LLM comparison 2026
Compare LLMs without mixing incompatible test setups
Claude Fable 5.2, Gemini 4 Carbon and Claude Opus 5.5 are practically level at the top of the overall index. Nemotron 3.5 Lightning is the fastest, at 180 tokens per second. Among the ten best placed models, Gemini 4 Carbon is the cheapest at $4.00 per million tokens.
- models
- 59
- head-to-head matchups
- 1,670
- active benchmarks
- 138
- active results
- 26,897
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
The best LLMs in the overall index
20 of 95 models
Performance groupsTop group (3)Upper group (7)Middle group (10)
Claude Fable 5.2max131.0
Gemini 4 Carbonhigh127.0
Claude Opus 5.5max125.7
Claude Fable 5.1max124.1
Claude Sonnet 5.5max121.8
Gemini 4 Argonhigh121.4
GPT-6 Astramax120.2
Muse Wonder Fullmax118.1
GPT-6.1 Solmax118.0
Claude Opus 5max117.4
Muse Wondermax113.4
Claude Fable 5max112.8
Kimi K3.1 Promax112.6
Muse Ember 1.0max110.4
GPT-6 Solmax109.1
GPT-5.6 Solmax109.1
Gemini 3.7 Flashhigh107.9
Muse Spark 1.3max107.3
Gemini 3.8 Flashhigh106.8
DeepSeek V4.1 Flash 1017max106.7
Embed this chart
Paste the code into your page. The chart adapts to the width and updates with our data.
You may use the chart with attribution (CC BY 4.0).
Compare up to five LLMs side by side
- Claude Fable 5.2
- Gemini 4 Carbon
- Claude Opus 5.5
In bold: the best value of the row. For the overall index and the areas only when the lead is certain.
Intelligence
Overall indexpoints, 100 = reference group average, higher is better
- Claude Fable 5.2
- 131.0
- Gemini 4 Carbon
- 127.0
- Claude Opus 5.5
- 125.7
Rank
- Claude Fable 5.2
- Rank 1 of 95
- Gemini 4 Carbon
- Rank 2 of 95
- Claude Opus 5.5
- Rank 3 of 95
Performance group
- Claude Fable 5.2
- Top group
- Gemini 4 Carbon
- Top group
- Claude Opus 5.5
- Top group
Areas of the overall index
Coding
- Claude Fable 5.2
- 131.5
- Gemini 4 Carbon
- 123.4
- Claude Opus 5.5
- 130.0
Reasoning
- Claude Fable 5.2
- 124.9
- Gemini 4 Carbon
- 122.7
- Claude Opus 5.5
- 121.8
Mathematics
- Claude Fable 5.2
- n/a
- Gemini 4 Carbon
- n/a
- Claude Opus 5.5
- n/a
Knowledge
- Claude Fable 5.2
- n/a
- Gemini 4 Carbon
- n/a
- Claude Opus 5.5
- n/a
Agents
- Claude Fable 5.2
- 124.1 (best value)
- Gemini 4 Carbon
- 119.1
- Claude Opus 5.5
- 116.1
General
- Claude Fable 5.2
- 132.8
- Gemini 4 Carbon
- n/a
- Claude Opus 5.5
- 125.3
Multimodal
- Claude Fable 5.2
- n/a
- Gemini 4 Carbon
- n/a
- Claude Opus 5.5
- 124.3
Finance
- Claude Fable 5.2
- 123.5
- Gemini 4 Carbon
- 123.8
- Claude Opus 5.5
- 120.7
Legal
- Claude Fable 5.2
- 120.9
- Gemini 4 Carbon
- 123.8
- Claude Opus 5.5
- 120.6
Health
- Claude Fable 5.2
- n/a
- Gemini 4 Carbon
- n/a
- Claude Opus 5.5
- 120.9
Price per 1M tokens
Blended price (3:1)3 parts input, 1 part output
- Claude Fable 5.2
- $16
- Gemini 4 Carbon
- $4 (best value)
- Claude Opus 5.5
- $8
Inputwhat you send
- Claude Fable 5.2
- $8
- Gemini 4 Carbon
- $2 (best value)
- Claude Opus 5.5
- $4
Outputwhat the model writes
- Claude Fable 5.2
- $40
- Gemini 4 Carbon
- $10 (best value)
- Claude Opus 5.5
- $20
Speed
Tokens per second
- Claude Fable 5.2
- n/a
- Gemini 4 Carbon
- n/a
- Claude Opus 5.5
- 75
Technical
Context windowhow much text the model reads at once
- Claude Fable 5.2
- 1 M tokens≈ 1,500 A4 pages
- Gemini 4 Carbon
- 1.05 M tokens (best value)≈ 1,570 A4 pages
- Claude Opus 5.5
- 1 M tokens≈ 1,500 A4 pages
Released
- Claude Fable 5.2
- Nov 9, 2026
- Gemini 4 Carbon
- Nov 10, 2026
- Claude Opus 5.5
- Sep 22, 2026
Open weights
- Claude Fable 5.2
- no
- Gemini 4 Carbon
- no
- Claude Opus 5.5
- no
Input types
- Claude Fable 5.2
- Text, Image
- Gemini 4 Carbon
- Text, Image, Audio, Video
- Claude Opus 5.5
- Text, Image
Output types
- Claude Fable 5.2
- Text
- Gemini 4 Carbon
- Text
- Claude Opus 5.5
- Text
