Overall leaderboard

Which current model does best across everything we publish? Each benchmark becomes a percentile among these models, then groups are weighted towards real business work. Hover, tap or tab to a bar to see what it is made of. For launches and pace, see models over time.

Models ranked: 23
Sources in the score: 8 licensed sources + 5 of our own benchmarks
Release: 13 Oct 2026
Score: 0–100, higher is better

Spring Prompt benches 35%
Agents and tool use 20%
Knowledge work 15%
Writing and preference 15%
Factuality 15%

Our own data: CatalogBench, ROASBench and BulletBench are in Spring Prompt benches (35%); DeckBench and SurveyBench are in Knowledge work (15%). Across these 23 models, 48% of a score comes from our own benchmarks on average (from 35% to 100%, depending on which results a model has).

Overall score from 0 to 100, higher is better, with each group's score, blended price and the date we last tested it
#	Model	Overall	Spring Prompt benches	Agents and tool use	Knowledge work	Writing and preference	Factuality	Price $ / M tokens	We last tested
1	Claude Opus 5.5Anthropic	85	73	93	94	90	88	$8.00	8 Oct 2026
Spring Prompt benches 73CatalogBench 69 · Reliably publish-ready 55.4%; Sales brief: reliably publish-ready 21.4%ROASBench 79 · Overall score 51.7 Agents and tool use 93Terminal-Bench (Artificial Analysis) 93 · Terminal-Bench 4.0 59.6% Knowledge work 94APEX-Agents 91 · Tasks passed 73.5%GDP.pdf 100 · Professional document tasks 30.6%DeckBench 84 · Deck rating 1,380 (95% range 1,278–1,507)SurveyBench 100 · Analysis rating 1,499 (95% range 1,449–1,590) Writing and preference 90Arena text 100 · Overall 1,504 (95% range 1,495–1,513)UGI Leaderboard 80 · Writing score 70.0 Factuality 88Arena factuality 100 · Overall 1,507 (95% range 1,499–1,515)SimpleQA Verified (Epoch AI) 77 · Correct answers 72.2%
2	GPT-6.1 SolOpenAI	79	93	85	73	59	67	$4.00	9 Oct 2026
Spring Prompt benches 93CatalogBench 95 · Reliably publish-ready 75.0%; Sales brief: reliably publish-ready 67.9%BulletBench 88 · Lightning 10+1: ladder Elo 584 (95% range 327–806); Bullet 60s: ladder Elo 928 (95% range 676–1,184) Agents and tool use 85Terminal-Bench (Artificial Analysis) 85 · Terminal-Bench 4.0 56.1% Knowledge work 73APEX-Agents 55 · Tasks passed 60.0%DeckBench 89 · Deck rating 1,411 (95% range 1,300–1,550)SurveyBench 75 · Analysis rating 1,431 (95% range 1,361–1,522) Writing and preference 59Arena text 68 · Overall 1,483 (95% range 1,473–1,494)UGI Leaderboard 50 · Writing score 67.7 Factuality 67Arena factuality 42 · Overall 1,463 (95% range 1,454–1,472)SimpleQA Verified (Epoch AI) 92 · Correct answers 73.9%
3	GPT-6 AstraOpenAI	78	87	79	86	61	66	$20.00	6 Oct 2026
