

• Harvey’s LAB: Contracts tests 13 models on contract review, redlining, and issue-spotting on real deal documents.
• Grok 4.6 leads with 10.0%, 1.7 points ahead of Kimi K3.
• Kimi K3 is the best open model and finishes ahead of every OpenAI, Anthropic, and Google model.

Legora BAR by Legora

• Legora BAR tests 17 models on end-to-end legal matters across 30 practice areas inside Legora’s agent harness.
• Claude Opus 5.5 leads with +6% vs the model average, 1 point ahead of Muse Wonder 1.0, which costs 80% less per case.
• The best open model, Kimi K3, is tied for sixth at +2% vs the model average, 4 points behind first, for 80% less per case.

DAYJOB: Law by Surge AI

• Surge AI’s DAYJOB: Law tests 15 models on legal work.
• GPT-6 Astra (36.9%) and Muse Wonder 1.0 (36.6%) are nearly tied.
• The best open model, Qwen 3.8 Max, is ninth at 16.9%, 20.0 points behind first.

CorporateBench by Epiq

• Epiq’s CorporateBench tests ten models on legal work.
• GPT-6 Astra leads with 82.0%, 3.0 points ahead of GPT-6 Sol, which costs 76% less per task.
• The best open model, DeepSeek V4.1 Flash, is fifth at 75.3%, 6.7 points behind first, for 97% less per task.

APEX Agents: Corporate Law by Mercor

• Mercor’s APEX Agents: Corporate Law tests 12 models on multi-step corporate-law assignments graded by practicing attorneys.
• GPT-6 Astra leads with 87.7%, 2.3 points ahead of Claude Opus 5.5.
• The best open model, GLM 5.3, is seventh at 81.4%, 6.3 points behind first, for 44% less per task.

Redline Bench by Crosby

• Crosby’s Redline Bench tests 11 models on negotiating a contract over four rounds of tracked changes.
• GPT-6 Astra leads with 62.8%, 8.1 points ahead of Grok 4.6.
• The best open model, Muse Spark Open, is sixth at 49.9%, 12.9 points behind first.

Finance AI model leaderboards

Big Finance Benchmark by Rogo

• Rogo’s Big Finance Benchmark tests 14 models on financial research tasks, scored on final-answer accuracy.
• Gemini 4 Argon (54.7%) and GPT-6 Astra (54.6%) are nearly tied.
• The best open model, Muse Spark Open, is third at 53.3%, 1.4 points behind first, for 29% less per task.

Financial Services Benchmark by Hebbia

• Hebbia’s Financial Services Benchmark tests eight models on finance work.
• Claude Opus 5.5 leads with 74.6%, 1.1 points ahead of GPT-6.1 Sol.
• The best open model, DeepSeek V4.1 Flash, is fourth at 63.1%, 11.5 points behind first.

DAYJOB: Finance by Surge AI

• Surge AI’s DAYJOB: Finance tests 15 models on finance work.
• Claude Opus 5.5 leads with 42.5%, 2.0 points ahead of Claude Sonnet 5.5.
• The best open model, DeepSeek V4.1 Flash (1017), is 14th at 12.5%, 30.0 points behind first.

ATLAS-Finance by Handshake

• Handshake’s ATLAS-Finance tests 11 models on finance work.
• The top three models are within 0.9 points of each other, led by Claude Opus 5 at 12.3%.
• The best open model, Kimi K3, is tied for sixth at 4.7%, 7.6 points behind first, for 33% less per task.

FrontierFinance by Samaya AI

• Samaya AI’s FrontierFinance tests 12 models on 220 expert-written investment research questions, from screening to earnings.
• Claude Fable 5.1 leads with 55.9%, 3.3 points ahead of Muse Wonder 1.0, which costs 51% less per task.
• The best open model, Kimi K3, is seventh at 46.4%, 9.5 points behind first, for 87% less per task.

SheetBench by HUD

• HUD’s SheetBench tests eight models on finance work.
• GPT-6 Astra leads with 64.7%, 2.2 points ahead of Claude Opus 5.5.
• Kimi K3 and DeepSeek V4.1 Flash, both open models, rank fourth and fifth, ahead of every Google and xAI model.

DAYJOB: Accounting by Surge AI

• Surge AI’s DAYJOB: Accounting tests 15 models on finance work.
• Claude Sonnet 5.5 (35.2%) and Claude Opus 5.5 (34.4%) are nearly tied.
• The best open model, GLM 5.3, is tied for 11th at 7.6%, 27.6 points behind first.