Model profiles: Claude Fable 5.2, Gemini 4 Carbon, Claude Opus 5.5
The fastest LLMs
20 of 95 models
Nemotron 3.5 Lightning180tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Claude Haiku 5.5164tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Gemini 3.5 Flash-Lite145tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Gemini 3.5 Flash142tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
GPT-5 nano136tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
GPT-5.4 mini134tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
DeepSeek V4.1 Flash 1017126tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Gemini 3.1 Flash-Lite125tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
GPT-5.6 Luna123tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Mistral Medium 3.5118tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
DeepSeek-V4.1-Flash118tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Mistral Small 4112tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Gemini 3.6 Flash110tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
GPT-4.1 nano110tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Ling 3.0 Flash106tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
GPT-5 mini105tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Claude Sonnet 5.5103tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Muse Spark 1.3102tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Gemini 3 Flash Preview101tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
GLM-5.3-Flash101tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-08
Output tokens per second, measured by OpenRouter
Intelligence vs. price
most attractive area: stronger and cheaper than half of the models shownbest intelligence for the price
More models in the chart (25)
- Claude Opus 5.5125.7 · $8
- GPT-6 Astra120.2 · $20
- GPT-6.1 Sol118.0 · $4
- Claude Opus 5117.4 · $10
- Muse Wonder113.4 · $8
- Claude Fable 5112.8 · $20
- Kimi K3.1 Pro112.6 · $4
- Muse Ember 1.0110.4 · $0.20
- GPT-6 Sol109.1 · $4
- GPT-5.6 Sol109.1 · $8
- Muse Spark 1.3107.3 · $2
- Gemini 3.8 Flash106.8 · $1.50
- DeepSeek V4.1 Flash 1017106.7 · $0.26
- GPT-6.1 Terra106.5 · $2
- Grok 4.7103.3 · $3
- Claude Haiku 5.5102.8 · $0.20
- Grok 4.6102.2 · $3
- GPT-5.5101.5 · $11.25
- Kimi K3101.1 · $6
- GPT-5.6 Terra100.8 · $4.50
- Qwen 3.8 Max 0902100.8 · $2.48
- Mistral Large 4100.8 · $1.05
- GLM-5.3100.6 · $2.15
- MiMo V2.6 Pro100.4 · $0.54
- MiniMax M3.1 Flash100.0 · $0.26
Performance vs. cost per task
These are the costs Vals AI paid per task of the Vals Index at the stated reasoning tier. Higher tiers cost more because the model thinks for longer.
most attractive area: stronger and cheaper than half of the models shownbest intelligence for the price
More models in the chart (24)
- Claude Opus 5.5125.7 · $24.77
- Claude Sonnet 5.5121.8 · $16.18
- GPT-6 Astra · max120.2 · $21.28
- Muse Wonder Full · max118.1 · $6.81
- GPT-6.1 Sol · max118.0 · $3.61
- Kimi K3.1 Pro · max112.6 · $6.44
- Gemini 3.8 Flash · high106.8 · $5.03
- DeepSeek V4.1 Flash 1017 · max106.7 · $0.66
- GPT-6.1 Terra · max106.5 · $2.94
- Grok 4.7 · xhigh103.3 · $11.63
- Claude Haiku 5.5102.8 · $3.17
- Kimi K3 · max101.1 · $7.17
- Mistral Large 4 · high100.8 · $3.72
- GLM-5.3 · max100.6 · $8.52
- MiMo V2.6 Pro100.4 · $0.43
- MiniMax M3.1 Flash100.0 · $0.66
- Muse Spark Open · xhigh97.5 · $4.72
- MiMo V2.6 Flash97.3 · $0.22
- Step 5 Preview97.1 · $2.73
- GPT-6 Luna · max95.8 · $0.45
- GLM-5.3-Flash · max94.3 · $2.05
- Inkling72.0 · $1.85
- Inkling Small68.7 · $0.40
- Nemotron 3.5 Super VL 120B A12B · high68.0 · $0.63
The best LLM per price band over time
For each price band, the line shows which model led in that month. The bands follow the blended price per million tokens (three parts input, one part output).
