The main LLMs today and what each is good at
A plain guide to the current Claude, GPT, Gemini, Llama, Qwen, DeepSeek and Mistral models, with list prices and what each vendor says they are for.
Updated 2026-11-19. "(source N)" refers to the numbered list at the end.
How to read this list
Model names change every few months, so treat this as a snapshot from 19 November 2026. Prices are list prices per million tokens of input and output, in US dollars, taken from each vendor's own pages. A million tokens is roughly the length of several long novels.
Where we say what a model is good at, we are reporting how the vendor positions it. Your own tests on your own tasks matter more than any label.
Anthropic Claude
Anthropic's current line-up has four main models, all with a 1 million token context window and up to 128,000 tokens of output (source 1).
- Claude Fable 5.2 ($8 in, $40 out): the top tier, aimed at hard reasoning and long agent tasks that run for many steps.
- Claude Opus 5.5 ($4 in, $20 out): Anthropic's suggested starting point for most work, including long coding and knowledge tasks.
- Claude Sonnet 5.5 ($2 in, $10 out): pitched as the best mix of speed and smarts.
- Claude Haiku 5.5 (from $0.10 in, $0.50 out): the fastest, meant for high-volume jobs like sorting, extraction and routing.
Restricted models
A Mythos line also exists but is only open to vetted organisations, for example through a cyber security programme (source 1).
OpenAI GPT
OpenAI's API lists four flagship models, each with about a 1.05 million token context window and 128,000 tokens of output (source 2).
- GPT-6.2 Astra ($8 in, $40 out): OpenAI's most capable model, for the hardest projects.
- GPT-6.2 Sol ($2 in, $10 out): the everyday model, including for running agents in volume.
- GPT-6.1 Terra ($1 in, $5 out): a fast model pitched at visual reasoning, computer use and clear writing.
- GPT-6 Luna ($0.10 in, $0.50 out): the low-cost option for narrow, high-volume jobs.
Google Gemini
Google's developer docs list four Gemini 4 models, each with a 1 million token context (source 3). Note the price changes coming soon (source 4).
- Gemini 4 Carbon: Google's strongest reasoning model, for complex software, science and security work. $2 in and $10 out until 30 November 2026, then $4 and $20.
- Gemini 4 Argon: an efficient reasoning model that aims to be faster and use fewer tokens than Carbon. Same pricing as Carbon.
- Gemini 4 Neon ($3 in, $15 out): pitched at quick, start-to-finish task work.
- Gemini 4 Hydrogen ($1.50 in, $7.50 out): for everyday coding and agent work where speed matters.
- Gemini 3.5 Flash-Lite ($0.30 in, $2.50 out): an older, cheap option still recommended for new low-cost projects.
Open-weight families
These models publish their weights, so you can run them yourself or rent them from a host. Licences differ, so read them before you build.
- Meta Llama: Llama 4 Scout and Maverick are still the newest open Llama models on Meta's Hugging Face page. Their files were last updated in 2025 (source 9). Many tools support them, but they are no longer at the front of the pack.
- Alibaba Qwen: Qwen3.8 is the newest open family. The 27B version is released under Apache 2.0, reads images and video, and handles 262,000 tokens natively, stretchable to 1 million (source 8).
- DeepSeek: DeepSeek V4.1 Flash is open under the MIT licence with a 1 million token context (source 7). Through DeepSeek's own API it costs $0.15 to $0.30 per million input tokens and $0.60 to $1.20 output, cheaper off-peak (source 6).
- Mistral: Mistral Large 4 ships under Mistral's own open-weight licence, while Mistral Small 4 and the small Ministral 3 models use Apache 2.0 (source 5).
What the gap looks like
The top labs are close. Stanford's 2026 AI Index found that, as of March 2026, the best US model led the best Chinese model by just 2.7% on its benchmark mix (source 10). In practice that means an open model from China may be good enough for many business tasks, at a fraction of the price.
The pricing spread is huge. The cheapest models on this page cost about 80 times less per token than the most expensive. For simple, repetitive jobs, the small models are usually the right call.
A sensible default
If you are starting today, pick one mid-tier model, such as Claude Sonnet 5.5, GPT-6.2 Sol or Gemini 4 Hydrogen. Build your task, write 20 to 50 test cases, and run them. Then try one tier down to see if quality holds, and one tier up only if it does not.
Keep your prompts and tools vendor-neutral where you can. Models are replaced often, and being able to switch in an afternoon is worth more than any single benchmark.
Sources
Written by Agents Lookup in our own words. Facts come from these pages:
- 1. Anthropic: Models overview: platform.claude.com/docs/en/about-claude/models/overview
- 2. OpenAI: Models: developers.openai.com/api/docs/models
- 3. Google AI Studio: Gemini models: aistudio.google.com/docs/models
- 4. Google AI Studio: Gemini API pricing: aistudio.google.com/docs/pricing
- 5. Mistral AI: Models overview: docs.mistral.ai/getting-started/models/models_overview
- 6. DeepSeek: Models and pricing: api-docs.deepseek.com/quick_start/pricing
- 7. Hugging Face: DeepSeek-V4.1-Flash model card: huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- 8. Hugging Face: Qwen3.8-27B model card: huggingface.co/Qwen/Qwen3.8-27B
- 9. Hugging Face: Meta Llama organisation page: huggingface.co/meta-llama
- 10. Stanford HAI: Inside the AI Index, 12 takeaways from the 2026 report: hai.stanford.edu/news/inside-the-ai-index-12-takeaways-from-the-2026-report