Which model is worth using
1 Oct 2026
Labs
Ten labs have a model in this snapshot. Marks use each lab’s public colors. They are not the official logos.
OpenAI
3 models on the board. GPT-6 is three API models, not one. Astra is the flagship. GPT-6.1 Sol is the model most teams should price first. Luna is for volume.
Anthropic
4 models on the board. Claude is split by job. Fable 5.1 is the top of the range. Opus 5.5 is the flagship most people should start with. Sonnet 5.5 is the fast complement. Haiku 4.5 is the older small model.
Google DeepMind
2 models on the board. Gemini 4 Argon is the new frontier model and is not on the public API yet. Gemini 3.8 Flash is the Gemini you can call today.
SpaceXAI
1 model on the board. SpaceXAI, formerly xAI, builds Grok. Grok 4.7 is the flagship for coding and knowledge work. The developer docs are still on docs.x.ai.
Meta
1 model on the board. Muse Spark 1.3 is Meta's API flagship in this snapshot. It is fast, and it takes text, image, audio, and video.
DeepSeek
2 models on the board. DeepSeek ships open weights and a very cheap cache. V4.1 Flash is the current fast model. The official price doubles during peak hours.
Alibaba
2 models on the board. Qwen is Alibaba's model family. Qwen3.8 Max is the flagship here. The license is the Qwen license, not MIT. Read it before you ship the weights.
Z.ai
2 models on the board. Z.ai, formerly Zhipu, ships GLM. GLM-5.3 Flash is the cheap open model on the cost frontier. GLM-5.3 is the larger open flagship.
Moonshot
1 model on the board. Kimi is Moonshot's family. Kimi K3 is the open flagship. It scores near the closed frontier and is priced like a closed model.
Xiaomi
2 models on the board. MiMo-V2.6-Pro is the value surprise in this snapshot: open weights, an index score in Grok 4.7's band, and a measured task cost of $0.13.
Also shipping, not on this board
- Mistral. No Mistral model is in the frontier cut of the 1 Oct 2026 Artificial Analysis list used here. Check mistral.ai before you reuse an older Medium or Large score.
- Amazon. Nova is not in this frontier cut. Amazon still matters as a place Claude, OpenAI, and other models are hosted.
- NVIDIA. Nemotron 3 Ultra is on the same leaderboard at index 23, about $0.60 per measured task, roughly 180 tokens per second, and a 262k window. It is a different job from the models on this board.
- Cohere. North Mini Code shows up on Artificial Analysis for latency, not for a frontier intelligence score.
- IBM. Granite 4.2 appears among the cheapest measured tasks, at a much lower index score than the models compared here.
- MiniMax. MiniMax-M3 scores 29 on the same index, with a measured task cost of about $0.51. Open weights, below the cut for this board.