Spring Prompt benches 87CatalogBench 99 · Reliably publish-ready 75.0%; Sales brief: reliably publish-ready 69.6%ROASBench 100 · Overall score 56.6BulletBench 38 · Ladder Elo 323 (95% range 0–528) Agents and tool use 79Terminal-Bench (Artificial Analysis) 92 · Terminal-Bench 4.0 59.6%; Terminal-Bench 2.1 89.9%τ-bench banking (Artificial Analysis) 45 · Τ-bench banking 43.1%Vending-Bench 2 100 · Money after a year $15,515 Knowledge work 86APEX-Agents 73 · Tasks passed 64.7%DeckBench 100 · Deck rating 1,492 (95% range 1,372–1,710) Writing and preference 61Arena text 53 · Overall 1,477 (95% range 1,470–1,484)UGI Leaderboard 70 · Writing score 68.9 Factuality 66Vectara Hallucination Leaderboard 50 · Hallucination rate 8.7%Arena factuality 47 · Overall 1,466 (95% range 1,460–1,472)SimpleQA Verified (Epoch AI) 100 · Correct answers 75.6%
4	Muse WonderMeta	70	68	–	75	–	–	$8.00	12 Oct 2026
Spring Prompt benches 68CatalogBench 85 · Reliably publish-ready 66.1%; Sales brief: reliably publish-ready 51.8%ROASBench 53 · Overall score 38.9BulletBench 46 · Ladder Elo 409 (95% range 100–609) Knowledge work 75DeckBench 63 · Deck rating 1,273 (95% range 1,175–1,385)SurveyBench 88 · Analysis rating 1,460 (95% range 1,400–1,545) No results: Agents and tool use, Writing and preference, Factuality
5	GPT-6 SolOpenAI	69	85	81	63	31	58	$4.00	6 Oct 2026
Spring Prompt benches 85CatalogBench 90 · Reliably publish-ready 66.1%; Sales brief: reliably publish-ready 67.9%ROASBench 95 · Overall score 54.1BulletBench 58 · Ladder Elo 445 (95% range 215–639) Agents and tool use 81Terminal-Bench (Artificial Analysis) 75 · Terminal-Bench 4.0 43.9%Vending-Bench 2 88 · Money after a year $14,428 Knowledge work 63APEX-Agents 27 · Tasks passed 54.3%GDP.pdf 67 · Professional document tasks 26.4%DeckBench 95 · Deck rating 1,464 (95% range 1,356–1,637) Writing and preference 31Arena text 32 · Overall 1,457 (95% range 1,450–1,464)UGI Leaderboard 30 · Writing score 55.6 Factuality 58Vectara Hallucination Leaderboard 100 · Hallucination rate 6.5%Arena factuality 21 · Overall 1,434 (95% range 1,428–1,440)SimpleQA Verified (Epoch AI) 54 · Correct answers 60.7%
6	Claude Fable 5.1Anthropic	66	47	75	80	67	82	$20.00	6 Oct 2026
Spring Prompt benches 47CatalogBench 31 · Reliably publish-ready 26.8%; Sales brief: reliably publish-ready 5.4%ROASBench 74 · Overall score 51.5 Agents and tool use 75Terminal-Bench (Artificial Analysis) 90 · Terminal-Bench 4.0 55.1%; Terminal-Bench 2.1 91.4%τ-bench banking (Artificial Analysis) 73 · Τ-bench banking 47.2%Vending-Bench 2 63 · Money after a year $5,422 Knowledge work 80APEX-Agents 82 · Tasks passed 68.6%GDP.pdf 89 · Professional document tasks 29.6%DeckBench 68 · Deck rating 1,278 (95% range 1,166–1,425) Writing and preference 67Arena text 95 · Overall 1,501 (95% range 1,495–1,508)UGI Leaderboard 40 · Writing score 63.9 Factuality 82Arena factuality 95 · Overall 1,500 (95% range 1,494–1,505)SimpleQA Verified (Epoch AI) 69 · Correct answers 70.8%
7	Muse Spark 1.3Meta	64	41	81	53	84	89	$2.00	6 Oct 2026
