One table of 25 current AI models from 10 vendors: what kind each is, what it costs and what it is good for, then three charts of Artificial Analysis's published scores. Last verified 2026-11-13; verify before you commit.
Snapshot verified 2026-11-13 against vendor pricing pages and Hugging Face model cards. Model names and prices change often.
Hosted prices are passthrough costs pulled from OpenRouter; each vendor group links its provider page. Open weights are free to download and self-host - hosting them always has a price. An asterisk (*) marks a promotional price at verification.
How to read it.Frontier models are used through the vendor's paid API; open-weight models can be downloaded and run on your own hardware, so the price you pay is the hardware or hosting. Prices are standard rates; batch, caching and long prompts change them. *Claude Haiku 5.5 prompts over 100K tokens cost $0.50 / $2.50. "Not checked" means we did not read that figure on the vendor page, and "Name only" rows are models we confirmed exist on the vendor's own site but whose figures we could not read there yet: follow the vendor link in the group header. The table carries no benchmark scores on purpose: they change monthly and will not tell you how a model does on your work. The charts below show Artificial Analysis's independent indices as context: use them to pick a shortlist, then test the shortlist on your own tasks.
What it measures: one weighted score from nine of Artificial Analysis's own tests of agentic knowledge work, coding, science and general knowledge (Intelligence Index v5.0.1). Higher is better. It is a broad average, not a score for your codebase.
Claude Fable 5.2 Max64.6
Gemini 4 Carbon High60.0
Claude Opus 5.5 Max59.1
GPT-6 Astra Max56.9
GPT-6.1 Sol Max54.6
Claude Sonnet 5.5 Max54.5
Gemini 4 Hydrogen Xhigh53.9
GPT-6.1 Terra Max51.3
Kimi K3.1 Pro46.5
Claude Haiku 5.5 Max36.5
DeepSeek V4.1 Flash Max32.6
GLM-5.3 Max31.7
GPT-6 Luna Max27.9
Mistral Large 4 High26.1
Qwen3.8 27B Xhigh20.1
Artificial Analysis Intelligence Index v5.0.1, sorted highest first, drawn on a 0 to 70 scale.
Capabilities
What it measures: six industry indexes Artificial Analysis builds from subsets of the same tests, one bar per index. Engineering is the closest to coding-agent work, so the rows are sorted by it. Higher is better.
Model
Finance & Accounting
Strategy & Ops
Legal
Healthcare & Medical
Engineering
Economics
Claude Fable 5.2 Max
64.7
67.4
63.3
65.4
72.6
66.5
Claude Opus 5.5 Max
59.4
61.4
58.0
57.9
65.3
60.3
Gemini 4 Carbon High
62.6
60.9
62.0
56.8
65.3
59.6
GPT-6 Astra Max
57.4
58.0
55.6
53.7
61.8
57.2
Gemini 4 Hydrogen Xhigh
58.1
57.7
56.7
53.3
60.1
55.0
GPT-6.1 Sol Max
56.2
57.6
53.7
52.0
59.4
56.3
Claude Sonnet 5.5 Max
55.8
58.1
52.2
54.2
58.5
54.1
GPT-6.1 Terra Max
49.0
50.0
46.4
44.7
54.7
48.5
Kimi K3.1 Pro
53.3
57.6
49.2
51.6
50.9
49.9
Claude Haiku 5.5 Max
43.6
39.1
38.7
37.5
42.5
42.7
GLM-5.3 Max
42.1
47.4
35.8
39.3
37.6
37.3
DeepSeek V4.1 Flash Max
42.2
49.2
37.7
41.3
37.6
38.5
Mistral Large 4 High
39.1
43.3
34.6
36.7
33.4
35.6
GPT-6 Luna Max
33.9
43.2
32.2
34.8
33.3
34.5
Qwen3.8 27B Xhigh
35.5
39.0
33.2
34.6
28.1
33.7
Artificial Analysis Capability Indexes, sorted by Engineering, drawn on a 0 to 80 scale.
Cost
What it measures: the average US dollars Artificial Analysis paid per Intelligence Index task, from each model's token prices times the tokens it actually used, reasoning and cache included. Lower is better, so the cheapest is at the top. A low per-token price can still cost more per task when a model uses many tokens: Artificial Analysis lists Qwen3.8 27B at $0.50 / $3 per million tokens, a quarter of GPT-6.1 Sol's $2 / $10, yet it cost $3.53 per task to Sol's $1.27.
GPT-6 Luna Max$0.15/task
Claude Haiku 5.5 Max$0.41/task
DeepSeek V4.1 Flash Max$0.60/task
GPT-6.1 Terra Max$0.96/task
GPT-6.1 Sol Max$1.27/task
Mistral Large 4 High$2.06/task
Gemini 4 Hydrogen Xhigh$3.16/task
Qwen3.8 27B Xhigh$3.53/task
Kimi K3.1 Pro$3.68/task
GLM-5.3 Max$5.01/task
Gemini 4 Carbon High$5.48/task
GPT-6 Astra Max$6.24/task
Claude Sonnet 5.5 Max$10.02/task
Claude Opus 5.5 Max$10.04/task
Claude Fable 5.2 Max$26.54/task
Artificial Analysis Cost per Intelligence Index Task (weighted average, US dollars per task), sorted cheapest first, drawn on a linear scale from $0.
