The Zeroic Journal

Gemini 2.5 Flash deprecation: plan for the Dec 1 reroute

From December 1, Google answers your 2.5 Pro and 2.5 Flash calls with different models. What changes, what it costs in January, and how to pick a replacement.

On this page 6 sections
  1. What the Gemini 2.5 Flash deprecation changes
  2. Rerouted requests change behavior without erroring
  3. The bill drops in December, then rises
  4. Pin a replacement before December 1
  5. How to choose your Gemini replacement
  6. How Better Analyst scaled its AI layer to 1.5M+ users

The Gemini 2.5 Flash deprecation takes effect on December 1, 2026: requests to gemini-2.5-flash start being answered by Gemini 3.5 Flash-Lite, and requests to gemini-2.5-pro by Gemini 3.8 Flash, billed at the new model’s rates. Nothing errors. A different model answers, and on January 1, 2027, Gemini 3.8 Flash’s price doubles.

That’s 14 days from today, November 17. Google announced both deprecations in its Gemini API changelog on November 2, so teams got 29 days of notice. If your product calls Gemini 2.5 from a backend route, a queue worker or an n8n node, the model name in that code will keep working. What sits behind the name won’t be the model you tested.

What the Gemini 2.5 Flash deprecation changes

Google’s deprecations page lists the dates and replacements for the Gemini API:

Model you callFrom December 1, 2026Billed atStatus
gemini-2.5-proRouted to gemini-3.8-flashGemini 3.8 Flash ratesDeprecated
gemini-2.5-flashRouted to gemini-3.5-flash-liteGemini 3.5 Flash-Lite ratesDeprecated
gemini-2.5-flash-liteNo changeIts own ratesNot deprecated, no shutdown date

Three more details on that page change the plan:

  • Both models are gone from AI Studio already. You can no longer pick them in the AI Studio interface, and Google says API access to the 2.5 models is limited to users who have used them before. A new project or a new API key may not get them at all.
  • The TTS previews leave later. gemini-2.5-pro-preview-tts and gemini-2.5-flash-preview-tts shut down on December 18, 2026, with Gemini 3.8 Flash TTS or 3.8 Flash-Lite TTS as replacements. If you run a voice feature, that’s a second date.
  • Image generation has its own date. gemini-2.5-flash-image shuts down on March 15, 2027, replaced by gemini-3.1-flash-lite-image.

If you call Gemini through Vertex AI on Google Cloud, these notices are not your schedule. The Vertex AI model lifecycle page lists October 20, 2026 as the retirement date for gemini-2.5-pro, gemini-2.5-flash and gemini-2.5-flash-lite, and says those dates can be extended but won’t move earlier. Check your Vertex project today; it may already be past its date.

Rerouted requests change behavior without erroring

A reroute is worse than a shutdown in one way: a shutdown fails loudly, and a reroute doesn’t. Your monitoring sees 200 responses. Your users see different answers.

Look at what each reroute does to the tier you picked:

  • 2.5 Pro becomes a Flash model. If you chose Pro for hard reasoning, long documents or careful extraction, you now get Google’s Flash tier. Google describes Gemini 3.8 Flash as its most intelligent Flash model, with a 1M-token context and up to 64k output tokens, but it is still a different model with different failure modes.
  • 2.5 Flash becomes a Flash-Lite model. Google’s own guide calls 3.5 Flash-Lite a strong migration path from 2.5 Flash, and that’s Google’s claim to test against your prompts. Classification and short extraction usually survive the move. Multi-step reasoning is where to look first.

Then the request settings. The Gemini 3 family doesn’t treat your old parameters the way 2.5 did:

  • temperature, top_p and top_k are ignored. Google’s 3.6 Flash and 3.5 Flash-Lite guide says these are deprecated and ignored, and that future generations will return an HTTP 400 when you send them. If you set temperature to 0 to get repeatable JSON, that setting stops doing anything on December 1. Google suggests explicit rules in the system instruction instead.
  • Thinking defaults change. Gemini 3.8 Flash defaults to medium thinking and rejects minimal; 3.5 Flash-Lite defaults to minimal. Both expect thinking_level in place of thinking_budget. A shared config that sends minimal to every Gemini model will fail on 3.8 Flash.
  • Prefilled replies and candidate counts go. The 3.8 Flash checklist says to drop prefilled model turns and candidate_count. If you start the model’s reply with { to force JSON, move that requirement into a structured-output schema.
  • Function calling gets stricter. On generateContent, every FunctionResponse needs call_id and name. Agents that rebuild chat history from plain text are the ones most likely to break.

