Best LLMs for AI agents
Models their providers position for agentic work: calling tools, following multi-step plans and recovering from errors over long sessions.
An agent multiplies whatever a model gets wrong. If each step succeeds 95% of the time, a 20-step task succeeds only about 36% of the time (0.95 to the 20th power), so reliability on tool calls and instruction following matters more than peak benchmark scores.
Long context lets an agent keep its plan, tool definitions and results in view, and prompt caching keeps that repeated material cheap. Price agents per completed task, not per token: a pricier model that finishes in fewer steps can cost less.
The shortlist
| Model | Tier | Weights | Context | Input / 1M | Output / 1M | Inputs |
|---|---|---|---|---|---|---|
| DeepSeek-V4-ProDeepSeek | Frontier | Open weights | 1M | $1.32 | $3.96 | Text |
| GLM-5.3Z.ai (Zhipu AI) | Frontier | Open weights | 1M | $1.40 | $4.40 | Text |
| Grok 4.7xAI | Frontier | Closed | 500K | $2 | $6 | Text, Images |
| Qwen3.8-MaxAlibaba (Qwen) | Frontier | Closed | 1M | $2 | $6 | Text, Images |
| Qwen3.8-2.4T-A95BAlibaba (Qwen) | Frontier | Open weights | 256K | $2 | $6 | Text |
| Kimi K3Moonshot AI | Frontier | Open weights | 1M | $3 | $15 | Text, Images, Video |
| Claude Opus 5.5Anthropic | Frontier | Closed | 1M | $4 | $20 | Text, Images, PDF |
| Claude Fable 5.1Anthropic | Frontier | Closed | 1M | $10 | $50 | Text, Images, PDF |
| GPT-6 AstraOpenAI | Frontier | Closed | 1.05M | $10 | $50 | Text, Images |
| gpt-oss-120bOpenAI | Advanced | Open weights | 128K | $0.15 | $0.60 | Text |
| DeepSeek-V4.1-FlashDeepSeek | Advanced | Open weights | 1M | $0.30 | $1.20 | Text, Images |
| MiniMax-M3MiniMax | Advanced | Open weights | 1M | $0.30 | $1.20 | Text, Images, Video |
| Mistral Large 3Mistral AI | Advanced | Open weights | 256K | $0.50 | $1.50 | Text, Images |
| NVIDIA Nemotron 3 UltraNVIDIA | Advanced | Open weights | 1M | $0.50 | $2.20 | Text |
| Gemini 3.8 FlashGoogle | Advanced | Closed | 1M | $0.75 | $3.75 | Text, Images, Video, Audio, PDF |
| Gemini 3.6 FlashGoogle | Advanced | Closed | 1M | $0.75 | $3.75 | Text, Images, Video, Audio, PDF |
| Mistral Medium 3.5Mistral AI | Advanced | Open weights | 256K | $1.50 | $7.50 | Text, Images |
| Claude Sonnet 5.5Anthropic | Advanced | Closed | 1M | $2 | $10 | Text, Images, PDF |
| GPT-6.1 SolOpenAI | Advanced | Closed | 1.05M | $2 | $10 | Text, Images |
| Claude Haiku 5.5Anthropic | Efficient | Closed | 1M | $0.10 | $0.50 | Text, Images, PDF |
| GLM-5.3-FlashZ.ai (Zhipu AI) | Efficient | Open weights | 1M | $0.15 | $0.50 | Text, Images, Video |
| Mistral Small 4Mistral AI | Efficient | Open weights | 256K | $0.15 | $0.60 | Text, Images |
| Granite 4.2 30BIBM | Efficient | Open weights | 128K | $0.16 | $0.65 | Text |
| Muse Glimmer 30BMeta | Efficient | Open weights | 128K | $0.35 | $1.50 | Text, Images, Video |
| Grok 4.3xAI | Efficient | Closed | 1M | $1.25 | $2.50 | Text, Images |
| Qwen3.8-27BAlibaba (Qwen) | Efficient | Open weights | 256K | $0.80 | $4 | Text, Images, Video |
| Amazon Nova 2 LiteAmazon (AWS) | Efficient | Closed | 1M | — | — | Text, Images, Video, PDF |
| NVIDIA Nemotron 3.5 Lightning 30B-A3BNVIDIA | Compact | Open weights | 1M | $0.05 | $0.20 | Text |
| Qwen3.6-35B-A3BAlibaba (Qwen) | Compact | Open weights | 256K | $0.10 | $0.95 | Text, Images, Video |
| Mistral Large 4Mistral AI | Frontier | Closed | 1M | $1.36 | $4.18 | Text, Images |
| Muse Spark 1.3Meta | Advanced | Closed | 1M | $1.25 | $4.25 | Text, Images, Video, PDF |
| Grok Build 0.1xAI | Efficient | Closed | 250K | $1 | $2 | Text, Images |
| Grok 4.6xAI | Frontier | Closed | 500K | $2 | $6 | Text, Images |
| GPT-5.6 SolOpenAI | Frontier | Closed | 1.05M | $4 | $20 | Text, Images |
| Claude Opus 5Anthropic | Frontier | Closed | 1M | $5 | $25 | Text, Images, PDF |
| GPT-5.5OpenAI | Frontier | Closed | 1.05M | $5 | $30 | Text, Images |
Ordered by status, capability tier and then price. Tiers are our editorial grouping of provider positioning and public benchmark standings, not our own test results. How we rank.
How to choose
- Require native tool calling and structured outputs; check the model page for both.
- Prefer models with prompt caching when your system prompt and tool list are long.
- Measure cost per finished task, including retries, rather than cost per token.
- Give agents least-privilege tools and a human approval step for anything irreversible. See LLM security and privacy.
Frequently asked questions
What makes an LLM good for agents?
Dependable tool calling, accurate instruction following over many steps, the ability to recover from errors, and enough context to hold the plan and tool results. Cost per completed task and latency per step also matter because agents make many calls.
Do I need a reasoning model for agents?
Not always. Reasoning modes help with planning and tricky steps but add output tokens and latency. Many teams use a reasoning model to plan and a cheaper fast model to execute routine steps.
Which framework should I use to build an agent?
Start with your provider's agent SDK or a lightweight framework, and add orchestration only when you need it. Our guide to AI agents compares the main options and when a simple workflow beats a full agent.
Other shortlists
Need hands-on help building it?
RAG, agents, fine-tuning or a production chatbot: describe the project and we'll suggest an approach and, if you want, partners who build it.
- Free and vendor-neutral: no model provider pays for placement.
- Your details go only to partners you agree to.