intermediate
Gemini 4 Argon: thinking_level and the 400s from prefill
Quick answer:
Argon went generally available on 13 October 2026 and is the only model in the fourth generation: there is no Gemini 4 Pro, no Gemini 4 Flash, no Gemini 4.0. What changes in the prompt is less than it looks and more than you feel: a renamed control, three parameters that are ignored, and a turn you can no longer fill in.
What has shipped
gemini-4-argon, GA on 13 October 2026. The models page lists no other Gemini 4, neither Pro nor Flash, so "Gemini 4" and Argon are the same thing.
A unified 1,048,576-token window, where input, thinking and output all come out of the same pool. Maximum output is 65,536 tokens in a synchronous response. Reasoning has three levels —low, medium, high— with medium as the default; Argon has no minimal. And caching is implicit only: explicit caching is not supported.
One note on the documentation, because it changes how the rest reads: Google's prompting page does not name Argon anywhere. "Gemini 4" does not appear, nor does "Argon", nor thinking_level, and its newest section is still titled "Gemini 3". What sits underneath are generation-3 rules that Argon inherits; there is no Argon-specific prompting guide published on that page.
What changes for reasoning
Reasoning is governed by thinking_level, not by a number: thinking_budget was replaced by an enum. The default varies by model, and that is the part that breaks other people's prompts: gemini-4-argon, gemini-3.8-flash and gemini-3.6-flash start at medium; gemini-3.5-flash-lite at minimal; gemini-3.1-pro-preview and gemini-3-flash-preview at high.
Thinking and the answer share a ceiling. max_output_tokens counts thought tokens too: if the model stops while reasoning, the response comes back truncated or empty with status incomplete. Lowering the ceiling to save money is the wrong move; the level is what comes down.
Thinking blocks go back exactly as they arrived, signatures included. Removing or editing them breaks the continuation of the reasoning, which makes the history one more thing you do not edit.
What changes for writing prompts
Sampling parameters are ignored. The Gemini 4 migration checklist settles it in one line: "If provided, they are ignored." Google's replacement for determinism is a system instruction with explicit rules, not a lower temperature. In practice: a grid search over those three numbers no longer measures the prompt, and the remedy the page used to give for a safety-filtered response —raising the temperature— does nothing.
text
system_instruction:
Always answer in three bullets, in this order: what happened, what it
means, what to do next. Maximum 20 words per bullet. If a fact is
missing, write "no data" in the bullet that needs it.
A prefilled model turn returns 400. If the last non-empty turn in the request is the model's, the API rejects it. The two classic reasons for prefilling —suppressing the preamble and forcing JSON— have a documented replacement: system_instruction and structured outputs. The prefix as text survives, but in the user turn.
With long inputs, context first and instructions last. A transition phrase goes between the material and the question to separate the evidence from the request, and with a one-million-token window that distance is not a style detail.
text
<context>
[paste the contract, the report or the files here]
</context>
Based on the information above, tell me in five bullets which obligations
the supplier accepts and which do not appear in the document.
What you stop doing
- Sending
thinking_budget. It no longer exists: the control isthinking_level. - Touching
temperature,top_portop_k, and worse, running a grid search to find "the setting". What you tune that way is noise, not the answer. - Trusting a prefilled model turn. It is a 400 on every call.
- Lowering
max_output_tokensto cut cost with reasoning on. It cuts the answer. - Assuming the rules on the prompting page are Argon's. They are generation 3's, and they are still unreconciled.
Sources
Summary
Argon replaces `thinking_budget` with `thinking_level` and a per-model default, puts thinking and output under the same token ceiling, and rejects prefilled model turns. Three habits go: the numeric budget, sampling, and prefill.
Ready to create better prompts?
Try our free AI prompt generator and join thousands of users creating better prompts.