None of this is hard to fix. The risk is fixing it after users report it.

The bill drops in December, then rises

All prices below are Google’s paid-tier Standard rates per million tokens, from the Gemini API pricing page on November 17, 2026. Output prices include thinking tokens.

ModelInputOutputFrom January 1, 2027
Gemini 2.5 Pro (prompts up to 200k)$1.25$10.00Rerouted
Gemini 2.5 Pro (prompts over 200k)$2.50$15.00Rerouted
Gemini 3.8 Flash$0.75$3.75$1.50 input, $7.50 output
Gemini 2.5 Flash (text)$0.30$2.50Rerouted
Gemini 3.5 Flash-Lite$0.30$2.50No change listed
Gemini 2.5 Flash-Lite$0.10$0.40No change listed

For 2.5 Flash users, text pricing is identical after the reroute, and audio input falls from $1.00 to $0.30. The cost question there is quality, not price.

For 2.5 Pro users, the picture flips at New Year. Take a feature that sends 3,000 input tokens and gets 500 output tokens back, 100 times per active user per month:

  • On 2.5 Pro today: $0.00875 a request, about $0.88 per active user per month.
  • On 3.8 Flash in December: $0.0041 a request, about $0.41 per user.
  • On 3.8 Flash from January: $0.00825 a request, about $0.83 per user.

That example assumes the same token counts on both models, and that’s the part to check. If 3.8 Flash at its default medium thinking bills 1,000 more output tokens per request than 2.5 Pro did, January’s cost rises to $0.01575 a request, about 1.8 times today’s. Pull billed input and output tokens per request from your own logs for a week on each model before you trust any per-user number, including ours.

Long-context work is the exception. Prompts over 200,000 tokens cost $2.50 input and $15 output on 2.5 Pro, and the pricing page lists no separate long-prompt rate for 3.8 Flash, so those requests stay cheaper even after January. If you do cost work across providers, we covered the same per-request method in cutting your OpenAI API bill.

Pin a replacement before December 1

Letting the reroute happen is a decision too, just an untested one. Change the model name yourself, in a branch, and ship it after your tests pass. That way the model in production is the one you evaluated, and the date passes without a surprise.

Your options on the Gemini API, at today’s Standard rates:

  • gemini-3.8-flash - $0.75 input and $3.75 output until December 31, then $1.50 and $7.50. The closest to what 2.5 Pro users get anyway.
  • gemini-3.6-flash - the same rates and the same January 1 doubling, per the pricing page. Defaults to medium thinking.
  • gemini-3.5-flash-lite - $0.30 and $2.50, with no listed change. Where 2.5 Flash traffic lands anyway.
  • gemini-2.5-flash-lite - $0.10 and $0.40, not deprecated. Worth testing for simple classification and routing calls, knowing it is an older model.
  • Gemini 4 Hydrogen - generally available since November 12 at $1.50 input and $7.50 output, the same as 3.8 Flash’s January price.
  • Gemini 4 Carbon - $2.00 input and $10.00 output on introductory pricing through November 30, then $4.00 and $20.00 from December 1. Budget at the higher number.

Whichever you pick, choose a fallback from a different model family too. After December 1, gemini-2.5-pro points at 3.8 Flash, so the old name is no rollback.

How to choose your Gemini replacement

  • You used 2.5 Flash for classification, tagging or short extraction: test 3.5 Flash-Lite at its default thinking and 2.5 Flash-Lite side by side. Keep the cheaper one that passes your evals.
  • You used 2.5 Flash for multi-step answers or tool use: test 3.8 Flash as well as 3.5 Flash-Lite. The Flash-Lite price only matters if the answers hold.
  • You used 2.5 Pro and your prompts are under 200k tokens: pin 3.8 Flash and budget at the January price, plus your measured thinking tokens.
  • You used 2.5 Pro for long documents over 200k tokens: 3.8 Flash is cheaper per token in both months; spend your time on quality checks for long inputs.
  • You set temperature to 0 for repeatable output: move to structured outputs and explicit rules in the system instruction before you switch.
  • You’re on Vertex AI: ignore December 1 and check the lifecycle page for each model your project calls.

