Data checked on 15 October 2026
Models and prices, no fanboying.
Numbers from independent benchmarks (Artificial Analysis), human-vote rankings (Arena) and official price lists. Every figure has a source. In a hurry? Start from the four picks below.
58 points, also #1 in WebDev and Agents on Arena.
52 points at $0.72 per task: an eighth of Opus.
43 points at $0.21, and 240 tokens per second.
46 points, runnable on your own servers.
Intelligence vs cost
The Intelligence Index summarises several benchmarks (reasoning, maths, code, knowledge, tool use). Cost per task includes reasoning tokens: a model pricey per token can be cheap per task, and vice versa.
All models compared
Click headers to sort
| Model | Index | Cost/task | Speed | API price in/out | Context | Best for |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 Anthropic · max | 58 | $5.98 | 95 t/s | $4 / $20 | 1M | The most capable today: long agentic work, complex code, hard analysis. |
| Claude Sonnet 5.5 Anthropic · max | 56 | $5.46 | 131 t/s | $2 / $10 | 1M | Near-Opus quality at half the token price: an excellent workhorse. |
| Muse Wonder Meta · max | 52 | $4.52 | 161 t/s | $4 / $20 | 1.05M | Meta's new agent model; the Full version reaches 56 points. |
| Claude Fable 5.1 Anthropic · max | 53 | $7.63 | 71 t/s | $10 / $50 | 1M | Flagship for writing and long reasoning; the priciest of the family. |
| GPT-6 Astra OpenAI · max | 53 | $3.26 | 44 t/s | $10 / $50 | 1M | OpenAI's most powerful model: 'Pro' reasoning and very hard tasks. |
| Gemini 4 Argon Google · high | 53 | $1.99 | — | $2 / $10 | 1M | Just released: #1 on Arena Text (preliminary) at a moderate price. |
| GPT-6.1 Sol OpenAI · max | 52 | $0.72 | 51 t/s | $2 / $10 | 1M | Best value among frontier models. It powers Lumen's chatbot. |
| Muse Spark 1.3 Meta · max | 48 | $1.6 | 128 t/s | $1.25 / $4.25 | 1M | Good balance of quality, speed and price. |
| Grok 4.7 SpaceXAI (xAI) · xhigh | 46 | $3.74 | 91 t/s | $2 / $6 | 500k | Integrated with X: useful for social-media current events. |
| MiMo-V2.6-Pro Xiaomi · — · open weights | 46 | $0.13 | 47 t/s | various hosts | 1M | The best open-weights model today, with a tiny cost per task. |
| Qwen3.8 Max Alibaba (Qwen) · 0902 | 45 | $5.41 | 40 t/s | $2 / $6 | 984k | Strong at multilingual and code; Chinese ecosystem. |
| GLM-5.3 Z AI (GLM) · max · open weights | 45 | $2.01 | 66 t/s | various hosts | 1M | Open weights: can be self-hosted for confidential data. |
| Kimi K3 Moonshot (Kimi) · max · open weights | 44 | $2 | 44 t/s | various hosts | 1.05M | Open weights with a very long context. |
| Claude Haiku 5.5 Anthropic · max | 43 | $0.21 | 240 t/s | $0.10 / $0.50Up to 100K input tokens; above that $0.50 / $2.50. | 1M | Very fast and cheap: classification, extraction, high-volume chatbots. |
| Gemini 3.8 Flash Google · high | 41 | $1.24 | 147 t/s | $0.75 / $3.75Promotional until 31 Dec 2026, then $1.50 / $7.50. | 1M | Multimodal (PDF, audio, video in) and fast; has a free API tier. |
| DeepSeek V4.1 Flash DeepSeek · max · open weights | 39 | $0.27 | 180 t/s | $0.30 / $1.2Peak rate; off-peak is half price. | 1M | Open weights, fast and cheap. |
| GPT-6 Luna OpenAI · max | 38 | $0.07 | 118 t/s | $0.10 / $0.50 | 1M | ChatGPT's free-tier model: near-zero cost for simple tasks. |
| Mistral Large 4 Mistral · preview | 38 | $1.13 | — | $0.68 / $2.09Promotional; regular $1.36 / $4.18. | 524k | The European option: useful if you need an EU provider. |
| DeepSeek V4 Pro DeepSeek · max · open weights | 36 | $0.67 | 101 t/s | $1.32 / $3.96 | 1M | Open weights, good low-cost reasoning. |
API prices in USD per million tokens (input / output), standard rate. Index, cost per task, speed: Artificial Analysis, highest reasoning setting measured. Prices: provider price lists and Arena.
The price of thinking
The more it thinks, the more it knows. And costs.