- under $1
- $1 to $5
- over $5
Changes at the top (25)
under $1
- GPT-5 miniAug 2025, 65.8
- DeepSeek-V4-ProApr 2026, 87.2
- GPT-5.6 LunaJul 2026, 93.8
- GLM-5.3-FlashAug 2026, 94.3
- MiMo V2.6 ProSep 2026, 100.4
- DeepSeek V4.1 Flash 1017Oct 2026, 106.7
- Muse Ember 1.0Nov 2026, 110.4
$1 to $5
- DeepSeek-R1Jan 2025, 44.6
- o3Apr 2025, 74.2
- GPT-5Aug 2025, 77.6
- Gemini 3 Flash PreviewDec 2025, 81.0
- Gemini 3.1 Pro PreviewFeb 2026, 91.4
- Claude Sonnet 5Jun 2026, 92.1
- GPT-5.6 TerraJul 2026, 100.8
- Grok 4.6Aug 2026, 102.2
- Claude Sonnet 5.5Sep 2026, 121.8
- Gemini 4 CarbonNov 2026, 127.0
over $5
- Claude Sonnet 4.5Sep 2025, 65.9
- Claude Opus 4.5Nov 2025, 74.4
- Claude Sonnet 4.6Feb 2026, 86.7
- GPT-5.5Apr 2026, 101.5
- Claude Fable 5Jun 2026, 112.8
- Claude Opus 5Jul 2026, 117.4
- Claude Opus 5.5Sep 2026, 125.7
- Claude Fable 5.2Nov 2026, 131.0
The lines use today's index values and today's prices. A model may have been rated or priced differently when it launched.
The overall index over time
How the ten best-placed models of today moved between the stored snapshots.
All values at a glance
- Claude Fable 5.2131.0
- Gemini 4 Carbon127.0
- Claude Opus 5.5125.7
- Claude Fable 5.1124.1
- Claude Sonnet 5.5121.8
- Gemini 4 Argon121.4
- GPT-6 Astra120.2
- Muse Wonder Full118.1
- GPT-6.1 Sol118.0
- Claude Opus 5117.4
Only snapshots of the same method version of the overall index are compared.
Who leads in LLMs?
Anthropic has the highest-scoring model, Claude Fable 5.2. The best models of Google are level with it.
Claude Fable 5.2Rank 1131.0
Claude Opus 5.5Rank 3125.7
Claude Fable 5.1Rank 4124.1
+ 12 more models in the index
Gemini 4 CarbonRank 2127.0
Gemini 4 ArgonRank 6121.4
Gemini 3.7 FlashRank 17107.9
+ 8 more models in the index
GPT-6 AstraRank 7120.2
GPT-6.1 SolRank 9118.0
GPT-6 SolRank 15109.1
+ 15 more models in the index
Muse Wonder FullRank 8118.1
Muse WonderRank 11113.4
Muse Ember 1.0Rank 14110.4
+ 2 more models in the index
Kimi K3.1 ProRank 13112.6
Kimi K3Rank 26101.1
Kimi K2.6Rank 5284.9
+ 3 more models in the index
DeepSeek V4.1 Flash 1017Rank 20106.7
DeepSeek-V4.1-FlashRank 3499.5
DeepSeek-V4-ProRank 4887.2
+ 1 more model in the index
Grok 4.7Rank 22103.3
Grok 4.6Rank 24102.2
Grok 4.5Rank 4294.8
+ 2 more models in the index
Qwen 3.8 Max 0902Rank 28100.8
Qwen 3.7 MaxRank 5185.4
Qwen 3.7 PlusRank 5483.4
+ 1 more model in the index
Mistral Large 4Rank 29100.8
Mistral Medium 3.5Rank 8946.3
Mistral Large 3Rank 9532.8
GLM-5.3Rank 30100.6
GLM-5.3-FlashRank 4394.3
GLM-5.2Rank 4789.0
+ 4 more models in the index
MiMo V2.6 ProRank 31100.4
MiMo V2.6 FlashRank 3897.3
MiMo-V2.5-ProRank 6673.0
+ 1 more model in the index
MiniMax M3.1 FlashRank 32100.0
MiniMax M3Rank 5583.0
MiniMax M2.7Rank 7865.1
+ 2 more models in the index
Other providers with models in the index: StepFun, Reflection AI, Thinking Machines Lab, NVIDIA, Upstage, Inception.
