intermediate
Gemini 3.5 Flash keeps its thinking: delete the recap
Quick answer:
On 19 May 2026 gemini-3.5-flash went generally available and became gemini-flash-latest. Two things arrived with it that are not the model: Managed Agents in preview, with stateful agents in Google-hosted Linux sandboxes, and the Antigravity agent, under the identifier antigravity-preview-05-2026.
What has shipped
Two entries in the release notes change what you write. The default thinking effort drops from high to medium. And the model maintains its intermediate reasoning across turns automatically: in Google's words, no API changes are needed.
The same day, Google said low thinking had improved enough to be used for coding and agentic work. A cheap level that works is a level you will use.
One note for the archive: gemini-3.5-flash has been deprecated since 8 October 2026 and requests are routed to gemini-3.6-flash. What follows holds for the model that replaced it.
What changes for reasoning
The two changes pull in opposite directions and are worth reading together. The default effort drops: without touching anything, the model thinks less than it did the day before. And the reasoning is preserved across turns: the model no longer needs you to remind it.
The first takes away something you did not ask for. The second takes away work you were doing yourself.
Until now, a multi-turn prompt carried a paragraph that existed only for the model: "as we agreed, we chose option B because the cost per unit was lower". That paragraph was not for the user. It was a reconstruction of earlier reasoning, written because the model, on the next turn, did not have it.
With automatic preservation, it has it. The paragraph pays tokens to tell the model something it already knows and, worse, it says it in your words rather than the model's. A paraphrase of reasoning is a second version of the decision, and two versions can disagree.
What changes for writing prompts
There is one test for what stays: does this paragraph contain something the model has not seen? If it does not, it is a recap and it goes.
What goes:
- the review of decisions already taken ("we chose B on cost");
- the summary of the plan before each step ("now we do X, then Y");
- the list of constraints that have been in
systemsince turn one; - the explanation of why a rule exists, when the rule is already written down.
What stays:
- what has changed. If the cost ceiling drops on turn four, that gets said, even though the model is carrying the earlier reasoning.
- what the model never saw. The contents of a file, a figure from the database, the result of a call. It is not in the conversation, so it is not in the reasoning.
- the stopping criterion. What counts as done and how it gets checked. That is not reasoning about the problem; it is the definition of the problem.
- instructions that arrive from outside. A policy, a deadline, a limit you did not decide inside the conversation.
text
Turn 4, instead of the recap:
Changes since the previous turn:
- The ceiling per unit drops to [figure]. It applies to everything you
decide from here; what is already decided does not get reopened.
- Still standing: [the constraint that has not changed, only if it is
the one under test on this turn].
If you are unsure, there is a check that needs no judgement: send the turn with the paragraph and without it, and compare. If the answer does not change, the paragraph was doing nothing. If it changes, look at where: almost always it is because the paragraph carried a fact, not because it carried a recap.
What to stop doing
- Rewriting the state of the conversation on every turn. It is the habit automatic preservation makes obsolete, and the one that costs the most tokens in a long session.
- Paraphrasing the model's reasoning. A second version of a decision is a decision that can contradict the first.
- Confusing preservation with memory. It preserves the reasoning of this conversation. What it does not do is bring you facts that were never in it.
- Assuming the effort you had. The default dropped to
mediumon 19 May. If you never wrote it down, the one you had washighand it is not any more. - Leaving the recap in "just in case". A paragraph that changes nothing is not free: it is paid on every turn and it pushes out what does matter.
Sources
Summary
The 19 May 2026 release made two changes that pull in opposite directions: the model thinks less by default and, at the same time, stops needing you to remind it what it already reasoned. What leaves the prompt is the recap; what stays is what the model never saw and what has changed since the previous turn.
Ready to create better prompts?
Try our free AI prompt generator and join thousands of users creating better prompts.