Chat Completion

Overview

The Chat Completion API generates AI responses given a list of messages. It's fully OpenAI-compatible — use the same SDK and request format.

Endpoint: POST /v1/chat/completions

How a request flows

Every chat completion takes the same path through the platform — your app never talks to the underlying provider directly:

1. 1

   Your app

   Send with +

2. 2

   CallMissed gateway

   Authenticate the key, check credits, route by model id

3. 3

   Provider

   Run inference on the best-fit backend — picked from the model id

4. 4

   CallMissed gateway

   Stream tokens back and deduct credits when the response completes

5. 5

   Your app

   Receive the completion (all at once, or token-by-token when streaming)

Tip: The model id decides routing automatically — you never pick a backend. See How CallMissed Works.

Make your first request

Basic Usage

from openai import OpenAI

client = OpenAI(
    api_key="cm_your_key",
    base_url="https://api.callmissed.com/v1"
)

response = client.chat.completions.create(
    model="sarvam-105b",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What is the capital of India?"}
    ]
)

print(response.choices[0].message.content)

Parameters

On GPT-5 and GPT-6 family models, temperature, top_p and logit_bias are not supported by the model and are left out of the request; the model runs at its default sampling.

OpenAI Python SDK note — The OpenAI client validates kwargs against its known parameters, so a CallMissed-specific field such as reasoning_effort raises TypeError: Completions.create() got an unexpected keyword argument. Pass it via extra_body instead:

client.chat.completions.create(
    model="kimi-k2.6",
    messages=[...],
    extra_body={"reasoning_effort": "none"},
)

Raw HTTP / curl users can keep it at the top level — only the OpenAI SDK gates kwargs.

Model Substitution

CallMissed never substitutes your model on its own. Send a model and you get that model, or a clean error (429/503 with Retry-After).

To opt in to failover, list fallbacks yourself in models:

{
  "model": "kimi-k2.6",
  "models": ["kimi-k2.5", "gpt-oss-120b"],
  "messages": [{"role": "user", "content": "Hello"}]
}

If model fails with a retryable error (an upstream outage or rate limit), the next id in models is tried. Each fallback must pass the same checks as model — your plan, the key's allowed_models, maintenance status, vision support and context window — and ids that fail them are skipped. The response's model field names the model that actually answered, and you are billed at that model's rate. Requests that send tools, tool_choice, response_format or structured_outputs never fall back, because a different model could change the result shape.

Need a model that is not in the catalog? See Models on demand.

Vision (Image Input)

Multimodal content (text + image parts) is accepted on any model whose supports_vision flag is true in GET /v1/models. Models without vision support reject image content with 400 unsupported_image_input before the upstream call, so you're not charged.

from openai import OpenAI

client = OpenAI(api_key="cm_your_key", base_url="https://api.callmissed.com/v1")

resp = client.chat.completions.create(
    model="gpt-5.6-sol",   # supports_vision: true
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "What is in this image?"},
            {"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}},
        ],
    }],
)

Vision-capable models: gpt-6.1-sol, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-4o, gpt-4.1, gpt-5-mini, grok-4.3, gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash, gemini-3.5-flash-lite, gemini-3.1-pro-preview, gemini-3.1-flash-lite, gemma-4-31b, gemma-4-26b-a4b-it, kimi-k2.5, kimi-k2.5-fast, kimi-k2.6, kimi-k2.7-code, mistral-small-3.1.

GET /v1/models is authoritative. Read supports_vision there rather than hard-coding this list.

An image_url may be an https:// URL or a base64 data URL. The Gemini models and gemma-4-31b only take image bytes, so CallMissed downloads a URL image for them: it must be publicly reachable without redirects, return an image content type, and be at most 20 MB (10 MB for gemma-4-31b). Gemini accepts PNG, JPEG, WEBP, HEIC and HEIF. If a download fails the request returns 400; send base64 to avoid the extra fetch.

Audio and Document Input

The Gemini models (gemini-3.8-flash, gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash, gemini-3.5-flash-lite, gemini-3.1-pro-preview, gemini-3.1-flash-lite) also take audio and PDF parts:

{"type": "input_audio", "input_audio": {"data": "<base64>", "format": "wav"}}
{"type": "file", "file": {"file_data": "data:application/pdf;base64,<base64>", "filename": "report.pdf"}}

Audio format may be wav, mp3, aac, flac, ogg, opus, m4a, aiff or webm. A file part must carry the PDF inline in file_data; file_id references are not supported. Any other model rejects these parts with 400 unsupported_audio_input or 400 unsupported_file_input before the upstream call, so you're not charged. Content that can't be passed on as sent is rejected with a 400 that names the part, rather than being dropped.

