Makers' prices beside independent test scores, the best model for each kind of task, and which ones you can download and run on your own computer.
Decided by independent testers, not the makers. Each card shows the top three on its main test and who leads the others. These lists change every few weeks, so each carries its date.
Many tests rolled into one score, and work tasks weighted by their share of the US economy.
Epoch Capabilities IndexEpoch AI · Oct 20, 2026
Epoch Capabilities Index (ECI) · CC BY 4.0
Building websites people prefer, fixing real software, and making working apps from scratch.
LMArena WebDevLMArena · Oct 19, 2026
Which answers people prefer in blind votes, and long back-and-forth conversations.
LMArena textLMArena · Oct 15, 2026
Carrying out long tasks on their own: real remote freelance work, a computer's command line, and playing games with a mouse.
LMArena Agent ArenaLMArena · Oct 19, 2026
Leads 3 of these 4 tests: Claude Opus 5.5
Math questions people vote on, checked proofs, and expert-level exam questions.
LMArena mathLMArena · Oct 15, 2026
Questions about pictures, and analysis of long documents.
LMArena visionLMArena · Oct 19, 2026
Ties are common: LMArena's and Epoch's margins of error often overlap for the top few. ProofBench has three models at 100%, listed in the order Vals gives. Coding scores on SWE-Bench Pro depend on the coding tool each model ran in, as Scale lists it.
Prices are what each maker charges developers per million tokens (a token is a word or part of a word), going in and coming out. People using a chat app pay a monthly plan instead, below. Prices read Oct 20, 2026.
| Model | Released | Price per million tokensin / out | Reads at oncetokens | Download | Epoch scoreECI | LMArenatext rank |
|---|---|---|---|---|---|---|
Claude Opus 5.5Anthropic · USFor long-running agentic coding and knowledge work | Sep 22, 2026 | $4 / $20 | 1M | No | 166.8#1 | #3 |
GPT-6 AstraOpenAI · USOur most capable model for the most demanding work. | Sep 3, 2026 | $10 / $50 | 1M | No | 166.4#2 | Below #20 |
GPT-6.1 SolOpenAI · USNear-Astra performance for complex work at a lower cost. | Sep 29, 2026 | $2 / $10 | 1M | No | 166.3#3 | Below #20 |
Claude Fable 5.1Anthropic · USFor demanding reasoning and long-horizon agentic work | Sep 1, 2026 | $10 / $50 | 1M | No | 164.7#4 | #6 |
Gemini 4 ArgonGoogle · USOur flagship reasoning model with a unified 1M context window and long decodes up to 1M tokens. | Sep 30, 2026 | $2 / $10Introductory price | 1M | No | 164.6#5 | #1 |
Claude Sonnet 5.5Anthropic · USHigh-performance model for coding and agents | Sep 28, 2026 | $2 / $10 | 1M | No | 164.4#6 | Below #20 |
Claude Haiku 5.5Anthropic · USFor high-volume, latency-sensitive tasks such as classification, extraction, and routing | Oct 7, 2026 | $0.10 / $0.50 | 1M | No | 159.0#14 | Below #20 |
Kimi K3Moonshot AI · ChinaKimi K3 is an open-weight, native multimodal agentic model and our most capable model to date. | Not stated | $3 / $15 | 1M | YesKimi K3 License | 157.4#17 | #18 |
Gemini 3.8 FlashGoogle · USGemini 3.8 Flash is our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows—all with the speed and cost efficiency of Flash. | Sep 2, 2026 | $0.75 / $3.75Introductory price | 1M | No | 157.2#19 | #9 |
Qwen3.8-MaxAlibaba (Qwen) · ChinaA 2.4-trillion-parameter MoE flagship with a major leap in coding and office productivity | Sep 2, 2026 | $2 / $6 | 1M | No | 156.9#20 | Below #20 |
Muse Spark 1.3Meta · USMuse Spark 1.3 is trained for agentic workflows and optimized for competitive coding performance. | Not stated | $1.25 / $4.25 | 1M | No | 156.8#22 | #10 |
GPT-6 LunaOpenAI · USOur most efficient model for focused, high-volume tasks. | Sep 22, 2026 | $0.10 / $0.50 | 1M | No | 156.3#27 | Below #20 |
GLM-5.3Z.ai (Zhipu AI) · ChinaGLM-5.3 is Z.ai's latest flagship model, delivering comprehensive advancements in complex software engineering | Not stated | $1.40 / $4.40 | 1M | Yesglm-5.3 | 155.9#30 | Below #20 |
DeepSeek-V4.1-Flash-1017DeepSeek · ChinaStronger benchmark performance, faster task completion, and shorter outputs than the previous version. | Oct 17, 2026 | $0.30 / $1.20Peak-hours price | 1M | YesMIT | Not scored | Below #20 |
Muse Spark OpenMeta · USThe most capable US open-weight model to date. Build your way, anywhere. | Oct 15, 2026 | Free to download; hosts set prices | 1M | YesMuse Model License | Not scored | Below #20 |
Muse WonderMeta · USFrontier coding, long-running agents, and advanced computer use. | Oct 12, 2026 | $4 / $20 | 1M | No | Not scored | #2 |
Mistral Large 4Mistral AI · FranceMistral Large 4 is a state-of-the-art, open-weight, general-purpose multimodal model with a granular Mixture-of-Experts | Oct 6, 2026 | $0.68 / $2.09Sale price | 1M | No | Not scored | Below #20 |
Grok 4.7SpaceXSI · USGrok 4.7 is our most capable model for coding and knowledge work. | Sep 21, 2026 | $2 / $6 | 500K | No | Not scored | Below #20 |
