Benchmarks / Overall leaderboard

Combined by Spring Prompt

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
36
Sources in the score
8 licensed sources + 5 of our own benchmarks
Release
16 Nov 2026
Score
0–100, higher is better

Choose a model for the work you do

The overall score is a starting point. Use task results, failure examples and measured costs to build your shortlist.

Choosing a listing model? Read our product-description model guide.

Our own data: CatalogBench, ROASBench and BulletBench are in Spring Prompt benches (35%); DeckBench and SurveyBench are in Knowledge work (15%). Across these 36 models, 62% of a score comes from our own benchmarks on average (from 35% to 100%, depending on which results a model has).

Want a score from our own tests only? The Spring Prompt Score averages our own benchmarks with equal weights and no third-party results, so it is free to cite, download and embed.

Overall score from 0 to 100, higher is better, with each group's score, blended price and the date we last tested it
#ModelOverallSpring Prompt benchesAgents and tool useKnowledge workWriting and preferenceFactualityPrice
$ / M tokens
We last tested
1 Claude Fable 5.2AnthropicProvisional · 2 of 5 groups
90
88–95–– $16.00 10 Nov 2026
2 Claude Opus 5.5Anthropic
82
6990908886 $8.00 8 Oct 2026
3= GPT-6.1 SolOpenAI
80
9581785866 $4.00 9 Oct 2026
3= GPT-6.1 TerraOpenAIProvisional · 3 of 5 groups
80
847774–– $2.00 27 Oct 2026
5 GPT-6 AstraOpenAI
77
8779856065 $20.00 6 Oct 2026
6 Gemini 4 ArgonGoogle
72
578535100100 $4.00 14 Oct 2026
7 Muse WonderMetaProvisional · 3 of 5 groups
70
666979–– $8.00 12 Oct 2026
8= GPT-6 SolOpenAI
68
8576633058 $4.00 6 Oct 2026
8= Kimi K3.1 ProMoonshot AIProvisional · 3 of 5 groups
68
6310035–– $3.81 27 Oct 2026
10= Muse EmberMetaProvisional · 2 of 5 groups
65
65–66–– $0.20 10 Nov 2026
10= Gemini 4 HydrogenGoogleProvisional · 2 of 5 groups
65
70–53–– $3.00 16 Nov 2026
12 Claude Fable 5.1Anthropic
64
4674806580 $20.00 6 Oct 2026
13 Claude Sonnet 5.5Anthropic
61
4596924042 $4.00 6 Oct 2026
14 Muse Spark 1.3Meta
59
3378498085 $2.00 6 Oct 2026
15 Gemini 3.8 FlashGoogle
58
4947449371 $3.00 10 Nov 2026
16 Qwen4-MaxAlibabaProvisional · 2 of 5 groups
56
41–91–– $2.25 16 Nov 2026
17 Gemini 4 CarbonGoogleProvisional · 2 of 5 groups
53
59–41–– $8.00 11 Nov 2026
18 DeepSeek-V4.1-FlashDeepSeek
51
5642–4555 $0.52 30 Sep 2026
19 Kimi K3Moonshot AI
47
3350337561 $6.00 6 Oct 2026
20 Gemini 3.1 Pro PreviewGoogle
46
5721138052 $4.50 8 Oct 2026
21 Qwen3.8-Max (0902)Alibaba
45
1876416052 $3.00 6 Oct 2026
22 MiniMax M3.1 FlashMiniMaxProvisional · 3 of 5 groups
42
483738–– $0.26 4 Nov 2026
23= Grok 4.7xAI
40
563766185 $3.00 6 Oct 2026
23= Claude Haiku 5.5AnthropicProvisional · 3 of 5 groups
40
274660–– $0.20 8 Oct 2026
25 GPT-6 LunaOpenAI
39
7017422013 $0.20 10 Nov 2026
26 DeepSeek-V4.1-Flash-1017DeepSeekProvisional · 3 of 5 groups
38
245451–– $0.52 19 Oct 2026
27 GLM-5.3Z.ai
35
773365529 $2.15 1 Oct 2026
28 Mercury 3InceptionProvisional · 2 of 5 groups
34
40–20–– – 19 Oct 2026
29 GLM-5.3 FlashX LegalZ.aiProvisional · 2 of 5 groups
31
29–36–– $0.59 4 Nov 2026
30 Particle-1Particle AIProvisional · 2 of 5 groups
30
19–54–– $0.16 4 Nov 2026
31 DeepSeek-V4-Pro (0423)DeepSeek
28
222904847 $1.77 6 Oct 2026
32= Mistral Large 4Mistral AIProvisional · 3 of 5 groups
25
243118–– $2.06 8 Oct 2026
32= Gemini 3.5 Flash-LiteGoogle
25
341692530 $0.85 8 Oct 2026
34 Mistral Medium 3.5Mistral AI
15
3076315 $3.00 8 Oct 2026
35 GLM-5V-TurboZ.aiProvisional · 3 of 5 groups
12
8––1025 $1.90 30 Sep 2026
36 Claude Haiku 4.5Anthropic
8
120858 $2.00 8 Oct 2026

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.

Knowledge work 15%

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

Writing and preference 15%

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

Factuality 15%

Sticking to the source when summarising, and factual answers.

  1. 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.
  2. 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.
  3. 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, and a hatched slot in the bar. With results in fewer than 4 groups a rank is provisional and marked so: a new model often has our benches before outside sources have tested it, and a missing group could move it a long way.
  4. 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.