Spring Prompt benches 41CatalogBench 56 · Reliably publish-ready 48.2%; Sales brief: reliably publish-ready 10.7%ROASBench 37 · Overall score 34.1BulletBench 8 · Ladder Elo 33 (95% range 0–270) Agents and tool use 81Terminal-Bench (Artificial Analysis) 62 · Terminal-Bench 4.0 33.3%; Terminal-Bench 2.1 85.4%τ-bench banking (Artificial Analysis) 100 · Τ-bench banking 50.5% Knowledge work 53APEX-Agents 45 · Tasks passed 57.8%GDP.pdf 78 · Professional document tasks 27.6%DeckBench 37 · Deck rating 879 (95% range 785–954) Writing and preference 84Arena text 84 · Overall 1,494 (95% range 1,488–1,500) Factuality 89Arena factuality 89 · Overall 1,480 (95% range 1,475–1,485)
8	Claude Sonnet 5.5Anthropic	63	49	100	89	42	43	$4.00	6 Oct 2026
Spring Prompt benches 49CatalogBench 44 · Reliably publish-ready 32.1%; Sales brief: reliably publish-ready 10.7%ROASBench 84 · Overall score 52.0BulletBench 8 · Ladder Elo 33 (95% range 0–270) Agents and tool use 100Terminal-Bench (Artificial Analysis) 100 · Terminal-Bench 4.0 63.6% Knowledge work 89APEX-Agents 100 · Tasks passed 75.5%DeckBench 79 · Deck rating 1,362 (95% range 1,240–1,518) Writing and preference 42Arena text 42 · Overall 1,471 (95% range 1,461–1,481) Factuality 43Arena factuality 63 · Overall 1,470 (95% range 1,461–1,479)SimpleQA Verified (Epoch AI) 23 · Correct answers 46.5%
9	Gemini 3.8 FlashGoogle	61	52	49	50	95	73	$3.00	6 Oct 2026
Spring Prompt benches 52CatalogBench 36 · Reliably publish-ready 35.7%; Sales brief: reliably publish-ready 3.6%ROASBench 68 · Overall score 51.4BulletBench 67 · Lightning 10+1: ladder Elo 197 (95% range 0–439); Bullet 60s: ladder Elo 1,068 (95% range 888–1,241) Agents and tool use 49Terminal-Bench (Artificial Analysis) 54 · Terminal-Bench 4.0 19.7%; Terminal-Bench 2.1 87.6%τ-bench banking (Artificial Analysis) 55 · Τ-bench banking 45.8%Vending-Bench 2 38 · Money after a year $5,094 Knowledge work 50APEX-Agents 64 · Tasks passed 64.3%GDP.pdf 56 · Professional document tasks 23.4%DeckBench 32 · Deck rating 816 (95% range 722–897) Writing and preference 95Arena text 89 · Overall 1,495 (95% range 1,490–1,500)UGI Leaderboard 100 · Writing score 78.6 Factuality 73Arena factuality 84 · Overall 1,477 (95% range 1,472–1,481)SimpleQA Verified (Epoch AI) 62 · Correct answers 69.7%
10	DeepSeek-V4.1-FlashDeepSeek	55	61	48	–	47	58	$0.52	30 Sep 2026
Spring Prompt benches 61CatalogBench 61 · Reliably publish-ready 41.1%; Sales brief: reliably publish-ready 23.2% Agents and tool use 48Terminal-Bench (Artificial Analysis) 48 · Terminal-Bench 4.0 26.8% Writing and preference 47Arena text 47 · Overall 1,474 (95% range 1,468–1,481) Factuality 58Arena factuality 58 · Overall 1,467 (95% range 1,461–1,473) No results: Knowledge work
11=	Kimi K3Moonshot AI	49	38	51	31	79	63	$6.00	6 Oct 2026
