Recommended Models
One model for each kind of work. Each is a starting point that people who build agents have tested in real use.
Prices are in USD for 1M input tokens / 1M output tokens. The list was updated on
. Coding agents read the same
list as Markdown at /models.md, or with agents current-models.
Recommendations
Frontier
The most capable general models, for work where quality decides and price is secondary.
GPT-6 Astra
OpenAI gpt-6-astra
- Price
- 10.00 / 50.00
- Input
- text, image, file
- Output
- text
- Weights
- Closed
Reserve for difficult reasoning, complex debugging, and heavy implementation where the extra capability justifies the cost.
Workhorse
Affordable, dependable models for the bulk of everyday agent and coding work.
GPT-6 Luna
OpenAI gpt-6-luna
- Price
- 0.10 / 0.50
- Input
- text, image, file
- Output
- text
- Weights
- Closed
Start here for routine coding, focused subagents, tool calls, classification, extraction, summaries, and everyday writing.
Open-weight workhorse
The best everyday model whose weights you can download and host yourself.
DeepSeek V4.1 Flash
Novita deepseek/deepseek-v4.1-flash
- Price
- 0.30 / 1.20
- Input
- text, image
- Output
- text
- Weights
- Open
The recommended low-cost option with open weights, for focused agent and coding work. Novita cache reads are $0.006 per million tokens.
Code review
Models that read code critically and find real bugs, security flaws, and regressions.
GLM 5.3
Novita zai-org/glm-5.3
- Price
- 1.40 / 4.40
- Input
- text
- Output
- text
- Weights
- Open
Preferred specialist for security analysis, vulnerability review, and security-sensitive changes. Use the full GLM 5.3 model, not GLM 5.3 Flash.
Orchestrator
Models that coordinate many agents, tools, and work streams over long runs.
GPT-6.1 Sol
OpenAI gpt-6.1-sol
- Price
- 2.00 / 10.00
- Input
- text, image, file
- Output
- text
- Weights
- Closed
Use when coordinating many agents, tools, dependencies, or work streams is the hard part.
UI and frontend
Models for demanding interface work: design fidelity, layout, and visual refinement.
Claude Opus 5.5
Anthropic claude-opus-5-5
- Price
- 4.00 / 20.00
- Input
- text, image, file
- Output
- text
- Weights
- Closed
Use for difficult UI implementation, design fidelity, and visual refinement. Routine UI changes still start with Luna.
Ultra-cheap
Low-intelligence, very cheap models for high-volume simple work such as extraction and classification.
No pick yet. No model has been judged for this category.
Decision
Small, fast models for the routing and yes-or-no judgements an agent makes between steps.
No pick yet. No model has been judged for this category.
Image generation
Models that generate or edit images.
Gemini 3 Pro Image Preview
Google gemini-3-pro-image-preview
- Price
- 2.00 / 12.00 text, 120.00 image out
- Input
- text, image
- Output
- text, image
- Weights
- Closed
Image generation with strong prompt understanding. Image output costs far more than text.
Vision and documents
Models that read images, screenshots, PDFs, and scanned documents.
No pick yet. No model has been judged for this category.
How a Model Is Chosen
A new model is judged at least three days after its release, so that people have used it for real work. It is compared with the current pick in its category on three things:
- What practitioners report counts for the most. Reports that show work outweigh opinions.
- Independent benchmarks count for about a third.
- Price decides between models of similar quality. The cheaper model wins.
A new model replaces a pick only when it is clearly the better default. A person reviews every change to this list.
Catalog
Other models worth knowing. They are alternatives, and the recommendations above take precedence. The Markdown guide has the scores and the pricing notes for each one.
| Model | Best for | Price |
|---|---|---|
| GPT-6 Luna | Default for most work, especially all cheap/small tasks | 0.10 / 0.50 |
| DeepSeek V4.1 Flash | Low-cost alternative for focused agent and coding work | 0.30 / 1.20 |
| GLM 5.3 | Security work and security review | 1.40 / 4.40 |
| GPT-6 Astra | Very expensive specialist for heavy work | 10.00 / 50.00 |
| Claude Opus 5.5 | Very expensive specialist for demanding UI work | 4.00 / 20.00 |
| GPT-6.1 Sol | Very expensive specialist for coordination-heavy efforts | 2.00 / 10.00 |
| Kimi K3 | Open-source heavy work | 3.00 / 15.00 |
| Gemini 3.1 Pro | Multimodal pro workflows | 2.00 / 12.00 |
| Tencent Hy3 | Cheap long-horizon agents | 0.14 / 0.58 |
| NVIDIA Nemotron 3 Ultra 550B A55B | Frontier open reasoning and coordinator/orchestration work at a value price. Strong tool discernment across large tool sets. | 0.50 / 2.50 |
| GLM 4.7 Flash | Good for conversation at a very low cost. Fine with small number of tool calls, but not appropriate for coordinators. | 0.06 / 0.40 |
| Gemma 4 31B | Conversation, tool calling, speed — open weights. | 0.10 / 0.35 |
| gemini-3-flash-preview | Fast grounded assistants | 0.50 / 3.00 |
| gemini-3-pro-image-preview | Image generation with strong prompt understanding | 2.00 / 12.00 text, 120.00 image out |
| NVIDIA Nemotron 3 Super 120B | Cheap open agent systems | 0.10 / 0.50 |
| Qwen3.5 397B A17B | Multilingual multimodal work | 0.39 / 2.34 |
| MiniMax M3 | Coding agents on a budget | 0.30 / 1.20 |
| Qwen3.5 35B A3B | Cheap deployable open model | 0.14 / 1.00 |