How Better Analyst scaled its AI layer to 1.5M+ users

Every AI product outlives the model it launched on. Better Analyst (formerly FormulaBot) is the clearest case on our books: two senior Zeroic engineers built it with the founder, from a rules-based formula engine to an English-to-SQL LLM pipeline and then AI analysis agents, used by 1.5M+ people.

Two pieces of that build make a model change routine instead of risky. A query validation layer checks the SQL the model writes before it runs, so a change in how a model writes queries is caught before it reaches a user’s data. And AI inference runs as its own service, split out when we moved the platform from a monolith to service-oriented architecture, so a model change ships and rolls back on its own, without touching data processing or the user-facing API.

That’s the setup we’d want in place before any reroute date: one place where the model is named, a check on what it returns, and tests that run on the new model before users do. If your Gemini calls are spread across functions with no tests behind them, an AI Audit maps every call site, tests the replacements and re-prices each feature. For a different provider on a similar clock, see our GPT-4o deprecation plan, and for the wider set of LLM features we build, our AI development services.

Frequently asked questions

What happens to gemini-2.5-flash on December 1, 2026?
From December 1, 2026, Google routes every Gemini API request that names gemini-2.5-flash to gemini-3.5-flash-lite and bills it at Gemini 3.5 Flash-Lite rates. Requests to gemini-2.5-pro go to gemini-3.8-flash. Your code keeps working without an error, but a different model answers, so test the replacement before the date rather than after it.
Is gemini-2.5-flash-lite being deprecated too?
No. As of November 17, 2026, Google's deprecations page says gemini-2.5-flash-lite is not deprecated and remains available through the API until further notice, with no shutdown date announced. It costs $0.10 per million input tokens and $0.40 per million output tokens on the paid tier. Vertex AI is different: its lifecycle page lists a retirement date for it.
Is Gemini 3.8 Flash more expensive than Gemini 2.5 Pro?
It depends on the month. Through December 31, 2026, Gemini 3.8 Flash costs $0.75 input and $3.75 output per million tokens, well under 2.5 Pro's $1.25 and $10. From January 1, 2027 it is $1.50 and $7.50, so input costs more than 2.5 Pro's rate for prompts under 200,000 tokens, while output stays cheaper. Thinking tokens bill as output.
Can I roll back to Gemini 2.5 Pro if the new model breaks my app?
No. After December 1, 2026, the gemini-2.5-pro name points to gemini-3.8-flash, so switching back to it changes nothing. Google also says access to the 2.5 models is limited to users who have used them before. If you want a fallback, pick and test a second model now, such as Gemini 4 Hydrogen or a model from another provider.
Does the December 1 reroute apply to Vertex AI?
Google's reroute notice covers the Gemini API documented in Google AI Studio. Vertex AI keeps its own schedule: its model lifecycle page lists October 20, 2026 as the retirement date for gemini-2.5-pro, gemini-2.5-flash and gemini-2.5-flash-lite, and says dates can be extended but not brought forward. Check that page for your project's models.

References

  1. Google AI Studio - Gemini API deprecations
  2. Google AI Studio - Gemini API pricing
  3. Google AI Studio - Gemini API changelog
  4. Google AI Studio - What's new in Gemini 3.8 Flash
  5. Google AI Studio - What's new in Gemini 3.6 Flash and 3.5 Flash-Lite
  6. Google Cloud - Vertex AI model versions and lifecycle

More from the journal

Hear from Prashant within 24 hours.

Prashant Abbi Partner @ Zeroic

Running Gemini 2.5 in production? Our AI Audit tests your prompts on the replacement models, re-prices each feature and hands you a ranked list of fixes before the December 1 switch.

See the AI Audit, or book a 30-minute call