Context Window

Every model in the catalog advertises a context_window (token count for the combined prompt + completion). The GET /v1/models response exposes it under two keys for cross-client compatibility:

• context_window (OpenAI/CallMissed canonical name)
• context_length (OpenAI SDK convention — same value)

from openai import OpenAI

client = OpenAI(api_key="cm_your_key", base_url="https://api.callmissed.com/v1")

for m in client.models.list():
    extra = m.model_extra or {}
    print(m.id, extra.get("context_window"), extra.get("supports_vision"))

Snapshot — GET /v1/models is authoritative:

Images, audio and files count toward the window by what the model reads, not by their encoded size, so a large base64 image does not by itself cause context_length_exceeded. A prompt that is clearly longer than the window is rejected with 400 context_length_exceeded before the model is called (on /v1/chat/completions, /v1/responses and /v1/messages); a prompt close to the limit is sent to the model, and if the model finds it too long you get the same 400 context_length_exceeded (as an error event mid-stream when stream: true). Either way nothing is billed.

Prompt caching

Models that support prompt caching reuse repeated prompt prefixes automatically — there is nothing to turn on. Cached prompt tokens are billed at the model's cached-input rate where one is published (see Models); a model with no cached rate bills them at its normal input rate. Tokens written to the cache are billed at the input rate, except on models that publish a separate cache-write rate (such as gpt-6.1-sol).

Every response reports the cached share of the prompt:

"usage": {
  "prompt_tokens": 3120,
  "completion_tokens": 42,
  "total_tokens": 3162,
  "prompt_tokens_details": { "cached_tokens": 2944 }
}

• prompt_tokens is the whole prompt, cached part included.
• prompt_tokens_details.cached_tokens is always present (0 on a miss or the first request).
• prompt_tokens_details.cache_write_tokens appears when the model reported tokens written to the cache (GPT-5.6 and later bill these at their own rate).
• Streaming: the same block is in the final usage chunk when you send stream_options: {"include_usage": true}.
• /v1/responses reports the same numbers as usage.input_tokens_details.cached_tokens and usage.input_tokens_details.cache_write_tokens.

To improve hit rates:

• Put stable content first (system prompt, tool definitions, reference documents) and the changing part last.
• Send a prompt_cache_key (up to 1,024 characters, on /v1/chat/completions and /v1/responses) and reuse it for requests that share a prefix. It is passed on as a cache-routing hint for kimi-k2.5, kimi-k2.6, kimi-k2.7-code, glm-4.7-flash, glm-5.2, gpt-oss-120b, nemotron-3-super, gemma-4-26b-a4b-it, mistral-small-3.1, DeepSeek-V4-Pro and DeepSeek-V4-Flash; other models ignore it.
• Explicit per-block cache breakpoints (cache_control / prompt_cache_breakpoint on a content part) are not applied today — caching works from the prompt prefix automatically.

{
  "model": "kimi-k2.6",
  "prompt_cache_key": "support-bot:policy-v3",
  "messages": [
    {"role": "system", "content": "<long, stable instructions>"},
    {"role": "user", "content": "Where is my order?"}
  ]
}

A prefix usually needs to be at least ~1,024 tokens before it is cached, and an idle cache expires after a few minutes. Hits are not guaranteed.

Responses API

For clients built on OpenAI's newer Responses API, CallMissed exposes a compatible POST /v1/responses endpoint. It accepts a Responses-shaped body and translates to the same chat engine under the hood — so you can point an OpenAI Responses client at https://api.callmissed.com/v1 without changes.

Endpoint: POST /v1/responses

from openai import OpenAI

client = OpenAI(api_key="cm_your_key", base_url="https://api.callmissed.com/v1")

resp = client.responses.create(
    model="gpt-4.1",
    input="Write a haiku about databases.",
)
print(resp.output_text)

• The same models, pricing, plan rules, vision and tool-calling support as /v1/chat/completions apply — this is a request/response-shape adapter, not a different model set.
• Responses are not stored: store is accepted for compatibility and ignored, and previous_response_id returns 400 unsupported_parameter. Send the full conversation in input each turn.
• An input_image given only as file_id, or an input_file given as file_id or file_url, returns 400 unsupported_content_part naming the part. Send the bytes inline instead.

If you're starting fresh, /v1/chat/completions is the most widely-supported surface; use /v1/responses when porting an existing Responses-API integration.

Errors

All errors return the OpenAI-compatible envelope:

{
  "error": {
    "message": "Invalid API key",
    "type": "invalid_request_error",
    "code": "invalid_api_key"
  }
}

Every response carries an X-Request-ID header — quote it when you contact support.