Spring Prompt benches 38CatalogBench 35 · Reliably publish-ready 41.1%; Sales brief: reliably publish-ready 1.8%ROASBench 63 · Overall score 44.1BulletBench 8 · Ladder Elo 33 (95% range 0–270) Agents and tool use 51Terminal-Bench (Artificial Analysis) 39 · Terminal-Bench 4.0 12.6%; Terminal-Bench 2.1 85.0%τ-bench banking (Artificial Analysis) 64 · Τ-bench banking 46.0%Vending-Bench 2 50 · Money after a year $5,165 Knowledge work 31APEX-Agents 18 · Tasks passed 50.6%GDP.pdf 22 · Professional document tasks 19.0%DeckBench 53 · Deck rating 1,230 (95% range 1,127–1,366) Writing and preference 79Arena text 79 · Overall 1,488 (95% range 1,484–1,493) Factuality 63Arena factuality 79 · Overall 1,472 (95% range 1,469–1,476)SimpleQA Verified (Epoch AI) 46 · Correct answers 50.6%
11=	Gemini 3.1 Pro PreviewGoogle	49	61	22	19	82	53	$4.50	8 Oct 2026
Spring Prompt benches 61CatalogBench 60 · Reliably publish-ready 44.6%; Sales brief: reliably publish-ready 16.1%ROASBench 58 · Overall score 43.5BulletBench 69 · Ladder Elo 529 (95% range 329–698) Agents and tool use 22Terminal-Bench (Artificial Analysis) 26 · Terminal-Bench 4.0 4.0%; Terminal-Bench 2.1 73.8%τ-bench banking (Artificial Analysis) 27 · Τ-bench banking 21.4%Vending-Bench 2 13 · Money after a year $911.21 Knowledge work 19APEX-Agents 0 · Tasks passed 35.3%GDP.pdf 11 · Professional document tasks 17.0%DeckBench 26 · Deck rating 775 (95% range 671–849)SurveyBench 38 · Analysis rating 792 (95% range 717–849) Writing and preference 82Arena text 74 · Overall 1,487 (95% range 1,484–1,490)UGI Leaderboard 90 · Writing score 72.2 Factuality 53Vectara Hallucination Leaderboard 0 · Hallucination rate 10.4%Arena factuality 74 · Overall 1,472 (95% range 1,470–1,474)SimpleQA Verified (Epoch AI) 85 · Correct answers 73.5%
13	Qwen3.8-Max (0902)Alibaba	47	22	78	43	63	53	$3.00	6 Oct 2026
Spring Prompt benches 22CatalogBench 20 · Reliably publish-ready 28.6%; Sales brief: reliably publish-ready 0.0%ROASBench 26 · Overall score 25.4 Agents and tool use 78Terminal-Bench (Artificial Analysis) 73 · Terminal-Bench 4.0 38.9%; Terminal-Bench 2.1 88.8%τ-bench banking (Artificial Analysis) 82 · Τ-bench banking 47.8% Knowledge work 43GDP.pdf 44 · Professional document tasks 23.2%DeckBench 42 · Deck rating 1,076 (95% range 995–1,177) Writing and preference 63Arena text 63 · Overall 1,482 (95% range 1,476–1,487) Factuality 53Arena factuality 68 · Overall 1,472 (95% range 1,467–1,476)SimpleQA Verified (Epoch AI) 38 · Correct answers 47.3%
14	Grok 4.7xAI	43	64	40	58	18	5	$3.00	6 Oct 2026
Spring Prompt benches 64CatalogBench 78 · Reliably publish-ready 57.1%; Sales brief: reliably publish-ready 33.9%ROASBench 42 · Overall score 36.6 Agents and tool use 40Terminal-Bench (Artificial Analysis) 40 · Terminal-Bench 4.0 25.8% Knowledge work 58DeckBench 58 · Deck rating 1,245 (95% range 1,161–1,373) Writing and preference 18Arena text 16 · Overall 1,442 (95% range 1,435–1,450)UGI Leaderboard 20 · Writing score 54.1 Factuality 5Arena factuality 5 · Overall 1,428 (95% range 1,421–1,435)
15	Claude Haiku 5.5Anthropic	42	28	55	55	–	–	$0.20	8 Oct 2026
