OhMyGPT

Parameters

How sampling and output parameters apply on each model family.

OhMyGPT adds no defaults or limits of its own to sampling and output parameters. The model's vendor sets them.

On the native protocol

On a model's native protocol, every parameter reaches the model as you sent it, including vendor-specific ones such as Mistral's safe_prompt or Qwen's enable_thinking. The vendor's defaults and limits apply, and if the vendor rejects a value, you get its error message. The exceptions are sampling parameters on OpenAI reasoning models and on recent Claude models.

Sampling on OpenAI reasoning models

OpenAI's reasoning models reject sampling parameters while they reason. On those requests OhMyGPT doesn't send temperature, top_p, logprobs or top_logprobs (on Responses: temperature, top_p, top_logprobs), the request runs without them, and the X-Parameters-Dropped response header lists the ones left out. This holds on every protocol you call these models through, and for each line of a Batch input file.

Whether a model is reasoning depends on the reasoning effort you send and the model's default, shown on each model's page:

  • Models that default to none, such as GPT-5.1, GPT-5.2 and GPT-5.4, keep the parameters unless you set another effort.
  • Models that reason by default, such as GPT-5.5, GPT-5.6 and GPT-6 Sol, keep them only when you set reasoning_effort (on Responses, reasoning.effort) to none.
  • Models without a none level, such as GPT-5, GPT-5.3-Codex and GPT-6 Astra, never receive them.
curl -i https://api.ohmygpt.com/v1/chat/completions \
  -H "Authorization: Bearer $OHMYGPT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-5.5",
    "temperature": 0,
    "messages": [{"role": "user", "content": "Say hi."}]
  }'
x-parameters-dropped: temperature

A stream that sent a keep-alive line before its first output carries it as a comment line ahead of the first event instead: : x-parameters-dropped: temperature.

Sampling on recent Claude models

Claude Opus 4.7 and the Claude models released after it don't accept temperature, top_p or top_k. These are Opus 4.7, Opus 4.8, Opus 5, Opus 5.5, Sonnet 5, Sonnet 5.5, Haiku 5.5 and the Fable 5 family. OhMyGPT doesn't send these parameters to them on any protocol, the request runs without them, and the X-Parameters-Dropped response header lists the ones left out. Earlier Claude models, such as Opus 4.6 and Haiku 4.5, receive them as sent.

Claude and Gemini on Chat Completions

When you call Claude or Gemini through Chat Completions, these parameters take effect under the vendor's own name. Cells marked — have no equivalent on that model:

Chat CompletionsClaudeGemini
max_tokens, max_completion_tokensmax_tokensmaxOutputTokens
temperature, top_p, top_ktemperature, top_p, top_ktemperature, topP, topK
stopstop_sequencesstopSequences
frequency_penalty, presence_penalty—frequencyPenalty, presencePenalty
seed—seed
response_format with json_schemaoutput_config.formatresponseMimeType and schema
response_format with json_object—responseMimeType: "application/json"
modalities with audio or image—responseModalities
logprobs, top_logprobs, logit_bias, verbosity, audio, web_search_options, min_p, top_a, repetition_penalty——
n greater than 1400400
tools, tool_choicetools, tool_choicefunctionDeclarations, toolConfig
parallel_tool_callsdisable_parallel_tool_use—
usermetadata.user_id—
reasoning_effort, thinkingsee Thinkingsee Thinking

Claude requires max_tokens; if you leave it out, OhMyGPT sends the model's maximum output.

A model without an equivalent doesn't receive the parameter, and the request runs without it, as Anthropic's own OpenAI SDK compatibility does for fields Claude lacks. So one request body works across models, but the output follows the model's defaults for that setting, so Claude ignores seed. The exception is n: these models return one choice per request, so n above 1 returns 400 naming n before the request reaches the model, and nothing is charged. If a setting must take effect, send it to a model that has it. Whether a model accepts a particular value of a parameter it has is up to the model; its own error comes back to you as is.

For JSON output from Claude, use response_format with a json_schema:

curl https://api.ohmygpt.com/v1/chat/completions \
  -H "Authorization: Bearer $OHMYGPT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-haiku-4-5",
    "max_tokens": 200,
    "messages": [{"role": "user", "content": "Give me a color and its hex code."}],
    "response_format": {
      "type": "json_schema",
      "json_schema": {
        "name": "color",
        "schema": {
          "type": "object",
          "properties": {"name": {"type": "string"}, "hex": {"type": "string"}},
          "required": ["name", "hex"],
          "additionalProperties": false
        }
      }
    }
  }'
"message": { "role": "assistant", "content": "{\"hex\": \"#FF6B6B\", \"name\": \"Coral Red\"}" }

Routing

provider on Chat Completions and Responses chooses which provider serves the model. It is read by OhMyGPT and does not reach the model. See Provider routing.

On this page