Models and pricing

#Models

USD per 1M tokens, markup included. Build snapshot at price version 2026-11-11.4, as of 2026-11-11. Prices follow the providers and may have changed since: the live list is GET /v1/models (no key), also as models.json and models.md. Model ids are exact; any other name is 404 model_not_found.

New, removed and renamed models, refusal labels, providers and price moves are dated in the model changelog; follow its Atom feed or read /changelog.json.

abliterated-model-large-v2: Based on GLM-5.3, abliterated and post-trained by abliteration.ai, so it can behave differently from GLM-5.3 Abliterated (glm-5.3-abliterated).

Until 2026-11-07, abliterated-model-large-v2 was also served by Refuseless, Experiential Labs and Silk Compute and priced at the cheapest of them. abliteration.ai now serves it alone, at abliteration.ai's price; their GLM-5.3 Abliterated is glm-5.3-abliterated.

abliterated-model-large-flash, abliterated-model, kimi-k3.1-pro-abliterated, deepseek-v4.1-flash-fast, minimax-m3.1-flash, deepseek-v4.1-flash, glm-5.3-flash, kimi-k3.1-pro and hy4 accept image input; the rest are text only. Context windows go up to 1,048,576 tokens.

#Reading /v1/models

GET /v1/models returns {"object": "list", "data": [...], "price_version", "prices_updated_at", "markup_bps"}. Each entry has the model's id, its cheapest provider's rates with the markup in pricing_per_mtok_micros, context_length, max_output_tokens, input_modalities, endpoints (which of /v1/chat/completions, /v1/responses and /v1/messages it is served on) and every provider's offer. A model is not served on an endpoint missing from its list: 400 unsupported_endpoint.

curl https://inference.dagon.so/v1/models

#How the charge is computed

Cost is exact integer micro-USD: ceil(Σ tokens × rate × (1 + markup)), at the rates of the provider that served the request. Cache reads bill at the discounted cached-input rate; cache writes and uncached input at the standard input rate. A request that falls back to another provider is billed at that provider's rate plus the markup, and its receipt names that provider and the price version.

#The hold

Before forwarding, the gateway holds the request's maximum cost: an upper bound of the prompt (its JSON size in UTF-8 bytes, plus a template allowance per turn and for tools; 16,384 tokens per image part) plus the full output cap, times n on chat. The bound is roughly 2 to 4 times the real input cost for ordinary English, and everything beyond the charge is released when the request settles. A non-text part other than an image (a document, file or audio part) holds the whole context window.

#Output cap

Without a cap the gateway writes 16,384 tokens for you. The largest cap is 65,536 tokens or the model's own limit, whichever is lower; a larger one is 400 max_tokens_too_large. A smaller cap means a smaller hold, so set it when your balance is tight.

#Worked example

A 1,200 character prompt to Abliterated Large Flash on abliteration.ai (abliterated-model-large-flash), answered with 412 input tokens (256 cached) and 238 output tokens:

Rates of the abliteration.ai route at price version 2026-11-02.2.

#What each outcome costs

The charge never exceeds the hold. If reported usage prices above it, you are charged the hold and the receipt carries held_micros.