Spring Prompt benches 28CatalogBench 38 · Reliably publish-ready 26.8%; Sales brief: reliably publish-ready 10.7%ROASBench 11 · Overall score 24.1BulletBench 31 · Ladder Elo 310 (95% range 0–575) Agents and tool use 55Terminal-Bench (Artificial Analysis) 55 · Terminal-Bench 4.0 32.8% Knowledge work 55DeckBench 47 · Deck rating 1,150 (95% range 1,055–1,259)SurveyBench 63 · Analysis rating 1,261 (95% range 1,211–1,328) No results: Writing and preference, Factuality
16	GPT-6 LunaOpenAI	40	70	23	39	21	13	$0.20	6 Oct 2026
Spring Prompt benches 70CatalogBench 78 · Reliably publish-ready 55.4%; Sales brief: reliably publish-ready 51.8%ROASBench 89 · Overall score 53.4BulletBench 23 · Ladder Elo 249 (95% range 0–453) Agents and tool use 23Terminal-Bench (Artificial Analysis) 23 · Terminal-Bench 4.0 12.6% Knowledge work 39APEX-Agents 9 · Tasks passed 44.3%GDP.pdf 33 · Professional document tasks 23.0%DeckBench 74 · Deck rating 1,288 (95% range 1,196–1,426) Writing and preference 21Arena text 21 · Overall 1,443 (95% range 1,436–1,449) Factuality 13Arena factuality 11 · Overall 1,430 (95% range 1,424–1,436)SimpleQA Verified (Epoch AI) 15 · Correct answers 41.4%
17	GLM-5.3Z.ai	38	13	75	36	58	30	$2.15	1 Oct 2026
Spring Prompt benches 13ROASBench 16 · Overall score 24.8BulletBench 8 · Ladder Elo 33 (95% range 0–270) Agents and tool use 75Terminal-Bench (Artificial Analysis) 58 · Terminal-Bench 4.0 41.9%; Terminal-Bench 2.1 83.9%τ-bench banking (Artificial Analysis) 91 · Τ-bench banking 50.3%Vending-Bench 2 75 · Money after a year $8,164 Knowledge work 36APEX-Agents 36 · Tasks passed 56.6% Writing and preference 58Arena text 58 · Overall 1,478 (95% range 1,473–1,484) Factuality 30Arena factuality 53 · Overall 1,467 (95% range 1,462–1,472)SimpleQA Verified (Epoch AI) 8 · Correct answers 41.0%
18=	Mistral Large 4Mistral AI	34	31	48	25	–	–	$2.06	8 Oct 2026
Spring Prompt benches 31CatalogBench 14 · Reliably publish-ready 10.7%; Sales brief: reliably publish-ready 1.8%ROASBench 47 · Overall score 37.0BulletBench 50 · Ladder Elo 249 (95% range 0–486) Agents and tool use 48Terminal-Bench (Artificial Analysis) 48 · Terminal-Bench 4.0 26.8% Knowledge work 25DeckBench 0 · Deck rating 560 (95% range 442–653)SurveyBench 50 · Analysis rating 797 (95% range 726–855) No results: Writing and preference, Factuality
18=	DeepSeek-V4-Pro (0423)DeepSeek	34	32	30	16	48	48	$1.77	6 Oct 2026
Spring Prompt benches 32ROASBench 32 · Overall score 28.3 Agents and tool use 30Terminal-Bench (Artificial Analysis) 29 · Terminal-Bench 4.0 14.6%; Terminal-Bench 2.1 64.8%τ-bench banking (Artificial Analysis) 36 · Τ-bench banking 30.1%Vending-Bench 2 25 · Money after a year $3,285 Knowledge work 16DeckBench 16 · Deck rating 690 (95% range 526–809) Writing and preference 48Arena text 37 · Overall 1,458 (95% range 1,454–1,462)UGI Leaderboard 60 · Writing score 68.4 Factuality 48Vectara Hallucination Leaderboard 75 · Hallucination rate 8.6%Arena factuality 37 · Overall 1,455 (95% range 1,452–1,458)SimpleQA Verified (Epoch AI) 31 · Correct answers 47.0%