DAYJOB: Tax by Surge AI

• Surge AI’s DAYJOB: Tax tests 16 models on finance work.
• Claude Sonnet 5.5 leads with 10.8%, 1.2 points ahead of Claude Fable 5.1.
• The best open model, DeepSeek V4.1 Flash (1017), is 11th at 1.2%, 9.6 points behind first, for 95% less per task.

TaxCalcBench by Column Tax

• Column Tax’s TaxCalcBench tests 14 models on finance work.
• Muse Wonder 1.0 leads with 77.5%, 5.5 points ahead of GPT-6 Sol, which costs 88% less per task.
• The best open model, Kimi K3, is eighth at 38.0%, 39.5 points behind first.

Web Search AI model leaderboards

Parallel Search Capability Leaderboard by Parallel

• Parallel Search Capability Leaderboard tests 16 models on web search work.
• Claude Opus 5.5 leads with 75.4%, 1.2 points ahead of Kimi K3.1 Pro, which costs 72% less per task.
• The best open model, DeepSeek V4.1 Flash (1017), is 11th at 68.9%, 6.5 points behind first, for 87% less per task.

ATLAS by Exa

• Exa’s ATLAS tests eight models on building complete tables of every entity that fits a request, searching with Exa.
• GPT-6.1 Sol leads with 69.5%, 15.7 points ahead of Gemini 4 Argon.
• The best open model, Kimi K3, is fifth at 42.1%, 27.4 points behind first.

Web Context by Nimble

• Nimble’s Web Context tests 11 models on web search work.
• GPT-5.6 Sol leads with 80.0%, 2.7 points ahead of Qwen 3.8 Max.

WARC-Bench by Uniphore

• Uniphore’s WARC-Bench tests eight models on short-horizon web subtasks on web-archive-based real websites, such as date pickers, menus, and scrollable containers. The web observation is represented via screenshots alone.
• Claude Opus 5.5 leads with 90.3%, 1.2 points ahead of GPT-6 Astra.
• The best open model, Kimi K3, is fourth at 85.9%, 4.4 points behind first, for 30% less per task.

Customer Support AI model leaderboards

τ³-Banking by Sierra

• Sierra’s τ³-Banking tests 18 models on banking support requests that must follow bank policy.
• Gemini 4 Argon leads with 58.2%, 2.2 points ahead of Claude Opus 5.5.
• GLM 5.3 and Kimi K3, both open models, rank third and fourth, ahead of every OpenAI and xAI model.

DuetBench–Diagnosis by Decagon

• Decagon’s DuetBench–Diagnosis tests nine models on diagnosing customer issues in real support conversations.
• Claude Opus 5 (1,035 Elo) and GLM 5.3 (1,032 Elo) are nearly tied, and GLM 5.3 costs 77% less per task.
• Open models take three of the top five places.

Cresta-Bench by Cresta

• Cresta-Bench tests 14 models on taking simulated calls as a contact-center agent for real-estate lending, senior health, a health-plan provider line, and airline support.
• Claude Opus 5.5 leads with 82.6%, 4.6 points ahead of Claude Sonnet 5.5, which costs 59% less per test run.
• The best open model, Qwen 3.8 Max, is eighth at 48.7%, 33.9 points behind first.

τ³-Voice by Sierra

• Sierra’s τ³-Voice tests nine models on live voice support calls in retail, airline, and telecom.
• gpt-live-1 leads with 81.7%, 1.3 points ahead of Ava Zero.

vCX Benchmark by Retell AI

• Retell AI’s vCX Benchmark tests 15 models on following instructions, staying grounded, calling tools, and speaking naturally as a voice agent.
• The top three models are within 0.8 points of each other, led by GPT-6.1 Sol at 93.4%.

Software AI model leaderboards

FrontierSWE V2 by Proximal

• Proximal’s FrontierSWE V2 tests 22 models on 34 expert-level software projects of up to 20 hours each.
• GPT-6 Astra leads with 65.5%, 3.2 points ahead of Claude Opus 5.5, which costs 90% less per trial.
• The best open model, GLM 5.3, is tenth at 30.2%, 35.3 points behind first.

