CHAT
Chat completions
Streaming and non-streaming chat completions on every model, with the OpenAI request and response schema. Existing SDKs work with a base URL change.
Models and pricing
Prices in USD per million tokens, published live at /v1/models. Every model on this page is priced under the cheapest listing we can find for it, and the API response is the source of truth, so what you see here is what you are billed.
The library
18 open-weight models behind one endpoint, all OpenAI-compatible. Search, filter by capability, and swap the model string; nothing else about your integration changes.
NEW
kimi-k3
$2.548/M in · $12.74/M out
1M context
deepseek-v4-flash
$0.14/M in · $0.60/M out
164K context
minimax-m3
$0.19/M in · $1.10/M out
204K context
deepseek-v3.2
$0.25/M in · $0.85/M out
164K context
llama-4-maverick
$0.15/M in · $0.55/M out
1M context
gpt-oss-120b
$0.0273/M in · $0.1547/M out
131K context
kimi-k2.7-code
$0.60/M in · $2.50/M out
262K context
qwen3-235b-a22b-2507
$0.0819/M in · $0.5005/M out
262K context
qwen3-coder-next
$0.20/M in · $0.80/M out
262K context
glm-5
$0.35/M in · $1.40/M out
203K context
mistral-large-3
$0.30/M in · $0.90/M out
262K context
gemma-4-26b-a4b-it
$0.0637/M in · $0.273/M out
262K context
qwen3-vl-30b-a3b
$0.20/M in · $0.91/M out
262K context
llama-4-scout
$0.11/M in · $0.34/M out
328K context
deepseek-r2
$0.55/M in · $2.19/M out
164K context
gemma-4-9b-it
$0.03/M in · $0.11/M out
131K context
gpt-oss-20b
$0.0273/M in · $0.1183/M out
131K context
minimax-m2.7
$0.06/M in · $0.31/M out
197K context
Full price list
14 open-weight models behind one endpoint,
https://nyxprovider.com/v1, all OpenAI-compatible. Context
is in tokens, prices are per million tokens. Swap the model string and
nothing else changes.
| Model | Context | Input | Output | Status |
|---|---|---|---|---|
kimi-k3 |
1,000,000 | $2.548 | $12.74 | live |
deepseek-v3.2 |
163,840 | $0.25 | $0.85 | live |
deepseek-r2 |
163,840 | $0.55 | $2.19 | live |
llama-4-maverick |
1,048,576 | $0.15 | $0.55 | live |
llama-4-scout |
327,680 | $0.11 | $0.34 | live |
gpt-oss-120b |
131,072 | $0.0273 | $0.1547 | live |
gpt-oss-20b |
131,072 | $0.0273 | $0.1183 | live |
qwen3-235b-a22b-2507 |
262,144 | $0.0819 | $0.5005 | live |
qwen3-coder-next |
262,144 | $0.20 | $0.80 | live |
qwen3-vl-30b-a3b |
262,144 | $0.20 | $0.91 | live |
glm-5 |
202,752 | $0.35 | $1.40 | live |
mistral-large-3 |
262,144 | $0.30 | $0.90 | live |
gemma-4-26b-a4b-it |
262,144 | $0.0637 | $0.273 | live |
gemma-4-9b-it |
131,072 | $0.03 | $0.11 | live |
Built for long context
Two models in the catalog take a million tokens or more: kimi-k3 at 1,000,000 and llama-4-maverick at 1,048,576. Long prompts are where caching pays. Repeated prefixes, a system prompt, a codebase, a document set, bill at the cached input rate published in the API, and the cache hits automatically on byte-identical prefixes with no request changes. A 400,000-token prompt that reuses a 380,000-token prefix bills 95 percent of its input at the cached rate.
How caching is billed →What every model supports
One request schema across the catalog, the OpenAI one. Streaming passes tokens through byte for byte, and routing adds 71µs at p50 and 235µs at p99 on top of model latency.
CACHE
Prompt caching
Byte-identical prefixes bill at the cached input rate automatically. No cache-control headers, no request changes, no minimum prefix length above 1,024 tokens.
JSON
Structured output
JSON mode and json_schema response formats with constrained decoding, so the response parses against your schema instead of failing in your parser.
TOOLS
Function calling
Parallel tool calls with the OpenAI tools schema. Tool call deltas stream in the order the model emits them, byte for byte.
Limits
Defaults on the public endpoint, per API key. Past a limit you get an
immediate 429 with a Retry-After header, never a request
that queues silently and streams late. Dedicated
capacity replaces all three defaults with numbers you pick.
| Limit | Public endpoint | Dedicated |
|---|---|---|
| Concurrent requests | 16 | Custom |
| Prompt tokens per minute, per model | 500,000 | Custom |
| Requests per minute, burst | 60 | Custom |
Questions
The short version of policies that live in full in the docs.
How do prices stay under every listing?
We track published price sheets for every model we serve and re-check them daily. When a competing listing drops below ours, the price at /v1/models moves the same day, and the API response is what bills, so there is no gap between the page and the invoice. The margin comes from utilization and 71µs routing overhead, not from marking up tokens.
How do new models get added?
Weights land, we tune batching and speculative decoding for the architecture, run an output-parity suite against the reference implementation, then publish the price. Bring-up runs two to five days for a new architecture and under 48 hours for a familiar one. Requests for specific models go to hello@nyxprovider.com.
What precision do you serve?
A mix of FP8 and BF16, chosen per model. Most weights and KV cache run in FP8, with BF16 kept for layers that measurably degrade under quantization in our parity suite. The served precision for each model is listed in its metadata at /v1/models, and it never changes without a changelog entry.
What happens when a model is deprecated?
30 days of notice by email and in a response header before any model comes off the endpoint. The price holds through the notice window, and the deprecation notice names a suggested replacement. Dedicated capacity can pin a model past its public retirement.