The same model, GPT-6.1 Sol, at five effort levels. From low to max the index rises 10 points, but cost per task grows fivefold and response time goes from 15 seconds to almost 5 minutes. For most daily tasks, medium is the sweet spot.
Read the reasoning guide →index cost per task · total response time
Arena
What people prefer
On Arena users compare two anonymous answers and vote for the better one. Millions of votes become an Elo score, as in chess. It measures human preference, not correctness: they often coincide, not always.
- 1Gemini 4 Argon (high)preliminary1,524
- 2Muse Wonder (max)1,509
- 3Claude Opus 5.5 (high)1,508
- 4Claude Fable 5 (high)1,504
- 5Claude Fable 5.1 (max)1,502
- 6Gemini 3.8 Flash (high)preliminary1,496
- 7Muse Spark 1.3 (max)1,494
- 8Kimi K3 (max)1,487
arena.ai · Text ↗Elo points15 October 2026
- 1Claude Opus 5.5 (max)1,811
- 2GPT-6 Astra (max)1,780
- 3Claude Sonnet 5.5 (xhigh)1,774
- 4GPT-6.1 Sol (max)1,760
- 5Claude Fable 5.1 (max)1,743
- 6Gemini 4 Argon (high)preliminary1,678
arena.ai · Web dev ↗Elo points15 October 2026
- 1Claude Opus 5.5 (max)17.18%
- 2GPT-6.1 Sol (max)14.78%
- 3Claude Fable 5.1 (max)14.12%
- 4Gemini 4 Argon (high)13.73%
- 5GPT-6 Astra (max)13.69%
- 6Claude Sonnet 5.5 (max)12.55%
arena.ai · Agents ↗Net Improvement: composite of success, satisfaction and steerability15 October 2026
- 1GPT Image 2.5 Sunburstpreliminary1,425
- 2GPT Image 2.5 Flarepreliminary1,398
- 3GPT Image 2 (medium)1,383
- 4Grok Imagine Image 2.0preliminary1,334
- 5MAI Image 2.6 (Microsoft)1,332
- 6Nano Banana 2.1 (Gemini)preliminary1,328
- 7Reve 2.11,302
arena.ai · Images ↗Elo points15 October 2026
- 1Gemini Omni 1.1 Flash1,516
- 2Gemini Omni Flash1,513
- 3FLUX 3 Video (Black Forest Labs)preliminary1,493
- 4Grok Imagine Video 1.5preliminary1,492
- 5Seedance 2.0 (ByteDance)1,479
- 6Wan 3.0 (Alibaba)1,476
- 7Muse Video (Meta)1,456
arena.ai · Video ↗Elo points15 October 2026
Calculator
What would it cost you, via API?
Set how many requests you make and how long they are. Uses standard token prices (no caching or batch discounts) and 1.3 tokens per English word.
Example: a small-site chatbot gets 500-2,000 questions a month; an automation reading 50 invoices a day makes about 1,000 requests a month.
Subscriptions
If you use the apps, not the APIs
Monthly prices shown to Italian customers on official pages. Claude is listed in dollars, excluding VAT.
ChatGPT
pricing ↗- Freefree
GPT-6 (Luna) with limits, limited images and deep research.
- Go€8
More messages, uploads and images. May include ads.
- Plus€23
Advanced GPT-6 reasoning, Codex, projects, scheduled tasks.
- Profrom €103
Pro reasoning on GPT-6 Astra, always-on Dot agent, top limits.
Claude
pricing ↗- Freefree
Sonnet and Haiku, web search, Artifacts, Claude Docs, Slides and Design.
- Pro$20
Opus, Claude Code, projects, Research, Claude in Chrome and Microsoft 365. $17/mo billed yearly.
- Max 5x$100
5x Pro usage, Fable access, $100/mo API credits.
- Max 20x$200
20x Pro usage, $200/mo API credits.
Google AI (Gemini)
pricing ↗- Freefree
Flash-Lite model, Deep Research, Gemini Live, Gemini Notebook, 15 GB.
- AI Plus€4.99
Flash model, 2x limits, 200 Flow video credits, 400 GB.
- AI Pro€21.99
Gemini 4 Argon and Pro, 4x limits, Antigravity, 1,000 Flow credits, 5 TB.
- AI Ultrafrom €99.99
5x or 20x Pro limits (€99.99 or €219.99), Project Genie.
My advice in three lines
- Occasional use: free plans are enough. Keep two (e.g. ChatGPT and Claude) and compare answers on important things.
- Daily work use: one ~€20/month plan. Claude Pro if you write and analyse documents, ChatGPT Plus for images, voice and data, Google AI Pro if you live in Gmail and Drive (and need 5 TB).
- Heavy use (agents, coding all day): €100+ plans only pay off if you often hit base-plan limits. Measure first.
