Gemini 2.5 Pro and Flash: change the string before December 1

Verdict, November 4, 2026. OP if a request you still send names gemini-2.5-pro or gemini-2.5-flash. Change the model string, and change the request, before December 1, 2026. Google's November 2 release note says those calls keep the old string and run a different model. Mid as a way to keep Pro quality on a Flash rate. The route is the model change. Mid if the only 2.5 id you call is gemini-2.5-flash-lite. That id is outside the November 2 note. Mid as a reason to leave the Gemini API for Claude or Codex, and Mid as a reason to move Gemini 4 Argon. Argon's December 1 change is a price, on a different row of the same pricing page. No letter moves on the Frontier Models board. Argon stays A, on the October 15 placement. S stays locked. OP or Mid has not sent a 2.5 request.

Pages below were read November 4, 2026. They are the Gemini API docs. This pass did not read a Vertex page.

The string stays. The model does not.

NERF The release notes for November 2, 2026 say gemini-2.5-pro and gemini-2.5-flash are deprecated, and neither can be selected in Google AI Studio. Effective December 1, 2026, requests to gemini-2.5-pro route to gemini-3.8-flash and bill at Gemini 3.8 Flash rates. Requests to gemini-2.5-flash route to gemini-3.5-flash-lite and bill at Gemini 3.5 Flash-Lite rates. Update the string before that date. The deprecations page repeats the route, limits gemini-2.5-pro to existing users who have actively used it, and says the same active-use limit covers the 2.5 models. It also says gemini-2.5-flash-lite is not deprecated, with no shutdown date in the table. Sources: release notes (opens in a new tab), deprecations (opens in a new tab).

December 1 is a swap on the id you already ship. That page defines a shutdown as the endpoint being turned off, and it says a Shutdown-column date is the earliest possible retirement, with the exact date sent later. Both rows still say December 1, 2026, and the replacement cell describes the automatic route. The callout is the instruction: that day the request is billed and answered as the other model. Source: deprecations (opens in a new tab).

The September 18, 2026 entry is still lower on the same release notes. It says the 2.5 models are not deprecated and will be served until further notice. November 2, above it, is the plan. Source: release notes (opens in a new tab).

What the bill actually does

PRICE These figures are the standard paid columns, read November 4, 2026. The November 2 note says "Gemini 3.8 Flash rates" and "Gemini 3.5 Flash-Lite rates" and does not name Standard, Batch, Flex, or Priority. gemini-2.5-pro is $1.25 per million input tokens and $10.00 per million output tokens at 200k prompt tokens or fewer, and $2.50 and $15.00 above that. Output includes thinking tokens. gemini-3.8-flash is $0.75 and $3.75 through December 31, 2026, then $1.50 and $7.50 starting January 1, 2027, with one rate on each side. Source: pricing (opens in a new tab).

A lower rate can still cost more. The 3.8 Flash guide says the model can use more tokens on longer and harder tasks, by design: smaller reasoning steps, tool calls, and checks. Lower the thinking level when the workflow does not need that. OP or Mid did not measure a task. Source: 3.8 Flash guide (opens in a new tab).

PRICE gemini-2.5-flash standard paid input is $0.30 per million tokens for text, image, and video, and $1.00 for audio. Output is $2.50 per million, thinking included. gemini-3.5-flash-lite standard paid input is $0.30 per million for text, image, video, and audio. Output is $2.50. On text, the two standard rates match. Audio input is the sticker that moves, from $1.00 to $0.30, once the route is what you are billed. Source: pricing (opens in a new tab).

VENDOR CLAIM The 3.5 Flash-Lite guide calls that model a strong migration path from Gemini 2.5 Flash and prints two comparisons: HLE 18.0% versus 11.0%, and CharXIV 74.5% versus 63.7%. The page gives no date, no sample size, and no method. OP or Mid did not run either test. Source: 3.6 and 3.5 Flash-Lite guide (opens in a new tab).

The request that returns 400

NERF A new model string with the old body is how a working client breaks. Replace thinking_budget with thinking_level. On Gemini 3.8 Flash, minimal returns an error. Use low, medium (the default), or high. Remove temperature, top_p, top_k, candidate_count, and prefilled model turns. The 3.6 page, which the 3.8 checklist points at, says those three sampling parameters are ignored, and that supplying them returns HTTP 400 in future model generations. A request whose last non-empty turn is a model turn returns HTTP 400. Sources: 3.8 checklist (opens in a new tab), API changes (opens in a new tab).

On gemini-3.5-flash-lite, the default thinking_level is minimal. Leave it for extraction, routing, or classification. Set medium or high for a subagent that uses tools, code execution, or several steps. The guide's reason is to keep tool use from stopping early. The same removals apply. Source: Flash-Lite checklist (opens in a new tab).

Function calling and thought_signature validation are the other breaking lines in the November 2 note. The checklist says to put multimodal assets inside the response payload, to separate inline instructions with a blank line (\n\n), and, on generateContent only, to put call_id and name on every FunctionResponse. A Malformed_Function_Call tied to text before a tool call has its own workarounds section. This article stops at that pointer. On the Interactions API, store: true plus previous_interaction_id lets the server keep thought blocks. A stateless resend must return every thought block unchanged, including signatures on built-in tools such as Google Search. On generateContent, the signature is metadata on a part, including a functionCall part or the last part. That page prints no replay sample for it. Sources: 3.8 checklist (opens in a new tab), thought signatures (opens in a new tab).

interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input=prompt,
    generation_config={"thinking_level": "medium"},
)

That shape is the 3.8 guide's request. prompt is a name. OP or Mid did not run it. A Flash call that should land on Flash-Lite uses gemini-3.5-flash-lite, where minimal is allowed. The same checklist offers /gemini-api-dev migrate my app to Gemini 3.8 Flash for an agent that already has the skill. Read the diff. The list above is the edit. Source: 3.8 guide (opens in a new tab).

Who should leave the string alone

Keep gemini-2.5-flash-lite until you pick a replacement on purpose. December 1 is also the day Gemini 4 Argon standard paid rates go from $2.00 and $10.00 per million tokens to $4.00 and $20.00. That is a price edit on gemini-4-argon, separate from this route. The 3.8 guide says the managed Antigravity agent now uses Gemini 3.8 Flash, and that the Antigravity SDK defaults to it. That sentence is a default, and it is a vendor statement. Antigravity with Gemini 4 Argon stays A on the AI Coding Agents list. Gemini 4 Argon stays A on the Frontier Models list, on the October 15 placement. The HLE line is a vendor claim. S stays locked. No letter moves. Sources: pricing (opens in a new tab), 3.8 guide (opens in a new tab), deprecations (opens in a new tab).

What to do today

Search the repo for gemini-2.5-pro and gemini-2.5-flash. Keep gemini-2.5-flash-lite until you pick a replacement on purpose.
Point Pro at gemini-3.8-flash with thinking_level of low, medium, or high. Point Flash at gemini-3.5-flash-lite, and raise that level when the call uses tools.
Delete thinking_budget, temperature, top_p, top_k, candidate_count, and any model turn you prefilled. On a stateless history, send thought signatures back unchanged.
Budget 3.8 Flash as a lower rate that can spend more tokens through December 31, 2026, then the January 1, 2027 rate. Leave Argon's December 1 price on its own row.

The weekly patch is on beehiiv: subscribe free (opens in a new tab).

The OP or Mid Team

Sources

Checked November 4, 2026.