On this page · 8 sections
Summary. Google's Gemini API documentation, checked on 20 November 2026, confirms that from 1 December 2026, 11 days from now, every API request naming gemini-2.5-pro will run on gemini-3.8-flash, and every request naming gemini-2.5-flash will run on gemini-3.5-flash-lite, each billed at the new model's rates. Google deprecated both models in its 2 November 2026 release notes, and neither can be selected in Google AI Studio any more. Gemini 2.5 Flash-Lite is not affected. Your code keeps its old model string, so nothing fails loudly. What changes sits underneath: different default thinking, ignored sampling parameters, stricter turn and function-calling rules, and new grounding charges. For 2.5 Pro users the December bill falls, because 3.8 Flash costs an introductory $0.75 input and $3.75 output per million tokens against $1.25 and $10. On 1 January 2027 it doubles to $1.50 and $7.50. For 2.5 Flash users the headline rates match at $0.30 and $2.50, but 3.5 Flash-Lite defaults to minimal thinking where 2.5 Flash thought by default. Our verdict: change the model string yourself before 1 December, set thinking_level explicitly, and run a golden-set regression this week instead of letting the route decide quality for you.
About this article. This is for engineering leads, platform teams and FinOps owners who call Gemini 2.5 Pro or 2.5 Flash through the Gemini API, directly or through a gateway or SDK wrapper. We checked every model ID, date, price and parameter change on 20 November 2026 against Google's Gemini API deprecations page, pricing page, release notes, model pages, migration guides, thinking guide, rate limits page and model selection guide, and against the Gemini Enterprise Agent Platform lifecycle page on Google Cloud. Rupee figures use European Central Bank reference rates of 19 November 2026. Vendors change dates and prices, so re-check before you commit; this is not legal or financial advice, and no vendor paid for inclusion. Researched with AI assistance and checked against the primary sources listed in the references.
What Google has announced
Two deprecations, one date, two different targets. The table below is taken from Google's deprecations page and release notes as of 20 November 2026.
| Model ID you call | Released | Deprecation notice | Effective date | Requests run on | Billed at |
|---|---|---|---|---|---|
| gemini-2.5-pro | 17 Jun 2025 | 2 Nov 2026 | 1 Dec 2026 | gemini-3.8-flash | 3.8 Flash rates |
| gemini-2.5-flash | 17 Jun 2025 | 2 Nov 2026 | 1 Dec 2026 | gemini-3.5-flash-lite | 3.5 Flash-Lite rates |
| gemini-2.5-flash-lite | 22 Jul 2025 | None | No date announced | gemini-2.5-flash-lite | Unchanged |
Four details shape the plan.
This is a route, not a shutdown. Google's deprecations page defines a shut-down model as one whose endpoint is no longer available. The 2.5 Pro and 2.5 Flash entries work differently: the endpoint keeps answering, but a different model produces the answer and a different price list applies. Google already did this to gemini-3.7-flash and gemini-3.5-flash, whose requests now route to 3.8 Flash and 3.6 Flash. An error would at least tell you something had changed. A route tells you nothing unless you measure.
Access is already restricted. Both models are gone from the AI Studio model picker, and the deprecations page says API access is limited to existing users who have actively used them. A new project or a new team member cannot reproduce a 2.5 baseline from scratch, so capture yours now.
gemini-2.5-flash-lite is not part of this. Google states that 2.5 Flash-Lite is not deprecated and stays on the API until further notice. Do not sweep it into the same change by accident.
Google gives two upgrade answers. The release notes tell you to change your string to the route target, 3.8 Flash or 3.5 Flash-Lite. Google's model selection guide instead recommends gemini-4-hydrogen as the upgrade from 2.5 Pro (or gemini-4-carbon for the hardest problems), and Hydrogen or 3.5 Flash-Lite from 2.5 Flash. The route is a safety net that keeps calls working, not necessarily the right target for your workload. Our Gemini 4 Hydrogen pricing and migration guide covers that option.