20	Gemini 3.5 Flash-LiteGoogle	26	36	16	12	26	32	$0.85	8 Oct 2026
Spring Prompt benches 36CatalogBench 23 · Reliably publish-ready 17.9%; Sales brief: reliably publish-ready 3.6%ROASBench 21 · Overall score 25.4BulletBench 92 · Lightning 10+1: ladder Elo 892 (95% range 641–1,081); Bullet 60s: ladder Elo 780 (95% range 586–956) Agents and tool use 16Terminal-Bench (Artificial Analysis) 14 · Terminal-Bench 4.0 1.0%; Terminal-Bench 2.1 53.6%τ-bench banking (Artificial Analysis) 18 · Τ-bench banking 17.5% Knowledge work 12GDP.pdf 0 · Professional document tasks 10.0%DeckBench 11 · Deck rating 688 (95% range 565–775)SurveyBench 25 · Analysis rating 632 (95% range 558–696) Writing and preference 26Arena text 26 · Overall 1,455 (95% range 1,450–1,459) Factuality 32Arena factuality 32 · Overall 1,447 (95% range 1,444–1,451)
21	Mistral Medium 3.5Mistral AI	16	31	7	9	3	16	$3.00	8 Oct 2026
Spring Prompt benches 31CatalogBench 33 · Reliably publish-ready 17.9%; Sales brief: reliably publish-ready 10.7%ROASBench 0 · Overall score 15.2BulletBench 72 · Lightning 10+1: ladder Elo 529 (95% range 361–669); Bullet 60s: ladder Elo 572 (95% range 428–708) Agents and tool use 7Terminal-Bench (Artificial Analysis) 6 · Terminal-Bench 4.0 0.0%; Terminal-Bench 2.1 50.6%τ-bench banking (Artificial Analysis) 9 · Τ-bench banking 15.1% Knowledge work 9DeckBench 5 · Deck rating 600 (95% range 437–739)SurveyBench 13 · Analysis rating 555 (95% range 471–623) Writing and preference 3Arena text 5 · Overall 1,427 (95% range 1,421–1,433)UGI Leaderboard 0 · Writing score 45.5 Factuality 16Arena factuality 16 · Overall 1,431 (95% range 1,425–1,436)
22	GLM-5V-TurboZ.ai	11	4	–	–	11	26	$1.90	30 Sep 2026
Spring Prompt benches 4CatalogBench 4 · Reliably publish-ready 1.8%; Sales brief: reliably publish-ready 0.0% Writing and preference 11Arena text 11 · Overall 1,433 (95% range 1,427–1,440) Factuality 26Arena factuality 26 · Overall 1,440 (95% range 1,435–1,446) No results: Agents and tool use, Knowledge work
23	Claude Haiku 4.5Anthropic	7	10	0	11	5	8	$2.00	8 Oct 2026
Spring Prompt benches 10CatalogBench 4 · Reliably publish-ready 1.8%; Sales brief: reliably publish-ready 0.0%ROASBench 5 · Overall score 17.5BulletBench 33 · Lightning 10+1: ladder Elo 33 (95% range 0–248); Bullet 60s: ladder Elo 445 (95% range 135–626) Agents and tool use 0Terminal-Bench (Artificial Analysis) 1 · Terminal-Bench 4.0 0.0%; Terminal-Bench 2.1 44.2%τ-bench banking (Artificial Analysis) 0 · Τ-bench banking 9.3%Vending-Bench 2 0 · Money after a year $458.89 Knowledge work 11DeckBench 21 · Deck rating 724 (95% range 615–816)SurveyBench 0 · Analysis rating 495 (95% range 406–562) Writing and preference 5Arena text 0 · Overall 1,414 (95% range 1,411–1,416)UGI Leaderboard 10 · Writing score 52.2 Factuality 8Vectara Hallucination Leaderboard 25 · Hallucination rate 9.8%Arena factuality 0 · Overall 1,418 (95% range 1,416–1,420)SimpleQA Verified (Epoch AI) 0 · Correct answers 13.2%

