Models

The API serves open-weight coding and reasoning models. Rates are public, and each model has its own page with the full spec, upstream weights, license, and reasoning controls.

Prices are USD per 1M tokens. Current promotional rates are shown; list rates, max output, and modalities are on each model page.

Models that were previously offered are listed under Retired models.

Choosing a model

• DeepSeek V4.1 Flash is the recommended default for coding agents: a million-token context, strong tool use, and low latency.
• Kimi K3 is the most capable option for hard, long-horizon agent work.
• GLM-5.3 Flash is the fastest and cheapest choice for high-volume or straightforward tasks.
• Dolly 1 Security is an uncensored variant for defensive and red-team workflows. It has no standard safety filters; only use it on your own systems or ones you are authorized to test.

Auto routing

auto is a virtual model that routes each request to a heuristic-selected tier, then applies session affinity to keep a conversation on one model:

Pass "model": "auto" and a tier is selected automatically. auto has no price of its own; each request bills at the rate of the model that handles it.

Model aliases

Every model is also exposed under a zro/ prefix. Harnesses that reserve well-known model IDs can still select it. The alias resolves to the same model, pricing, and routing:

zro/deepseek-v4.1-flash
zro/kimi-k3
zro/glm-5.3
zro/glm-5.3-flash
zro/dolly1-security
zro/auto

Availability by region

Inference can run in Europe, the United States, or both, and the setting is per API key. If a model is not served in any region a key allows, the request fails rather than routing elsewhere. See Regions.

Prompt caching

All served models support prompt caching. Repeated prefixes are cached, and billing uses the model's cache read rate, a fraction of the input rate, instead of the full input rate. Cache is namespaced per API key. See Prompt caching.