Gemini 2.5 calls re-route on 1 December; Pro users' input price is 20% higher by January

01A re-route, not a shutdown

Google’s deprecations table lists 1 December under “Shutdown date”, but the same row says requests “will automatically route” to the replacement.³ So code that names the old model keeps getting answers, from another model, unless it uses something the new one rejects. Google did the same on 8 October, when calls to 3.7 Flash moved to 3.8 Flash and calls to 3.5 Flash to 3.6 Flash.¹ Access to the 2.5 models has been limited since 18 September to users who had “actively used them in the past”; gemini-2.5-flash-lite is not deprecated and stays available “until further notice”.¹³

This covers the Gemini Developer API, the one behind Google AI Studio. Google Cloud’s Agent Platform, formerly Vertex AI, runs a separate timetable: its documentation, updated on 7 November, gives 2.5 Pro, 2.5 Flash and 2.5 Flash-Lite a retirement date of 20 October 2026, and the pages we read describe no automatic re-route. It notes that listed dates “refer to retirement of support” on that platform.⁴ If your calls go through Google Cloud, check there rather than assume the December date.

02What each call will cost

A separate change arrives the same day. Gemini 4 Argon, released on 13 October, doubles on 1 December from US$2 to US$4 per million input tokens and from US$10 to US$20 for output, the end of a price Google had published with its expiry date. The same rise appears on Agent Platform’s price list.¹²⁵ For 2.5 Flash users, the headline rates do not change: US$0.30 input and US$2.50 output on both models, though audio input falls from US$1.00 to US$0.30.²

03What to test before 1 December

Token counts may not stay equal: thinking tokens are billed as output on both models, and 3.8 Flash’s default thinking level is medium.²⁶ Google tells developers to update the model string before 1 December and calls these changes breaking:¹

• Thinking settings. thinking_level replaces thinking_budget.⁶
• Sampling parameters. temperature, top_p and top_k are “deprecated and ignored”, so an application that relies on them loses that control without an error; Google says future generations will return one.⁷
• Prefilled model turns. A request ending in a model turn returns an HTTP 400 error.⁷
• Function calling. Google lists thought signature validation and, on generateContent, call_id and name on every response.¹⁶
• Grounding allowance. 2.5 Pro and 2.5 Flash have 1,500 free grounded prompts a day, then US$35 per 1,000. Gemini 3 and later share 5,000 free search requests a month, then US$14 per 1,000, and Google says one request may run more than one search.²
• Output quality. 2.5 Pro users move from a Pro model to a Flash one, and 2.5 Flash users to a Flash-Lite one. Run your own evaluation set; prices do not measure this.

Google does not say, in what we read, which of these will fail on a re-routed request and which will be tolerated, so test each one against the new model by name.