On 1 December 2026, every Gemini API call to gemini-2.5-pro is served by gemini-3.8-flash, and every call to gemini-2.5-flash by gemini-3.5-flash-lite, billed at the new model's rates. The API changes that come with those models can break function calling and thinking settings. Don't let it happen by default: switch the model name in a staging copy this month, run your tests, and budget the new bill.
What changes on 1 December, and what can you not do?
Google offers no way to pin the old model past that date. Its changelog entry of 2 November announces both routes, and the deprecations page lists 1 December as the shutdown date for both.
Both models are already gone from AI Studio's model picker, and since 18 September access has been limited to users who have actively used them. Google's 18 September line that the 2.5 models were "not deprecated and will continue to be served until further notice through the API" is stale.
What will the new bill be?
Google's pricing page lists these standard-tier rates per 1M tokens. Output prices include thinking tokens.
| Model | Input | Output |
|---|---|---|
gemini-2.5-pro (to 200k prompt) | $1.25 | $10.00 |
gemini-2.5-pro (over 200k) | $2.50 | $15.00 |
gemini-3.8-flash, to 31 Dec 2026 | $0.75 | $3.75 |
gemini-3.8-flash, from 1 Jan 2027 | $1.50 | $7.50 |
gemini-2.5-flash (text/image/video input) | $0.30 | $2.50 |
gemini-3.5-flash-lite (text/image/video input) | $0.30 | $2.50 |
If you call gemini-2.5-pro
December is cheaper: input drops 40% and output 62.5%. January erases most of it. Input ends up 20% above today's rate, and output stays 25% below. Our Gemini 3.8 Flash pricing piece covers that 1 January step in full.
An illustrative workload: 100M input and 20M output tokens a month, prompts under 200k tokens.
- Today on 2.5 Pro: $325 a month.
- December on 3.8 Flash: $150.
- From January: $300.
If your prompts run past 200k tokens, you pay the higher 2.5 Pro tier today, and 3.8 Flash lists no context tiers at all, so the saving is larger.
If you call gemini-2.5-flash
For text, image and video input, nothing changes on price. The same 100M and 20M workload costs $80 before and after (100 x $0.30 plus 20 x $2.50). Audio input drops from $1.00 to $0.30. What changes is the model answering.
Which code breaks, and how do you test it?
Google's changelog names three breaking areas: function calling with thought_signature validation, thinking_level replacing thinking_budget, and removed sampling parameters and prefilled model turns. The only way to learn which of your calls are affected is to run them against the new model.
A test plan for a staging copy:
- Change the model name to
gemini-3.8-flash(orgemini-3.5-flash-litefor Flash callers). - Run your function-calling flow end to end, since
thought_signaturevalidation applies to function calling, which agent loops depend on. - Replace
thinking_budgetwiththinking_levelwherever you set it. - Delete sampling parameters and prefilled model turns from requests.
- Replay a sample of production prompts and compare answers, latency and billed output tokens against your current model.
Step 5 answers the question Google doesn't. It publishes no quality, latency or thinking-token comparison between 2.5 Pro and 3.8 Flash, so your own replay is the only evidence. Thinking tokens are billed as output, which makes the replay's output-token count the number that settles your December and January bill.
Migrate, accept the reroute, or leave Gemini?
Choose by what the workload does.
- Migrate now, to
gemini-3.8-flashif you call Pro orgemini-3.5-flash-liteif you call Flash. It's our recommendation for anything customer-facing or agentic. You pick the date, you see the failures in staging, and you aren't debugging in production on 1 December. - Move to
gemini-2.5-flash-lite, a separate model the deprecations page lists as not deprecated, with no shutdown date announced. It's worth considering only if its quality is enough for the task. Check its rates on the pricing page first. - Accept the reroute and monitor. It's the least effort today and the riskiest for agents, since the first sign of a break is a failed run in production. Reasonable for low-stakes batch jobs.
- Move to another vendor, but only if staging shows the breaking changes cost more than a vendor switch. What Sonnet 4.5's retirement breaks is the same deadline-driven migration at another vendor, with its own list of breaks, so compare both lists first.
What this means for your AI budget and release calendar
Put two dates in the plan. Before 1 December, the migration is done and the function-calling tests pass. On 1 January, if you're on the Pro route, the AI line rises again, so forecast December and January separately instead of carrying December's saving into the new year.
Another Google price change lands on 1 December, covered in our Gemini 4 Argon pricing piece. If you use both, budget them together.
Ask one engineer to own the staging replay this month, and have them report quality, latency and billed output tokens against the old model. That report is what lets you choose between the options above on evidence. For more model changes with dates attached, the AI news hub tracks them as they land.
FAQ
Can I keep using gemini-2.5-pro after 1 December?
Not under that name. Google lists 1 December 2026 as the shutdown date, and calls are served by gemini-3.8-flash and billed at its rates. The deprecations page offers no pin or extension.
Will the new model cost more?
For gemini-2.5-flash callers on text, image and video, no: the rates match gemini-3.5-flash-lite. For gemini-2.5-pro callers, December costs less and January costs more on input than today ($1.50 against $1.25 per 1M tokens), though output stays cheaper.
What is thinking_level?
It's the setting that replaces thinking_budget on the new models, per Google's changelog. If your requests set thinking_budget, change them in staging and rerun your tests.