Legacy-Bench by Factory

• Factory’s Legacy-Bench tests 12 models on maintaining enterprise code in COBOL, Fortran, Java 7, and other legacy languages.
• Claude Opus 5.5 and Muse Spark 1.3 tie for first at 62.0%, and Claude Opus 5.5 costs 75% less per task.
• DeepSeek V4.1 Flash and DeepSeek V4.1 Flash (1017), both open models, rank third and fourth, ahead of every OpenAI, Google, and xAI model.

sudo L7 by Surge AI

• Surge AI’s sudo L7 tests 16 models on software work.
• Claude Sonnet 5.5 leads with 50.6%, 5.0 points ahead of Claude Opus 5.5.
• The best open model, Qwen 3.8 Max, is tied for sixth at 16.7%, 33.9 points behind first.

Senior SWE-Bench by Snorkel AI

• Snorkel AI’s Senior SWE-Bench tests seven models on software work.
• GPT-6 Astra leads the ranked models at 36.8%; Claude Opus 5.5 (39.3%, reported by Anthropic) is shown but not ranked.
• The best open model, Kimi K3, is sixth at 20.2%, 16.6 points behind first.

SWE-Clar by Upwork

• Upwork’s SWE-Clar tests nine models on building greenfield software from real Upwork jobs after clarifying incomplete client requests.
• GPT-6 Astra leads with 67.0%, 4.0 points ahead of Claude Opus 5.
• The best open model, Qwen 3.8 Max, is fifth at 55.0%, 12.0 points behind first.

SWE-Bench Pro V2 by Scale AI

• Scale AI’s SWE-Bench Pro V2 tests 16 models on resolving 642 real issues in professional open-source repositories.
• The top three models are within 0.3 points of each other, led by Claude Opus 5 at 99.4%.
• The best open model, Kimi K3, is fourth at 97.7%, 1.7 points behind first.

APEX-SWE by Mercor

• Mercor’s APEX-SWE tests 11 models on repository-scale engineering tasks graded against a maintainer’s patch.
• Claude Opus 5.5 leads with 67.6%, 3.8 points ahead of Claude Opus 5, which costs 79% less per task.
• The best open model, DeepSeek V4.1 Flash, is fifth at 51.9%, 15.7 points behind first.

DeepSWE v1.1

• DeepSWE v1.1 tests 16 models on debugging failures whose cause sits several layers from the symptom.
• Claude Opus 5 leads with 72.3%, 1.3 points ahead of GPT-6 Astra, which costs 59% less per task.
• Kimi K3 and DeepSeek V4.1 Flash (1017), both open models, rank third and fourth, ahead of every Google and xAI model.

ViBench 2.0 by Replit

• Replit’s ViBench 2.0 tests nine models on software work.
• GPT-6.1 Sol leads with 64.7%, 4.4 points ahead of GPT-6 Astra.
• The best open model, DeepSeek V4.1 Flash, is tied for sixth at 0.0%, 64.7 points behind first.

DesignArena: Full-Stack by DesignArena

• DesignArena: Full-Stack tests 15 models on software work.
• GPT-6 Astra leads with 1,357 Elo, 31 Elo points ahead of Claude Fable 5.1.
• The best open model, Kimi K3, is third at 1,315 Elo, 42 Elo points behind first.

DesignArena: Agentic Game Dev by DesignArena

• DesignArena: Agentic Game Dev tests 15 models on software work.
• Claude Fable 5 leads with 1,248 Elo, 16 Elo points ahead of GPT-5.6 Sol.
• The best open model, Kimi K3, is fourth at 1,218 Elo, 30 Elo points behind first.

MacroscopeBench by Macroscope

• MacroscopeBench tests 12 models on catching bugs in real code changes without false reports.
• Claude Opus 5.5 leads with 83.9%, 3.6 points ahead of GPT-6 Sol, which costs 68% less per task.
• The best open model, GLM 5.3, is sixth at 77.4%, 6.5 points behind first, for 67% less per task.

AI Code Review Benchmark by Qodo

• Qodo’s AI Code Review Benchmark tests 11 models on software work.
• Claude Opus 5.5 leads with 56.8%, 2.8 points ahead of Claude Sonnet 5.5.
• GLM 5.3, DeepSeek V4 Pro 0813, and Kimi K3, all open models, rank fourth, fifth, and sixth, ahead of every OpenAI and Google model.

ORCA Benchmark by Traversal

• Traversal’s ORCA Benchmark tests 12 models on finding the root cause of live production incidents.
• GPT-6 Astra leads with 50.2%, 5.9 points ahead of Claude Opus 5.
• The best open model, Kimi K3, is fourth at 31.2%, 19.0 points behind first.

