Models & Pricing
Chat models
Prices are per million tokens, input and output priced separately. Where a model has a cached-input price, the input tokens the provider reports as cache hits bill at that rate instead of the input rate; the receipt shows the split. deepseek-v3.2 has no cached tier.
The Claude models cache prompts automatically (no cache_control needed) and report cache writes as well as reads. Prompt tokens written to the cache on a turn bill at a cache-write price instead of the input price — claude-opus-5-5 1.50, claude-haiku-5-5 7.50 per million — and GET /v1/models reports it as cache_write_per_million_usd. Every other model has no cache-write price (null): written prompt tokens simply bill at the input price.
Prices marked promotional are introductory and can change; GET /v1/models reports promotional_pricing: true on them. Occasionally a model runs a short free window — GET /v1/models then shows a free_until timestamp on it, and requests completed before that moment are not charged.
The GPT models (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-6-astra, gpt-6-sol, gpt-6-luna, and gpt-6.1-sol) and the Claude models (claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5, and claude-fable-5.1) serve /v1/responses in addition to chat completions. The other chat models serve chat completions only.
gpt-6.1-sol always reasons, and its provider serves function tools only on the Responses surface — so a chat-completions request with tools is served through /v1/responses for you and comes back as an ordinary chat completion (the same way max is handled; see the table below). temperature other than 1 is rejected on it. Image inputs on the Claude models must be inline data: URLs; remote image URLs are rejected.
Requests that omit the output-limit field get a per-model default: 4,096 tokens for deepseek-v3.2, 8,192 tokens for every other chat model listed above.
Reasoning effort by model
Every chat model takes the OpenAI reasoning_effort field (reasoning.effort on /v1/responses) to set its thinking level. The accepted values differ per model. A value the model does not accept is rejected with a 400 — Invalid request. when the platform catches it, The selected model could not complete the request. when the provider does — and nothing is charged either way. Verified against production on 2026-09-24; the gpt-6.1-sol and Claude rows on 2026-10-21.
Reasoning tokens are billed as output tokens at the model’s output rate and are included in usage.completion_tokens; the receipt shows the exact charge. The DeepSeek, Kimi, and GLM models return their thinking as message.reasoning_content (and delta.reasoning_content in streams) next to the normal content. The GPT models never return reasoning text — only the count in usage.completion_tokens_details.reasoning_tokens. The Claude models return reasoning text inconsistently (sometimes message.reasoning / reasoning_content, sometimes nothing) while always billing the reasoning tokens; do not depend on the text being present.
Image generation models
Four models serve POST /v1/images/generations. Image requests are never billed by tokens — the price is fixed by the request itself and the receipt settles at exactly that amount.
FLUX — billed per megapixel
FLUX models accept any WIDTHxHEIGHT size with each side 256–2560 px and an area up to 4 megapixels, billed as ceil(width × height / 2²⁰) megapixels — a 1024x1024 image is 1 MP; 1600x1024 rounds up to 2 MP.
GPT Image — billed per image (quality × size)
gpt-image-2
gpt-image-1.5
Omitted size and quality default to 1024x1024 and high. Requesting n images multiplies the per-image price by n (up to 4). See Image Generation for the full parameter reference.
Both GPT Image models also serve POST /v1/images/edits, sharing this same per-image (quality × size) grid: an edit costs exactly what a generation of the same quality, size, and n costs. FLUX models do not support editing.
GET /v1/models is the authoritative live list, including current pricing. Prices on this page are current as of 2026-09-24. See Models & Receipts for the endpoint reference.