Gemini 2.5 Pro Guide: Pricing Tiers, Retirement and Migration

Gemini 2.5 Pro is the model a lot of production Gemini code was written against, and in October 2026 it is in an awkward place: still served, no longer recommended, and on a retirement clock on at least one of Google's platforms. If you run it today, the useful questions are what it costs you, when you have to move, and what changes when you do.

This guide covers all three. It also covers the one billing detail that catches people out on this model more than any other.

What 2.5 Pro was

Google announced Gemini 2.5 on 25 March 2025 with an experimental 2.5 Pro build, describing it as a thinking model that reasons through a problem before responding. At launch Google reported 18.8% on Humanity's Last Exam without tool use, 63.8% on SWE-bench Verified with a custom agent setup, and a first-place debut on LMArena. Those were vendor figures at the time, and the field has moved a long way since, so read them as history rather than as a current ranking.

The stable gemini-2.5-pro identifier arrived on 17 June 2025, after a run of preview snapshots (03-25, 05-06, 06-05). It takes text, images, audio, video and PDF input, returns text, and has a 1,048,576-token input limit with 65,536 output tokens. Google's model page lists a January 2025 knowledge cutoff, function calling, structured outputs, code execution, context caching and grounding with Google Search among its supported features.

The preview snapshots matter now only because some aggregators still list them, for example google/gemini-2.5-pro-preview pointing at the 06-05 build. Google redirected its 03-25 and 05-06 preview identifiers to the stable model in June 2025 and shut down all three previews in December 2025. There is no reason to target a preview string in new code.

The 200K pricing cliff

Gemini 2.5 Pro is priced in two tiers by prompt length. On Google's paid tier it costs $1.25 per million input tokens and $10 per million output tokens for prompts up to 200K tokens. Above 200K, the rates rise to $2.50 input and $15 output. Output prices include thinking tokens.

The trap is that the tier is decided by the size of the prompt, and the higher rate applies to the whole request. A 190K prompt and a 210K prompt differ by about 10% in size but by roughly double in input cost. If you use the 1M window for whole-repository or long-document work, you are paying the upper tier on every call.

Two levers help. Context caching cuts repeated input to $0.125 per million under 200K, with an hourly storage charge on top. And retrieval keeps prompts under the threshold in the first place. Context window cost trade-offs walks through the arithmetic.

Google also offers 2.5 Pro on its Batch API at half the standard rate, which aggregators expose as google/gemini-2.5-pro:batch. It is the same model on an asynchronous endpoint: you submit jobs and collect results later. For evaluation runs and back-fills it is the obvious choice.

The retirement timeline

This is the part to act on. Google Cloud's model page for 2.5 Pro states that the model is being retired on 20 October 2026 on its enterprise platform. On the Gemini API, a September 2026 changelog entry limited access to the 2.5 models to users who have used them recently; Google says they are not deprecated there and will be served until further notice, and its model page now points new projects at newer models.

Put plainly: if you are on Vertex, the date is days away. If you are on the Gemini API, you have time, but no new project should start here. Model deprecation and migration covers how to run the switch without a fire drill.

Moving to Gemini 3.1 Pro

The natural successor is Gemini 3.1 Pro, still listed as a preview at the time of writing. It keeps the same 1M input and 64K output limits and the same input types, so most of the work is behavioural, not structural. Google's own Gemini 3 migration notes call out a few changes that will bite if you ignore them:

• Thinking control changed. Gemini 3 models use a thinking_level setting rather than a token budget. Sending both in one request returns an error.
• Leave temperature at 1.0. Google warns that lowering it on Gemini 3 can cause looping or worse results on reasoning tasks. Many 2.5 integrations set a low temperature by habit.
• Simplify the prompt. Elaborate chain-of-thought scaffolding written for 2.5 is often redundant with a higher thinking level.
• Re-check document token counts. Default media resolution changed, which can raise PDF token usage.

The price goes up: Google lists 3.1 Pro Preview at $2 input and $12 output per million up to 200K, $4 and $18 above. Whether that is worth it depends on how often 2.5 Pro was failing your tasks. The Gemini 3.1 Pro Preview guide covers the newer model in detail.

When staying put is reasonable

There are honest reasons to delay. A pipeline with a validated prompt and a stable eval score is worth something, and migrating onto a preview model swaps a known quantity for an unknown one. If you are on the Gemini API and still have access, it is defensible to stay while you build an eval set for the replacement.

What is not defensible is staying without a plan. Before you touch anything:

1. Check which platform you call 2.5 Pro through, and that platform's retirement date.
2. Log what share of your requests exceed 200K tokens, because that changes the cost comparison.
3. Build twenty representative tasks and score both models on them.
4. Pin the replacement identifier explicitly; an alias is the wrong thing to depend on mid-migration.

If your 2.5 Pro traffic is mostly short prompts and routine work, also price the move down rather than up. Gemini 2.5 Flash was always the cheaper sibling, and its successors may cover the job for a fraction of Pro's rate.