Gemini 2.5 Reroutes on December 1: What Happens to Your Bill

Nothing in your code breaks on December 1. That is the trap. Google is not shutting off the gemini-2.5-pro and gemini-2.5-flash strings — it is quietly sending them somewhere else, and billing you at the new model's rates. If you only watch for errors, you will miss the month your bill first drops by half, then climbs back, on the same traffic.

Here is the short version, straight from Google's official Gemini API pricing page: effective December 1, 2026, every API request to gemini-2.5-pro automatically routes to Gemini 3.8 Flash and bills at 3.8 Flash rates, and every request to gemini-2.5-flash routes to Gemini 3.5 Flash-Lite at 3.5 Flash-Lite rates. The two models are already marked deprecated and can no longer be picked in AI Studio. What follows is what that does to a real monthly bill — and what the same workload costs on GPT-6, since December is a perfectly good moment to re-price the whole stack.

The rate card, before and after

All figures below are paid-tier Standard rates per 1M tokens, for prompts of 200K tokens or fewer. Google counts thinking tokens as output, so an output-heavy reasoning workload is exactly where a reroute bites hardest.

Two things jump out before any workload math. First, the Pro leg has two price changes, not one: an introductory 3.8 Flash rate through December 31, then a standard rate from January 1 that doubles it. Anyone who budgets December's bill forward into January will under-budget by 2× on that leg. Second, the Flash leg has no headline change at all — $0.30 / $2.50 before, $0.30 / $2.50 after. If your bill moves on the Flash route, it moves because the new model uses a different number of tokens, not because the rate changed.

Worked example: 1,000 requests a month

Take a steady workload: 1,000 requests a month, each 10,000 input tokens and 2,000 output tokens. That is 10M input and 2M output tokens a month, with every request under both Google's 200K prompt band and GPT-6's 272K short-context band, so short-context rates apply throughout. No caching in this example; caching only widens the gaps below, since cached reads get cheaper on the same schedule.

December feels like a gift. January is the bill you actually have to plan for — and at this 5-to-1 input-to-output mix, January still lands 7.7% below today. But that is mix-dependent, and the mix is where people get caught. January's rates move input up $0.25 per million and output down $2.50 per million against today's Pro rates. The two cancel out at exactly 10 input tokens per output token. Run more input-heavy than 10-to-1 — long retrieved context, short answers — and January costs more than today. Run more output-heavy, and January costs less. Check your own ratio in usage_metadata before you assume December's discount survives the new year.

One more quiet saving on the Pro leg: context-cache storage falls from $4.50 per million tokens per hour on 2.5 Pro to $0.50 through December and $1.00 from January on 3.8 Flash. If you hold large cached contexts for hours, that line item drops even while the headline input rate rises in January. Cache-read rates follow the same two-step schedule: $0.125 today, $0.075 in December, $0.15 in January.

The same workload on GPT-6

This is a GPT-6 cost site, so here is the honest comparison, using our site's pricing data at short-context Standard rates (verified against OpenAI's pricing page, November 3, 2026):

Read that against the reroute: December's rerouted Pro bill ($15.00) sits between Luna ($2.00) and Terra ($20.00). January's ($30.00) sits between Terra and 6.1 Sol ($40.00). Luna is not a like-for-like substitute for a Pro reasoning workload — tier choice is a quality decision, not just a price sort — but the spread is the point. If you are re-testing prompts for a new model anyway, which the reroute forces you to do, that is the cheapest moment to also test whether a cheaper GPT-6 tier clears your bar. Our GPT-6 vs Gemini comparison runs the same kind of side-by-side for other workloads.

What to do before December 1

• Switch the model string yourself, this month. Pointing at gemini-3.8-flash or gemini-3.5-flash-lite directly lets you pick the day, run your evals, and compare token counts — instead of waking up to a different model on December 1 with no baseline.
• Record a token baseline now. Log input, output (including thinking), and cached tokens per request for a week on 2.5. The reroute changes the model, so token counts can move even where rates do not — the Flash leg is the case to watch.
• Budget January, not December. December's $15.00 in the example above is a one-month introductory rate. Put $30.00 — and your own input-to-output ratio against the 10-to-1 breakeven — in the forecast.
• Do not confuse this with Argon. December 1 is also the day Gemini 4 Argon doubles from $2/$10 to $4/$20. Same date, different model, opposite direction. If you run both, price them separately.

One honest caveat: a reroute is not a like-for-like price cut. 3.8 Flash is a different model with different defaults and behavior, and if it reasons longer on your prompts, the extra thinking tokens — billed as output — can eat the rate saving. The rates above are exact; your bill is rates times tokens, and only your own traffic settles the second half. Prices verified November 3, 2026 — always confirm on the official pricing page before making decisions.

Sources

• Google Gemini API pricing — 2.5 Pro / 2.5 Flash deprecation notices, reroute targets, and 3.8 Flash / 3.5 Flash-Lite rates.
• OpenAI API pricing — cross-check for the GPT-6 tiers quoted from this site's pricing data.