The best choice per price tier
Highest overall index within the input price range.
Under $0.50 input
- 2.Claude Haiku 5.5102.8
- 3.DeepSeek-V4.1-Flash99.5
- 4.GPT-6 Luna95.8
$0.50 to under $2 input
- 2.Muse Spark 1.3107.3
- 3.Gemini 3.8 Flash106.8
- 4.GPT-6.1 Terra106.5
$2 to under $5 input
- 2.Claude Opus 5.5125.7
- 3.Claude Sonnet 5.5121.8
- 4.Gemini 4 Argon121.4
$5 input and above
- 2.Claude Fable 5.1124.1
- 3.GPT-6 Astra120.2
- 4.Claude Opus 5117.4
The best LLM for your task
Terminal coding
Terminal-Bench 4.0Vals AI38 models measured
Which model completes programming tasks on its own in the command line?
Retrieved: November 7, 2026. Only comparison models with an exactly matched result.
The published leaderboard documents model-specific test settings. 1 comparison model has more than one incompatible setup. It is omitted from this ranking instead of selecting the best result.
Who leads across several tasks?
Place in the top 8 of each task. Empty cells mean: not in the top eight or not measured.
| Model | Terminal coding | Building apps | Expert knowledge | Chat quality | Finance tasks | Hits |
|---|---|---|---|---|---|---|
| Claude Fable 5.1 | 6. | 6. | 7. | 5. | 5. | 5 |
| Gemini 4 Argon | 8. | 2. | 1. | 1. | 4 | |
| Claude Opus 5.5 | 1. | 5. | 3. | 7. | 4 | |
| Claude Sonnet 5.5 | 2. | 1. | 8. | 3 | ||
| Gemini 3.8 Flash | 4. | 6. | 2. | 3 | ||
| Gemini 3.7 Flash | 5. | 8. | 4. | 3 | ||
| Claude Fable 5 | 4. | 4. | 2 | |||
| Muse Wonder Full | 6. | 3. | 2 | |||
| Gemini 4 Carbon | 4. | 8. | 2 | |||
| GPT-6 Astra | 5. | 7. | 2 | |||
| Claude Opus 5 | 7. | 6. | 2 |
At a glance
Latest releases
Release date
- 1.Gemini 4 CarbonNovember 10, 2026
- 2.Claude Fable 5.2November 9, 2026
- 3.GPT-6.1 TerraOctober 26, 2026
- 4.Kimi K3.1 ProOctober 23, 2026
- 5.DeepSeek V4.1 Flash 1017October 17, 2026
- 6.Gemini 4 ArgonOctober 13, 2026
Cheap and strong
Input per 1M tokens · overall index 100+ (reference group average)
- 1.Claude Haiku 5.5$0.10 · 102.8
- 2.DeepSeek V4.1 Flash 1017$0.15 · 106.7
- 3.Mistral Large 4$0.70 · 100.8
- 4.Gemini 3.7 Flash$0.75 · 107.9
- 5.Gemini 3.8 Flash$0.75 · 106.8
- 6.GPT-6.1 Terra$1.00 · 106.5
Largest context
Context window in tokens
- 1.GPT-6 Astra1.05M
- 2.GPT-6.1 Sol1.05M
- 3.GPT-6 Sol1.05M
- 4.GPT-5.6 Sol1.05M
- 5.GPT-6.1 Terra1.05M
- 6.GPT-5.51.05M
Popular comparisons
The most searched matchups, at most three per model.
