Gemini parameter and protocol differences

For Gemini models, adjust thinking, tools and caching using this page. See Generate Content, Stream Generate Content and Chat Completions for full structures; native standard and streaming requests share these rules.

Applicable models

Model	Usage notes
gemini-4-argon	Native thinking levels: `LOW/MEDIUM/HIGH`; default: `MEDIUM`; implicit caching only; long output requires continuation
gemini-3.8-flash	Also uses three thinking levels, defaulting to `MEDIUM`; native `MINIMAL` is unsupported. Supports implicit and explicit caching
gemini-3.7-flash	Also uses three levels, defaulting to `MEDIUM`; native `MINIMAL` is unsupported. Google Developer API deprecated this model in favor of 3.8; this concerns Google's service, not platform removal or automatic substitution

Generate Content differences

Field	Difference
`thinkingConfig.thinkingLevel`	None of the three can disable thinking with `MINIMAL` or an old zero budget; use `LOW` to reduce effort, not Gemini 2.5's disabling configuration
`thinkingConfig.thinkingBudget`	Use `thinkingLevel` for new integrations; the two fields are mutually exclusive. For Argon, Gemini Developer API maps a standalone old budget to the nearest level and warns
`temperature/topP/topK`	These models ignore old sampling settings; omit them, as they cannot control sampling
`candidateCount`	Gemini 3 and later do not support this setting; remove it from requests
`frequencyPenalty/presencePenalty`	Google Cloud's 3.7 Flash guide states these return an error, unlike the silently ignored sampling fields above
`thinkingConfig.includeThoughts`	When enabled, `parts[].thought` distinguishes thinking summaries from answers; summaries are not original thinking and cannot replace signatures in history
`parts[].thoughtSignature`	Retain returned Parts completely and unchanged; tool-call history must include signatures. Do not merge signed and unsigned blocks
`functionCall/functionResponse`	Match function-result `id` and `name` to actual calls one by one, including every parallel result
Model prefilling	Do not end with an unfinished model message; 3.8 also requires nonempty text in the final user turn
`maxOutputTokens`	Includes thinking and the answer. Argon direct REST allows `32768` per non-streaming hop or `65536` per streaming hop; larger output requires continuation, not simply a larger number for one return
`continuationToken`	Only Argon uses long-output continuation here: with `CONTINUATION`, retain original input, tools and total budget and supply the latest token. To continue after `MAX_TOKENS`, raise the budget above tokens already used. Do not append intermediate output to `contents`

Chat Completions differences

These rules describe Google's official Chat compatibility layer; platform channels may use different conversions.

Field	Difference
`reasoning_effort`	These models use `low/medium/high`, defaulting to `medium`; Argon also accepts `minimal`, mapping it to `low`
`extra_body.google.thinking_config`	Pass Gemini thinking settings such as `thinking_level` and `include_thoughts`; mutually exclusive with `reasoning_effort`
`tool_calls[].extra_content`	Retain `google.thought_signature` and replay complete tool-call history; names and arguments alone lose the signature
`response_format`	Supports the vendor's JSON structured output; use the Gemini Schema subset, not every OpenAI Schema behavior. See Generate Content Reference for native format fields
Long-output continuation	Google's official Chat-compatible endpoint does not support long-output continuation; use native Generate Content or Stream Generate Content

Cache differences

Field	Difference
Implicit caching	Enabled by default without a switch; minimum input prefix for Argon and 3.8 Flash: `4096` tokens. Matching prefixes are still required; hits are not guaranteed
`cachedContent`	Argon rejects explicit-cache creation or references; 3.8 Flash supports them. Google Cloud also supports explicit caching for 3.7 Flash
Chat cache references	Google's compatibility layer references native cache through `extra_body.google.cached_content`, not OpenAI's `prompt_cache_options`. Chat does not add explicit-cache support to Argon
Cache usage	Inspect native `usageMetadata.cachedContentTokenCount` to confirm actual hits; enabled settings alone are not proof

Media input differences

Input	Protocol content block	Difference
Image	Native `inlineData`	Use Base64 data and the correct `mimeType`
File	Native `fileData`	For images, audio, video and PDF; provide a service-readable resource URI, not a local path. Native Files references must come from the corresponding file service
Image	Chat `image_url`	Use supported image URLs or Base64 Data URLs, not native Parts
Audio	Chat `input_audio`	Official examples use Base64 audio and `format`; this does not configure audio generation

Field	Difference
`mediaResolution`	Controls input media processing precision. Request-level and Part-level settings have different structures; per-media settings override request settings. This is not output image resolution
`responseModalities`	All three output text only; use `TEXT`. General Schema fields `IMAGE/AUDIO`, `imageConfig` and `speechConfig` do not enable image or speech generation for these models

Request examples

Replace MODEL_ID with the full platform model ID. Both examples use medium thinking and an explicit budget; they do not automatically continue output.

curl 'https://api.agtcloud.ai/v1beta/models/MODEL_ID:generateContent' \
  -H 'Authorization: Bearer your-api-key' \
  -H 'Content-Type: application/json' \
  --data '{
    "contents": [{"role": "user", "parts": [{"text": "Explain how to troubleshoot an API timeout in three points"}]}],
    "generationConfig": {
      "thinkingConfig": {"thinkingLevel": "MEDIUM"},
      "maxOutputTokens": 32768
    }
  }'