Z.ai's flagship for complex, long-horizon coding. Text only.
- 1M context
- 128K output
- Tools
- Reasoning
- Intelligence
- 45
- DeepSWE
- 69%
- Terminal-Bench
- 83.9%
- Input
- $1.40
- Output
- $4.40
- Cache read
- $0.26
USD per 1M tokens
Models
Pick the right model.
Every model we serve, side by side: list prices, context windows, and scores on the benchmarks we trust for agentic coding.
6 models in production
Ranked by the Artificial Analysis Intelligence Index, highest first.
Moonshot's 2.8T flagship for repository-scale agentic work, with native vision. The premium tier.
- 1M context
- 128K output
- Vision
- Tools
- Reasoning
- Intelligence
- 44
- DeepSWE
- 68.5%
- Terminal-Bench
- 85%
- Input
- $3.00
- Output
- $15.00
- Cache read
- $0.30
USD per 1M tokens
Z.ai's fast multimodal coder: 18B active parameters on a 1M window, at a flash price.
- 1M context
- 128K output
- Vision
- Tools
- Reasoning
- Intelligence
- 42
- DeepSWE
- 63.4%
- Terminal-Bench
- 84.3%
- Input
- $0.15
- Output
- $0.50
- Cache read
- $0.03
USD per 1M tokens
DeepSeek's latest flash model, built for fast agentic coding, with native image input.
- 1M context
- 384K output
- Vision
- Tools
- Reasoning
- Intelligence
- 39
- DeepSWE
- 74.2%* reported by the publisher
- Terminal-Bench
- 90.6%* reported by the publisher
- Input
- $0.15
- Output
- $0.60
- Cache read
- $0.028
USD per 1M tokens
The lowest price in the lineup for real agentic work. Text only.
- 1M context
- 384K output
- Tools
- Reasoning
- Intelligence
- 34
- DeepSWE
- 54.4%* reported by the publisher
- Terminal-Bench
- 78.7%
- Input
- $0.14
- Output
- $0.28
- Cache read
- $0.028
USD per 1M tokens
Qwen3.6-35B-A3B, open weights by Qwen
A light, fast complement for the roles around your main coder: scouting, summaries, quick edits.
- 256K context
- 256K output
- Vision
- Tools
- Reasoning
- Intelligence
- 18
- DeepSWE
- Not published
- Terminal-Bench
- 44.9%
- Input
- $0.15
- Output
- $1.00
- Cache read
- $0.05
USD per 1M tokens
In Labs
Experiments open for a short window, free with a Labs seat while they run. Seats are limited and rotate, and an experiment is not meant for production work.
MiMo-V2.6-Pro-RL, open weights by Xiaomi
Xiaomi's flagship open model, post-trained for agentic coding. Free to try while the experiment runs.
- 1M context
- 128K output
- Vision
- Tools
- Reasoning
- Intelligence
- 46
- DeepSWE
- 71.9%* reported by the publisher
- Terminal-Bench
- 89.9%* reported by the publisher
Free with a Labs seat while the experiment runs
Free to try while the experiment runs. For production work, use umans-deepseek-v4.1-flash.
- 1M context
- 384K output
- Vision
- Tools
- Reasoning
- Intelligence
- 39
- DeepSWE
- 74.2%* reported by the publisher
- Terminal-Bench
- 90.6%* reported by the publisher
Free with a Labs seat while the experiment runs
About the scores
Scores as of October 3, 2026. Each model's page links every score to its source.
Artificial Analysis Intelligence Index
By Artificial Analysis
An independent composite of reasoning, knowledge, maths, coding and agentic evaluations, run by Artificial Analysis on every model the same way.
MethodologyDeepSWE
By Datacurve
Long-horizon software engineering: real repository tasks an agent has to finish end to end.
MethodologyTerminal-Bench 2.1
By tbench.ai
Hard, realistic tasks an agent completes in a terminal sandbox: building, debugging, configuring.
Methodology* Reported by the model's publisher on its model card. Publishers run their own harnesses, so these compare less cleanly than the independent measurements, which carry no mark. No number here is estimated: a model without one has not been measured on that benchmark yet.
One key, every model.
Create a key once, then switch models by changing one string in your tool.
Looking for a retired model? Its history stays on the status page.