- 1Claude Opus 5.5 vs. GPT-6 Astra
- 2GPT-6 Luna vs. GPT-5.6 Luna
- 3GPT-6 Luna vs. GPT-5.6 Terra
- 4Claude Sonnet 5.5 vs. Gemini 3.8 Flash
- 5Claude Opus 5.5 vs. DeepSeek-V4-Pro
- 6Claude Opus 5.5 vs. Gemini 3.8 Flash
- 7Gemini 3.8 Flash vs. DeepSeek-V4.1-Flash
- 8Claude Sonnet 5.5 vs. DeepSeek-V4.1-Flash
- 9DeepSeek V4.1 Flash 1017 vs. GLM-5.3-Flash
Data and feeds
The comparison matrix shows 59 current models. 95 models have a place in the overall index, because the overall index covers the whole model directory, not only the comparison matrix.
The 59 most relevant models
The selection includes current frontier models, lower-cost options, and selected earlier generations. You can compare Claude Opus 4.5 to 4.8, Sonnet 4.5 and 4.6, and Haiku 4.5 with newer models.
For the full technical profile of every model, browse the AI model directory. Which open-weight models run on your own hardware is shown by the open-source LLM ranking.
Looking for older models? GPT-3.5 vs. GPT-4: What's the Difference?
Anthropic
OpenAI
Google DeepMind
Moonshot AI
Xiaomi
The last 45 days
New models and price changes, newest first. Prices in dollars per 1M tokens (input / output); a new price only counts once it holds for three days.
PriceDeepSeek-V4-Pro $1.31 / $3.96 · not confirmed yet
PriceMiMo-V2.5 $0.40 / $2 · not confirmed yet
PriceClaude Sonnet 5 $2 / $10 · not confirmed yet
PriceGLM-4.7 $0.40 / $1.75 · not confirmed yet
PriceGLM-5 $0.95 / $2.55 · not confirmed yet
PriceGLM-5.2 $1.30 / $4.40 · not confirmed yet
PriceGrok 4.7 $2 / $6 · not confirmed yet
PriceKimi K2.5 $0.50 / $2.50 · not confirmed yet
PriceKimi K2.7 Code $0.75 / $3.50 · not confirmed yet
PriceLlama 4 Scout $0.10 / $0.30 · not confirmed yet
PriceMiMo-V2.5-Pro $0.435 / $0.87 · not confirmed yet
NewERNIE 5.5 Flash Preview · not in the overall index yet
PriceQwen 2.5 7B cheaper $0.05 / $0.05 · before $0.05 / $0.13
PriceQwen 3.5 397B A17B more expensive $0.60 / $3.60 · before $0.55 / $3.50
NewBielik 11B v3.1 Instruct · not in the overall index yet
PriceMiniMax M1 more expensive $0.55 / $2.20 · before $0.40 / $2.20
PriceGLM-5.2 more expensive $1.40 / $4.40 · before $1.30 / $4.40
PriceQwen 3.8 27B more expensive $0.42 / $3 · before $0.20 / $2.50
PriceGLM-5.3 more expensive $1.40 / $4.40 · before $0.90 / $4
PriceInkling more expensive $1 / $4.05 · before $0.95 / $4.05
PriceLlama 3.3 70B input more expensive, output cheaper $0.59 / $0.79 · before $0.45 / $0.90
PriceGemini 3.6 Flash cheaper $0.75 / $3.75 · before $0.825 / $4.125
PriceMiniMax M2 more expensive $0.30 / $1.20 · before $0.255 / $1.02
NewERNIE 5.5 Flash Base, GPT-6 Cyber, Inkling 2 Small · not in the overall index yet
NewHy4 · not in the overall index yet
PriceDeepSeek-V3.2 cheaper $0.27 / $0.41 · before $0.27 / $0.42
PriceDeepSeek-V4-Pro more expensive $1.32 / $3.96 · before $1.31 / $3.96
PriceKimi K2.7 Code cheaper $0.912 / $3.84 · before $0.95 / $4
PriceNemotron 3 Nano more expensive $0.06 / $0.24 · before $0.05 / $0.20
PriceQwen 3 Coder 480B A35B input more expensive, output cheaper $0.29 / $1.20 · before $0.22 / $1.80
PriceQwen 3.6 35B A3B cheaper $0.10 / $0.95 · before $0.10 / $1
PriceLaguna XS 2.1 more expensive $0.10 / $0.20 · before $0.06 / $0.12
PriceSolar Pro 3 input more expensive, output cheaper $0.25 / $0.25 · before $0.15 / $0.60
NewGPT-Rosalind-6 Research · not in the overall index yet
