Gemini 4 Argon: thought-only MAX_TOKENS with preserved signatures, including thinkingLevel=low

Google Cloud Support directed us here because our reproduction uses the Gemini Developer API / Google AI Studio.

We are investigating a long tool-using conversation where the model exhausts its output budget in thinking and returns neither a visible answer nor a function call. We would appreciate guidance on expected behaviour and supported thinking controls.

Environment

  • Native REST request to https://generativelanguage.googleapis.com/v1alpha/models/gemini-4-argon:generateContent, bypassing proxies and the evaluation SDK.
  • History contains 48 function calls with matching tool responses. Historical thought signatures, function names and argument objects were preserved.
  • Signed-history prompt token count: 295,553.
  • Output cap: 8,192 tokens. Signed replays explicitly used thinkingConfig.thinkingLevel=low or high, with includeThoughts=true. No separate numeric thinking budget was imposed.

Native replay results

Tests were run on November 3, 2026; all returned HTTP 200. One attempt per condition, without retries.

Condition Duration Thinking tokens Candidate tokens Result
Signatures preserved, low 91.769s 8,189 3 MAX_TOKENS; thought text only; no answer/functionCall
Signatures preserved, high 103.775s 8,189 3 Same outcome
Signatures preserved, high; remove only 48 empty text parts before function calls 131.389s 8,189 3 Same outcome
Diagnostic control: high; replace historical signatures with skip_thought_signature_validator 12.346s Not reported 241 STOP with a functionCall

The third condition reported 293,504 cached input tokens. The diagnostic control had 254,209 prompt tokens, so it changes historical thought state and prompt size. It is not an equivalent-quality comparison or a proposed workaround.

These capped requests demonstrate budget exhaustion, not an indefinite hang. We understand that thinking consumes output budget; we have not established that this behaviour is a product defect or that a larger cap would fail.

Questions for the Gemini team

  1. Is it expected that thinkingLevel=low can consume essentially the entire 8,192-token output budget without leaving an answer or next tool call on this history? Are there supported controls to bound thinking separately for this exact model?
  2. Are there known degradation or looping cases with long signed tool histories? What history-preservation or compaction rules are recommended without discarding reasoning context?
  3. How should a tool-using client recover from a thought-only MAX_TOKENS response while preserving valid signatures and conversation state?
  4. Both v1alpha and v1beta models.get returned version=3.0, inputTokenLimit=1048576, and outputTokenLimit=32768. Please confirm the supported output ceiling and how it relates to thinking tokens for this model/API.

The complete conversation contains confidential task data and cannot be posted publicly. We have native response IDs for all four requests. Could a Google team member provide a private channel to share those IDs and, subject to authorization, the exact reproduction? No API keys, prompts or signature values are included here.

Hello @Igor_Demin_External ,

Here are the details on how thinkingLevel, output limits, MAX_TOKENS recovery, and thoughtSignature compaction work on gemini-4-argon:

  • Thinking bounds and thinkingLevel=low: On gemini-4-argon, thinkingLevel (low, medium, or high) guides reasoning effort rather than enforcing a hard numeric cap (thinkingBudget is deprecated for this model and cannot be combined with thinkingLevel), so complex multi-step tool histories can still use more than 8,192 thinking tokens even at thinkingLevel: "low". Because maxOutputTokens caps combined thinking and output tokens, raise maxOutputTokens to at least 32,768 (65,536 or higher for long agentic and tool-calling workflows) as outlined in Gemini thinking and Long decode — Set max_output_tokens.
  • models.get outputTokenLimit=32768 and output ceiling: Gemini 4 Argon uses a unified 1,048,576-token window shared across input, thinking, and output. In models.get, outputTokenLimit: 32768 reflects the single-hop cap for one direct non-streaming REST generateContent response (65,536 per hop on streamGenerateContent); when maxOutputTokens is set higher (up to 1,048,576 minus input tokens), Long decode continues generation across hops via finishReason: "CONTINUATION" and continuationToken.
  • Recovering from a thought-only MAX_TOKENS response: When a gemini-4-argon response stops with finishReason: "MAX_TOKENS", the candidate still includes a continuationToken. As documented in Resume when the model uses all output tokens, resend the exact same original request (contents and tools unchanged, without appending the partial thought-only candidate) with top-level continuationToken set to that token and a higher maxOutputTokens (above the tokens used so far) to resume from the stopping point and obtain the completed functionCall and thoughtSignature.
  • History preservation and compaction: Per Thought signatures — Function calling, strict thoughtSignature validation applies only to functionCall steps within the current turn (after the most recent user message containing standard text content rather than functionResponse). Once a multi-step tool turn completes, you can compact older completed turns into a concise text summary before starting a new user text turn while keeping exact thoughtSignature parts intact on the active turn’s functionCall steps.

If you still run into unexpected behavior after raising maxOutputTokens (32,768+), please let us know here and we can follow up via forum private message (DM) to collect the four responseIds.