SecIT Bench by Cribl

• Cribl’s SecIT Bench tests 17 models on software work.
• DeepSeek V4.1 Flash leads with 82.9%, 1.5 points ahead of GLM 5.3.
• Open models take two of the top five places.

o11y-bench by Grafana Labs

• Grafana Labs’ o11y-bench tests 18 models on querying metrics, logs, and traces and investigating incidents in a live Grafana stack.
• Claude Opus 5.5 leads with 87.5%, 1.6 points ahead of GPT-6.1 Sol, which costs 97% less per trial.
• The best open model, Kimi K3, is tied for sixth at 81.3%, 6.2 points behind first, for 89% less per trial.

ML Research AI model leaderboards

FrontierMemory by Letta

• Letta’s FrontierMemory tests seven models on writing memory from past agent runs that helps a task agent solve new coding and knowledge-work tasks.
• Claude Opus 5.5 and GLM 5.3 tie for first at 73.3%, and Claude Opus 5.5 costs 22% less per task.
• GLM 5.3 is the best open model and finishes ahead of every OpenAI, Google, and xAI model.

DAYJOB: Data Science by Surge AI

• Surge AI’s DAYJOB: Data Science tests 15 models on ml research work.
• Claude Opus 5.5 leads with 37.3%, 2.0 points ahead of Claude Sonnet 5.5.
• The best open model, Qwen 3.8 Max, is eighth at 7.3%, 30.0 points behind first.

WitnessBench by Fleet

• Fleet’s WitnessBench tests 13 models on working out an unfamiliar puzzle game’s rules by playing it, then applying them to harder levels.
• Claude Opus 5.5 leads with 46.5 out of 100, 20.9 points ahead of GPT-6 Astra Pro.
• The best open model, Kimi K3, is fourth at 18.8 out of 100, 27.7 points behind first.

Healthcare AI model leaderboards

Bedside Bench by Doximity

• Doximity’s Bedside Bench tests 15 models on 500 physician-written clinical cases on medical reasoning and safety.
• Claude Opus 5 leads the ranked models at 94.6%; Doximity Ask V7 (95.4%, reported by Doximity) is shown but not ranked.
• The best open model, Muse Spark Open, is second at 94.3%, 0.3 points behind first, for 86% less per task.

RCM-Bench-1 by Commure

• Commure’s RCM-Bench-1 tests 20 models on 97 revenue cycle tasks such as claims, billing, and eligibility checks.
• Claude Opus 5 leads with 92.7%, 1.2 points ahead of Gemini 4 Argon.
• The best open model, DeepSeek V4.1 Flash, is ninth at 89.0%, 3.7 points behind first, for 93% less per task.

DAYJOB: Healthcare by Surge AI

• Surge AI’s DAYJOB: Healthcare tests 15 models on healthcare work.
• Claude Opus 5.5 leads with 48.3%, 14.3 points ahead of Claude Sonnet 5.5.

APEX-1: General Practitioner by Mercor

• Mercor’s APEX-1: General Practitioner tests 12 models on primary-care consults graded by practicing GPs.
• Claude Opus 5.5 (73.0%) and Claude Opus 5 (72.1%) are nearly tied, and Claude Opus 5 costs 65% less per task.
• The best open model, Kimi K3, is fourth at 64.3%, 8.7 points behind first, for 88% less per task.

HealthBench Professional

• HealthBench Professional tests 13 models on 525 clinical consults graded against physician-written rubrics.
• GPT-6 Astra leads with 64.4%, 5.6 points ahead of GPT-5.6 Sol.
• The best open model, Muse Spark Open, is third at 57.3%, 7.1 points behind first, for 67% less per task.

Life Sciences AI model leaderboards

BioBench by LatchBio

• LatchBio’s BioBench tests 13 models on end-to-end analysis of real biology data, from single-cell omics to drug discovery and oncology.
• Claude Opus 5.5 leads with 58.0%, 5.0 points ahead of Claude Fable 5.1.
• The best open model, DeepSeek V4.1 Flash, is eighth at 42.1%, 15.9 points behind first.

VIALS by Handshake

• Handshake’s VIALS tests 13 models on life sciences work.
• GPT-6 Astra (49.9%) and GPT-6.1 Sol (49.1%) are nearly tied, and GPT-6.1 Sol costs 83% less per task.
• The best open model, Kimi K3, is 11th at 20.9%, 29.0 points behind first.