NewLFM3-1.7B, LFM3-1.7B-Base, LFM3-Omni-2B, LFM3-VL-2B · not in the overall index yet
PriceDeepSeek-V4-Flash cheaper $0.13 / $0.26 · before $0.14 / $0.28
PriceMiMo-V2.5 cheaper $0.14 / $0.28 · before $0.168 / $0.336
PriceMiMo-V2.5-Pro more expensive $0.48 / $1.80 · before $0.435 / $0.87
PriceMuse Glimmer 30B cheaper $0.30 / $1.20 · before $0.35 / $1.50
PriceQwen 3 235B A22B Instruct 2507 cheaper $0.09 / $0.55 · before $0.20 / $0.60
PriceQwen 3 Next 80B A3B Thinking more expensive $0.15 / $1.20 · before $0.0975 / $0.78
NewNemotron 3.5 Bolt E4B · not in the overall index yet
NewFlint · not in the overall index yet
NewMercury 3 · not in the overall index yet
PriceGemini 3.7 Flash more expensive $0.825 / $4.125 · before $0.75 / $3.75
PriceGPT-5.6 Luna cheaper $0.20 / $1.20 · before $0.22 / $1.32
PriceNemotron 3 Super input cheaper, output more expensive $0.08 / $0.45 · before $0.085 / $0.40
NewHy-Memory-30B-A3B, Nemotron 3.5 Bolt 74B A9B, Sage-A1.5 · not in the overall index yet
PriceClaude Sonnet 5 cheaper $2.20 / $11 · before $2.50 / $12.50
NewMiniCPM-V 4.7 1B, MiniCPM-V 4.7 35B-A3B · not in the overall index yet
PriceQwen 3 235B A22B Thinking 2507 cheaper $0.30 / $2.90 · before $0.30 / $3
PriceMistral Medium 3.5 more expensive $1.50 / $7.50 · before $0.40 / $2
PriceGPT OSS 120B more expensive $0.10 / $0.50 · before $0.07 / $0.30
PriceLlama 3.1 8B more expensive $0.03 / $0.05 · before $0.02 / $0.05
PriceLlama 4 Maverick cheaper $0.20 / $0.60 · before $0.20 / $0.80
PriceLlama 4 Scout more expensive $0.11 / $0.34 · before $0.10 / $0.30
PriceQwen 3 32B cheaper $0.14 / $0.40 · before $0.14 / $0.57
PriceQwen 3 Coder 30B A3B cheaper $0.10 / $0.30 · before $0.15 / $0.60
PriceQwen 3 Coder Next cheaper $0.11 / $0.80 · before $0.20 / $1.50
PriceQwen 3.7 Flash more expensive $0.0345 / $0.1495 · before $0.03 / $0.13
NewFalcon-Emirati · not in the overall index yet
PriceGemma 4 31B more expensive $0.20 / $0.40 · before $0.15 / $0.40
PriceGLM-4.7 cheaper $0.42 / $2.20 · before $0.60 / $2.20
NewKolibri 1 · not in the overall index yet
NewIndex-Translate-2B, Index-Translate-35B-A3B Preview, Index-Translate-9B · not in the overall index yet
NewMercury Voice · not in the overall index yet
PriceNemotron 3.5 Lightning cheaper $0.06 / $0.16 · before $0.08 / $0.20
Frequently asked questions about the LLM comparison
How the benchmark matching, winners, and data freshness work.
There is no single best LLM for every workload. A coding agent, a low-cost extraction pipeline, and a long-context research task reward different strengths. Choose two models above to compare. The matchup page shows only shared measurement series and keeps benchmarks separate from pricing, context, architecture, and availability.
It compares large language models using shared benchmark results, pricing, technical specifications, availability, and source dates.
The comparison does not merge different benchmark versions, tasks, metrics, cohorts, reasoning levels, or agent setups into one score. A row appears only when both model results belong to the same documented measurement series.
No. A measured lead applies to the documented task and setup. Price, latency, context size, tool access, and your workload can change the practical choice.
The benchmark snapshot was retrieved on November 12, 2026. Volatile model facts also include their own source and verification date.