15 models tracked by Artificial Analysis as of 2026-11-13. Each is charted at the setting Artificial Analysis lists (Max, High or Xhigh); hover a bar for its exact name there. In our table but not in Artificial Analysis's public model view, so no bar: Claude Fable 5.1, gpt-oss-120b / 20b, Gemini 4 Argon, Gemma 4 (up to 31B), GLM-5.3-Flash, DeepSeek V4 Pro, Llama 4 Scout / Maverick, MiniMax M3.1 Flash.
Intelligence, capability and cost figures are from Artificial Analysis, last verified 2026-11-13. We show their published indices; we do not run our own benchmarks.
What Is the Difference?
A frontier model is one of the most capable models available, offered only as a service: you pay per token and never see the model itself. An open-weight model is one whose trained weights you can download and run yourself, under the license the publisher chose. Frontier models are usually the most capable and the least work; open-weight models give you control over where the data goes, what it costs at scale and what you may change.
Which Should You Pick?
If you are prototyping or using a coding agent, start with a frontier model. It is the fastest way to learn what is possible, and most coding agents are tuned for one.
If your data cannot leave your machines, for legal or contractual reasons, use an open-weight model you run yourself. Aider and Cline, among others, can point at a local model server.
If you make millions of similar requests, price an open-weight model on hosted or owned hardware against the frontier per-token cost.
If you need to fine-tune deeply or ship a model inside your product, an open-weight model with a permissive license is usually the only option.
Whatever you pick, measure it on your own tasks with a small eval, not on someone else's leaderboard. The agent reads the same AGENTS.md either way.
How Much Hardware Does an Open-Weight Model Need?
About 2 GB of memory per billion parameters at 16-bit precision, or about half a gigabyte when quantized to 4 bits, plus room for the context. A 27B model at 4 bits fits in roughly 16 to 20 GB, one high-end consumer GPU or a laptop with plenty of unified memory. Mixture-of-experts models run only part of their weights per token, so they are faster than their size suggests, but all the weights must still fit in memory.
Which Frontier Models Are Current?
As of 2026-11-13, from the vendors' own pages:
Claude (Anthropic): Fable 5.2 is the flagship; Opus 5.5, Sonnet 5.5 and Haiku 5.5 step down in price. Fable 5.1 is still active but no longer in the latest lineup. All five have a 1M-token context. Source: Anthropic models overview.
GPT (OpenAI): GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna, from most to least expensive. Source: OpenAI API pricing.
Gemini (Google): the Gemini 4 line, including Carbon, Argon and Hydrogen. Google's pages need JavaScript and could not be read for this snapshot, and Google has announced price changes from December 1, 2026, so check Gemini API pricing directly.
How Do They Compare Side by Side?
Frontier (closed)
Open-weight
Who it is for
Teams that want the most capable model with no infrastructure work.
Teams that need data on their own hardware, predictable cost at volume, offline use or fine-tuning.
What it costs
Pay per token, from about 0.10 to 10 US dollars per million input tokens.
Open weights; free to download and self-host, and hosted inference has a passthrough price.
How you run it
An API key and an HTTPS request.
llama.cpp, Ollama, LM Studio, vLLM, or a host that serves open models.
Where your data goes
To the vendor, under its data policy.
Wherever you run it.
Upgrades
The vendor ships and retires versions on its schedule.
You choose when; old weights keep working.
Customization
Prompts, tools, sometimes vendor fine-tuning.
Full fine-tuning, quantization and distillation, within the license.
Which Open-Weight Families Matter?
As of 2026-11-13, from the publishers' Hugging Face cards and docs:
Qwen (Alibaba): Qwen3.8, for example Qwen3.8-27B, August 2026, Apache 2.0.
DeepSeek: V4 Pro and V4 Flash, with V4.1 Flash builds from October 2026 under MIT.
GLM (Z.ai): GLM-5.3 and GLM-5.3-Flash, with weights published by zai-org.
Mistral: Mistral Large 4, October 2026, under Mistral's own license.
Gemma (Google): Gemma 4, up to 31B parameters, Apache 2.0.
Llama (Meta): Llama 4 Scout and Maverick, April 2025, Llama 4 Community License, gated.
gpt-oss (OpenAI): gpt-oss-120b and gpt-oss-20b, Apache 2.0, weights only.
Open-weight is not open source: you get the weights, usually not the training data or code, and the license decides what you may do. Source: Hugging Face.
Frequently Asked Questions
Is an open-weight model the same as open source?
Not quite. Open-weight means you can download and run the trained weights. Open source usually also means the code and training data are available. Most open-weight models share weights only, under a license that may restrict use.
Are open-weight models free?
The weights are free to download under their license, but running them is not: you pay for the hardware or hosting. At low volume a frontier API is often cheaper; at high volume your own hardware can be.
Can I run a good model on a laptop?
Yes, within limits. Quantized models of roughly 7 to 30 billion parameters run on recent laptops with enough memory. They are useful for many tasks but generally less capable than the largest frontier models.
Why do model names change so often?
Labs release new versions every few months and retire old ones. That is why this page is dated and why you should check the vendor page before you build on a specific model name.
Which model is best for coding?
It depends on your language, codebase and budget, and the answer changes often. Try two or three on real tasks from your repository and measure which one produces changes you would merge.