BioBench-Refusal by LatchBio

• LatchBio’s BioBench-Refusal tests 12 models on refusing harmful biology requests disguised as research while still completing routine dual-use work.
• Grok 4.7 leads with 62.4%, 13.7 points ahead of Gemini 4 Argon.
• The best open model, DeepSeek V4.1 Flash, is sixth at 11.3%, 51.1 points behind first.

Productivity AI model leaderboards

Genspark Slides Benchmark by Genspark

• The Genspark Slides Benchmark tests 11 models on building finished slide decks from 200 real production tasks.
• Claude Fable 5.1 leads with 83.6%, 1.4 points ahead of Genspark’s specialized Gen-1 Slides, which costs 92% less per deck.
• The best open model, Kimi K3, is sixth at 72.6%, 11.0 points behind first.

DesignArena: Python-PPTX Slides by DesignArena

• DesignArena: Python-PPTX Slides tests 15 models on productivity work.
• GPT-6 Astra leads with 1,414 Elo, 96 Elo points ahead of Claude Fable 5.1.
• The best open model, Kimi K3, is fifth at 1,219 Elo, 195 Elo points behind first.

Business Operations AI model leaderboards

AutomationBench by Zapier

• Zapier’s AutomationBench tests 15 models on business operations work.
• Gemini 4 Argon leads with 51.3%, 6.6 points ahead of Claude Sonnet 5.5.
• The best open model, Kimi K3, is 11th at 22.7%, 28.6 points behind first.

Clipping Bench by Naïve

• Naïve’s Clipping Bench tests 19 models on running a short-form clip farm for a simulated month.
• GPT-6 Astra leads with $9,514, $3,807 ahead of GPT-6.1 Sol, which costs 81% less per season.
• The best open model, GLM 5.3, is ninth at $644, $8,870 behind first.

Short Form Bench by Naïve

• Naïve’s Short Form Bench tests nine models on running a small brand’s short-form video business for a simulated week.
• GPT-6.1 Sol leads with $2,949, $146 ahead of GPT-6 Astra.

Marketing AI model leaderboards

AIM Routing by Zeta

• Zeta’s AIM Routing tests seven models on picking exactly the tools a marketer’s analytics request needs.
• AIM-R1 leads with 94.7%, 1.9 points ahead of Zeta’s specialized AIM-R2.
• The best open model, GLM 5.3 Flash, is sixth at 87.2%, 7.5 points behind first, for 79% less per 1,000 requests.

AIM Grounding by Zeta

• Zeta’s AIM Grounding tests eight models on summarizing customer reports, with each statement supported by claims grounded in the data.
• The top three models are within 0.2 points of each other, led by GPT-6 Sol at 96.3%.
• Zeta’s specialized AIM-S3 is second at 96.2%, for 94% less per 1,000 requests than GPT-6 Sol.

AIM Consistency by Zeta

• Zeta’s AIM Consistency tests eight models on summarizing customer reports, assessing whether statements within each summary contradict one another.
• The top three models are within 0.6 points of each other, led by GPT-6 Sol at 98.9%.
• Zeta’s specialized AIM-S3 is third at 98.3%, for 94% less per 1,000 requests than GPT-6 Sol.

Hardware AI model leaderboards

CAD Arena by Normal Research

• Normal Research’s CAD Arena tests 12 models on rebuilding engineering drawings as editable CAD models.
• Claude Opus 5.5 leads with 74.6 out of 100, 7.5 points ahead of GPT-6 Astra, which costs 80% less per task.

DataCenterBench Tier 1 by Robocurve

• Robocurve’s DataCenterBench Tier 1 tests nine models on physical data-center maintenance with a robot, such as connecting cables and operating controls.
• GPT-6 Astra leads with 40.0%, 1.7 points ahead of GPT-6.1 Sol, which costs 87% less per task.
• The best open model, GLM 5.3 Flash, is tied for sixth at 15.0%, 25.0 points behind first.

Chip Design AI model leaderboards

Build-A-Chip-Bench by ChipAgents

• ChipAgents’ Build-A-Chip-Bench tests eight models on taking a chip design from a workload specification to a routed, verified layout.
• Claude Fable 5.1 leads with 23.5%, 4.9 points ahead of Gemini 4 Argon, which costs 60% less per task.
• The best open model, GLM 5.3, is fourth at 2.2%, 21.3 points behind first.