The overall index is the Gradually AI benchmark meta-score (method benchmark-meta-score-v5), the same value shown in the AI model directory. It is not a benchmark of its own. A one-factor statistical model estimates one shared strength per model from all included test series by Artificial Analysis, Epoch AI benchmark run dataset snapshot, LMArena, Princeton HAL, Scale AI, and Vals AI. For the overall value, core areas carry 75% (coding, reasoning, mathematics, knowledge, agents, general, and multimodal) of the weight, specialist domains 25% (finance, law, and medicine). Supplementary areas such as cybersecurity or search also inform the shared-strength estimate, but have no separate area weight in the overall value. Price, speed, and context window are not included. The value is given in points, much like an IQ score: a fixed reference group of models averages 100 points and 15 points are one standard deviation of that group. The scale is open at both ends and is not a percentage.
Within each series all values are standardized, which makes different units comparable. Percentages and rates are stretched first, so gaps near 0 and 100% are not compressed. The model also estimates how sharply each series separates the models and weights it accordingly. Area scores move away from the overall strength only as far as the model's own results in that area support. Missing results are never entered as zero. Differing published model settings within a result series remain disclosed and make the value provisional.
A model receives an overall score with at least 8 test series from at least 2 independent evaluators and 3 areas. It is ranked once its standard error is small enough. Currently 95 models hold a rank. The rank follows the order of the scores among all ranked models, such as rank 14 of 95. Whether two models are level is checked separately with 200 recalculations with randomly redrawn test families. Where a model has no own results in a core area, its overall strength stands in, and the note “estimated” names those areas.
The database combines 30 sources. Its data was retrieved on November 12, 2026. Only reviewed independent evaluators count toward the overall index, currently Artificial Analysis, Epoch AI benchmark run dataset snapshot, LMArena, Princeton HAL, Scale AI, and Vals AI. Vendor-reported results do not count.
The database contains published result series for areas such as coding, reasoning, mathematics, knowledge, finance, law, agentic tasks, and multimodal capabilities. Individual leaderboards and matchup tables keep their documented task, metric, and cohort separate.
The matchup table lists each developer’s base rate per 1 million tokens for input and output. Tiered, regional and special rates appear as notes below it, so that unlike offers are not treated as equivalent. Caching and batch discounts are left out.
Many providers publish neither the total nor the active parameter count. In that case, we show no value. We do not estimate figures from rumors or older model generations.
The log only takes over a new price once it has held for three days in a row in the daily price index. Shorter swings appear there only for models in the overall index and are marked as not confirmed yet. Prices right after a model launch do not count as a change, because rates often settle during the first days.
Two models are level when the gap in the overall index is not certain: across 1,000 recalculations with randomly redrawn test families, neither model is ahead in at least 95% of them, or there is too little data. The provider section therefore puts every provider whose best model is level with the leading model into the top group.
Changelog
The latest updates and improvements to our LLM comparison
November 12, 2026v1.21
Compare up to five LLMs on this page
- The multi-model comparison is now a section of this page, directly below the overall index
- Alternatives to a model are now listed in its model profile
November 11, 2026v1.20
Overall index, version 5
- Models with few test series are pulled less towards the middle and now land close to their value with full data
- Gemini 4 Carbon is now in second place
- Five models no longer have a place in the overall index
- Added a block with data and feeds
November 2, 2026v1.19
Recommendations by popularity
- Duel pages, model profiles, vendor pages and this hub now recommend comparisons in one shared block, ordered by how often people compare the models
- Each model's direct predecessor and successor are always listed, and every block has a picker for your own comparison
Show older versions (19)
November 2, 2026v1.18
One rank per model
- Ranks in the overall index and in the areas now show one place in score order, such as rank 14 of 95, instead of a range; whether two models are level is still decided by the 200 recalculations
October 28, 2026v1.17
Intelligence vs. price, provider race and recent changes
- Shows the overall index as columns in the providers' colors, with provider logos and a model picker; the full list with rank ranges stays below
- Replaced the price-performance map with the intelligence vs. price chart: overall index against blended price, the most attractive area and the line of the best intelligence for the price
- Added a provider section with the best model of every provider that has at least three models in the index