Claude Mythos 5.1Anthropic · USOur most capable model for cybersecurity and biology research. | Sep 1, 2026 | No public price | Not stated | No | Not scored | Below #20 |
| DeepSeek-V4-Pro-0813DeepSeek · China | Aug 13, 2026 | $1.32 / $3.96Peak-hours price | 1M | YesMIT | Not scored | Below #20 |
Qwen3.8-2.4T-A95BAlibaba (Qwen) · ChinaFor the first time, Qwen3.8 brings a Qwen-Max-class model to open release. | Aug 8, 2026 | Free to download; hosts set prices | 262K | Yesqwen3.8-max | Not scored | Below #20 |
Qwen3.8-27BAlibaba (Qwen) · Chinaa native vision-language model that understands images and videos, with flexible thinking control | Aug 5, 2026 | Free to download; hosts set prices | 262K | YesApache 2.0 | Not scored | Below #20 |
Muse GlimmerMeta · USBuilt to enable always-on, local agents, available on Hugging Face. | Aug 2026 | Free to download; hosts set prices | 131K | YesApache 2.0 | Not scored | Below #20 |
Gemini 3.5 Flash-LiteGoogle · USOur fastest, most cost-effective 3.5 model for high-throughput execution. | Jul 21, 2026 | $0.30 / $2.50 | 1M | No | Not scored | Below #20 |
Mistral Medium 3.5Mistral AI · FranceOur frontier-class multimodal model optimized for agentic and coding use cases. | Apr 28, 2026 | $1.50 / $7.50 | 256K | YesModified MIT | Not scored | Below #20 |
Gemma 4 31BGoogle · USToday, we are introducing Gemma 4 — our most intelligent open models to date. | Apr 2, 2026 | Free to download; hosts set prices | 256K | YesApache 2.0 | Not scored | Below #20 |
Gemma 4 26B A4BGoogle · USToday, we are introducing Gemma 4 — our most intelligent open models to date. | Apr 2, 2026 | Free to download; hosts set prices | 256K | YesApache 2.0 | Not scored | Below #20 |
gpt-oss-120bOpenAI · USproduction, general purpose, high reasoning use cases | Aug 5, 2025 | Free to download; hosts set prices | 128K | YesApache 2.0 | Not scored | Below #20 |
MiniMax-M3MiniMax · ChinaMiniMax-M3 is a native multimodal model with 1M context. | Not stated | $0.30 / $1.20 | 1M | Yesminimax-community | Not scored | Below #20 |
Makers' prices and descriptions, so company-reported; quoted lines are each maker's own words. Many charge more for very long prompts, and less for repeated prompts and batch work. Google's two Gemini prices are introductory and double from December 1 and January 1; Mistral Large 4's is a sale price (regular $1.36 / $4.18), and its download is promised but not out yet; DeepSeek's is the peak-hours rate, half off-peak. Epoch score: Epoch AI, Epoch Capabilities Index, Oct 20, 2026, CC BY 4.0, top 30 shown. LMArena: text leaderboard, Oct 15, 2026, top 20.
Each dot is a model Epoch AI scores, placed by its maker's price. Higher is more capable; further left is cheaper. The thin lines are Epoch's range of uncertainty: where two overlap, the scores can't be told apart.
The top score, Opus 5.5, at $20 per million output tokens.
The cheapest price among the 6 models whose range overlaps the top score: GPT-6.1 Sol, Gemini 4 Argon and Sonnet 5.5 (Gemini's is introductory).
The score of the cheapest model Epoch rates, Haiku 5.5, at $0.50 per million output tokens.
Open models you can download and use without the internet, sorted by whether they fit your machine. Pick where it would run and how much memory there is.
Sizes, speeds, machines and appsThe file takes up to three quarters of 16 GB.
Fits, with little room left. Close other apps and keep conversations short.
The file alone is larger than 16 GB. Smallest first.
and 9 larger, up to 1.51 TB
How we sort: by the size of each model's usual download (its smallest 4-bit file, from Ollama, LM Studio or Unsloth). Our rule of thumb, not the makers': a file up to three quarters of the memory runs comfortably, leaving room for the system, other apps and the conversation. Ollama says longer conversations need more memory. What makers and apps say
Monthly prices in US dollars from each company's own pages.
OpenAI
OpenAI: ChatGPT pricing page, Oct 20, 2026 · company-reported
Anthropic
Anthropic pricing page, Oct 20, 2026 · company-reported
Google: Gemini subscriptions page, Oct 20, 2026 · company-reported
SpaceXSI · The X Subscription
SpaceXSI, Oct 20, 2026 · company-reported
Mistral AI · formerly Le Chat
Mistral AI: Pricing page, Oct 20, 2026 · company-reported
Team and business plans are per person. Claude's Max plan is shown from $100. ChatGPT's own pricing page showed no amounts, so its prices come from OpenAI's help pages. Grok now comes in “The X Subscription”, launched October 12. We couldn't find plan prices for Meta AI, DeepSeek, Qwen or Kimi.
Makers' own pages for prices and specs; independent testers for every score. We don't use any company's marketing benchmark to rank models.
Not shown yet: Epoch AI's per-test tables, Terminal-Bench's and OSWorld's own leaderboards, and LiveBench, which our tools couldn't read reliably. How we check · AI and coding · AI agents