Each row shows every group's score; tap a bar for its breakdown.

How the score works

Spring Prompt benches 35%

Catalogue listings, ad budgets and fast chess: our own tests.

CatalogBench · 50% of the groupCatalogBench: reliably publish-ready; CatalogBench (sales brief): reliably publish-ready
ROASBench · 30% of the groupROASBench: overall score
BulletBench · 20% of the groupBulletBench Lightning 10+1: ladder Elo; BulletBench Bullet 60s: ladder Elo

Agents and tool use 20%

Multi-step tasks with tools: support desks, coding, function calls.

tau2-benchtau2 v1.0.1, retail: task success (pass^1); tau2 v1.0.1, airline: task success (pass^1); tau2 v1.0.1, telecom: task success (pass^1)Not counted yet: it has tested fewer than 5 of these models.
Berkeley Function Calling Leaderboard (BFCL) V4BFCL: overall accuracyNot counted yet: it has tested fewer than 5 of these models.
OpenHands IndexOpenHands Index: averageNot counted yet: it has tested fewer than 5 of these models.
Microsoft STATE-BenchSTATE-Bench v0.7: task success (pass@1)Not counted yet: it has tested fewer than 5 of these models.
Terminal-Bench (Artificial Analysis)Artificial Analysis: Terminal-Bench 4.0; Artificial Analysis: Terminal-Bench 2.1
τ-bench banking (Artificial Analysis)Artificial Analysis: τ-bench banking
Vending-Bench 2Vending-Bench 2: money after a year

Knowledge work 15%

Real professional tasks: banking, consulting and legal work, business documents and decks, user surveys, paid freelance projects.

APEX-AgentsAPEX-Agents: tasks passed
GDP.pdfGDP.pdf: professional document tasks
Remote Labor IndexRemote Labor Index: projects done to client standardNot counted yet: it has tested fewer than 5 of these models.
DeckBenchDeckBench: deck rating
SurveyBenchSurveyBench: analysis rating

Writing and preference 15%

What people prefer in blind comparisons, and judged writing quality.

Arena textArena Text: overall
UGI LeaderboardUGI: writing score

Factuality 15%

Sticking to the source when summarising, and factual answers.

Vectara Hallucination LeaderboardVectara: hallucination rate
Arena factualityArena Text factuality: overall
SimpleQA Verified (Epoch AI)SimpleQA Verified: correct answers

Same field for every number. Only the models in this table are compared, so a model is not lifted by a source that tested it against older ones. A benchmark counts once it has tested 5 of them.
Percentiles, not raw scores. On each headline metric a model scores the share of the others it beats (ties count half). Where a model was run in more than one configuration, such as different reasoning efforts, its best result on that metric counts; benchmark pages list every configuration. Arena ratings, pass rates and Elo can then be averaged.
Missing is not zero. A model needs CatalogBench or ROASBench and at least 2 groups. Groups it has no results in are left out and the others re-weighted; a dash marks them.
What it does not show. Scores are rounded to whole points, halves up, and models on the same whole number share a rank (marked =). Differences of a few points are within noise, and a percentile hides how far apart two models are. Ratings such as DeckBench, SurveyBench and BulletBench come with 95% ranges, shown in each model's breakdown; where two models' ranges overlap, the order between them may not be real. For a decision, read the benchmark pages; for your own product, test it on your data.

Derived from results reported by APEX-Agents: Published with permission (Mercor results via Epoch AI); Artificial Analysis: Artificial Analysis commercial data licence; GDP.pdf: Published with permission (Surge AI results via Epoch AI); Arena (formerly LMArena): Creative Commons Attribution 4.0 International; SimpleQA Verified (Epoch AI): Creative Commons Attribution 4.0 International; UGI Leaderboard: Apache License 2.0; Vectara Hallucination Leaderboard: Apache License 2.0; Vending-Bench 2: Published with permission (Andon Labs results via Epoch AI); converted to percentiles by Spring Prompt. Also in the method, and counted once they have tested 5 of these models: Berkeley Function Calling Leaderboard (BFCL) V4: Apache License 2.0; OpenHands Index: Apache License 2.0; Remote Labor Index: Published with permission (Scale AI and the Center for AI Safety results via Epoch AI); Microsoft STATE-Bench: MIT License; tau2-bench: MIT License.