- Added a log of new models and confirmed API price changes from the last 45 days
- The former AI leaderboard now redirects to the overall ranking on this page
- Replaced the radar chart question in the FAQ with questions on price changes and on “level”
- Removed Claude Mythos 5 and Mythos 5.1 from the comparison because neither model is publicly available; their former matchup pages redirect to this page
October 20, 2026v1.16
One overall index for the whole site
- The overall index is now the benchmark meta-score v3 and matches the AI model directory
- A one-factor model combines every reviewed test series; core areas carry 75% and specialist domains 25% of the overall weight
- Every score shows its standard error, and rank ranges come from 200 recalculations with randomly redrawn test families
October 17, 2026v1.15
Overall index on the widest result series again
- Per-model run settings such as temperature or reasoning effort no longer split an evaluator's result series
- Explained in the FAQ which settings may differ within a result series
October 13, 2026v1.14
Popular matchups as ranked cards
- Showed the 20 most popular matchups as numbered cards with the overall index and input price of both models
- Moved the index cohorts and their source from the leaderboard into the FAQ
October 12, 2026v1.13
Price-performance map, task rankings, and index methodology
- Added a price-performance map with an efficiency frontier and the best choice per budget tier
- Added task rankings for five use cases and an all-rounder matrix
- Added quick lists for the latest releases, cheap and strong models, and the largest context windows
- Explained the methodology of the overall index in the FAQ
September 30, 2026v1.12
Claude Sonnet 5.5 added
- Added Claude Sonnet 5.5 with official specifications, pricing, context window, knowledge cutoff, and source links
- Expanded the comparison to 50 models and 1,225 head-to-head matchups, including GPT-6.1 Sol, Grok 4.7, and DeepSeek-V4.1-Flash
September 22, 2026v1.11
Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna added
- Added Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna with official specifications, pricing, context windows, knowledge cutoffs, and source links
- Expanded the comparison to 46 models and 1,035 head-to-head matchups
September 22, 2026v1.10
46-model roster and provider-balanced hosting prices
- Expanded the comparison to 46 models and 1,035 head-to-head matchups
- Switched hosting-market tariffs to the daily provider-balanced OpenRouter observation, so profiles and comparison pages show the same figure as the price index
- Refreshed deprecation and lifecycle status against current provider pages
September 4, 2026v1.9
GPT-6 Astra added
- Added GPT-6 Astra with official specifications, pricing, context window, availability, and source links
- Expanded the comparison to 36 models and 630 head-to-head matchups
September 3, 2026v1.8
Muse Spark 1.3 and Qwen 3.8 Max 0902 added
- Added Muse Spark 1.3 with current pricing, context, availability, and source links
- Updated Qwen 3.8 Max to the 0902 snapshot and added it to the comparison roster
- Expanded the comparison to 35 models and 595 head-to-head matchups
September 2, 2026v1.7
Gemini 3.8 Flash added
- Added Gemini 3.8 Flash with official specifications, pricing, knowledge cutoff, and source links
- Expanded the comparison to 33 models and 528 head-to-head matchups
September 2, 2026v1.6
Claude 5.1 and GLM-5.3 added
- Added GLM-5.3 with official specifications, pricing, availability, and source links
- Expanded the comparison to 32 models and 496 head-to-head matchups, including Claude Fable 5.1 and Mythos 5.1
August 29, 2026v1.5
GLM-5.3-Flash and cleaner leaderboards
- Added GLM-5.3-Flash with official specifications, pricing, and source links
- Expanded the comparison to 29 models and 406 head-to-head matchups
- Removed models without a published score from leaderboard rows
August 28, 2026v1.4
Complete and cross-benchmark leaderboards
- Added all 28 comparison models to every individual benchmark leaderboard
- Added an orientation index from four setup-comparable benchmark cohorts
- Excluded ambiguous model variants and kept missing results visible without estimation
August 24, 2026v1.3
Related matchups, provider icons, and expanded FAQ
- Expanded each LLM matchup page to 14 related matchups and placed the hub button below the list
- Added local provider icons to all 28 model badges without placeholders
- Expanded the FAQ with benchmark, pricing, parameter, and radar methodology details
August 21, 2026v1.2
Benchmark leaderboards and capability profiles
- Added five separate leaderboards for overall preference, reasoning, coding, agents, and finance
- Added evidence-based capability radar charts to the head-to-head comparison pages
- Kept every ranking within one published cohort, task, metric, and source snapshot
August 18, 2026v1.1
FAQ, changelog, and internal links
- Added an FAQ about benchmark coverage, methodology, and database freshness
- Added this changelog so material updates remain visible
- Added the hub and all matchup pages to navigation, the sitemap, and related comparisons
August 17, 2026v1.0
Initial release
- Published 28 current models and 378 head-to-head matchups
- Combined shared benchmarks, API pricing, specifications, and primary sources
- Added the comparison picker and ten prioritized matchups