What changes in the output
The model pages show the same broad envelope on both sides: a 1,048,576-token input limit, a 65,536-token output limit, text output, and support for function calling, structured outputs, code execution, context caching, Search and Maps grounding, URL context, Batch, Flex and Priority. The differences sit in how the new models think and which request shapes they accept.
| Behaviour | 2.5 Pro and 2.5 Flash | 3.8 Flash (from 2.5 Pro) | 3.5 Flash-Lite (from 2.5 Flash) |
|---|---|---|---|
| Thinking control | thinking_budget (guide also lists low, medium, high) | thinking_level: low, medium, high | thinking_level: minimal, low, medium, high |
| Default thinking | On | medium | minimal |
| minimal level | Not listed | Returns an error | Supported, default |
| temperature, top_p, top_k | Not flagged | Remove (migration checklist) | Deprecated and ignored |
| Prefilled model turn | Not flagged | HTTP 400 error | HTTP 400 error |
| candidate_count | Not flagged | Unsupported | Unsupported |
| Search grounding billing | Per grounded prompt | Per search query | Per search query |
From 2.5 Pro to 3.8 Flash: longer reasoning, not shorter
Google describes 3.8 Flash as its most capable Flash model for long-horizon coding, agents and enterprise workflows. Its migration guide also warns that the model can use more tokens on long and complex tasks by design, taking smaller reasoning steps, calling tools repeatedly and checking its own work, and suggests a lower thinking level for everyday tasks. Teams that used 2.5 Pro for document analysis or code review should expect answers that differ in length and structure, and thinking token counts that move in either direction. The default level is medium, and minimal is not allowed: a request that sends thinking_level set to minimal fails.
From 2.5 Flash to 3.5 Flash-Lite: less thinking by default
This is the change most likely to lower quality without anyone noticing. Google's thinking guide lists 2.5 Flash with thinking on by default and 3.5 Flash-Lite with thinking at minimal by default, which the guide describes as near-off. Google positions Flash-Lite for high-volume extraction, translation and simple data processing, and its migration guide tells teams with subagents, tool calls or multi-step reasoning to raise thinking_level to medium or high to avoid premature tool termination. Google's own benchmark comparison favours Flash-Lite over 2.5 Flash on reasoning and chart understanding (18.0% against 11.0% on HLE, 74.5% against 63.7% on CharXIV), but those scores are vendor-reported and say nothing about your prompts at the default setting.
Request shapes that break
Three changes apply to both targets, because Google says they apply to 3.6 Flash, 3.5 Flash-Lite and every later release.
- Sampling parameters. temperature, top_p and top_k are deprecated and ignored, and Google says future models will reject them with HTTP 400. Any workflow that relied on temperature 0 for repeatable output loses that lever; Google suggests explicit rules in a system instruction instead.
- Prefilled model turns. If the last non-empty turn in a request is a model turn, the API returns a 400 error. Prompts that prefill an opening brace to force JSON, or a label to suppress a preamble, will fail rather than degrade. Structured outputs or a system instruction replace them.
- Function calling and thought signatures. The 2.5 model pages and release notes list function calling and thought_signature validation as breaking changes. On the generateContent API, every FunctionResponse must carry call_id and name, and in stateless conversations you must send thought blocks back exactly as received.
Google's documentation does not say what happens to a thinking_budget value sent in a routed request. Treat that as untested until you have tested it.
What changes in cost
Google bills routed requests at the target model's rates. The paid-tier standard prices below are per million tokens, from Google's pricing page on 20 November 2026.
| Rate, per 1M tokens | gemini-2.5-pro today | 3.8 Flash to 31 Dec 2026 | 3.8 Flash from 1 Jan 2027 |
|---|---|---|---|
| Input, prompts up to 200k | $1.25 | $0.75 | $1.50 |
| Input, prompts over 200k | $2.50 | $0.75 | $1.50 |
| Output incl. thinking, up to 200k | $10.00 | $3.75 | $7.50 |
| Output incl. thinking, over 200k | $15.00 | $3.75 | $7.50 |
| Batch input and output | $0.625 and $5.00 | $0.375 and $1.875 | $0.75 and $3.75 |
| Cached input | $0.125 | $0.075 | $0.15 |
| Cache storage per hour | $4.50 | $0.50 | $1.00 |
3.8 Flash has no long-prompt surcharge, and its cache storage costs a fraction of 2.5 Pro's. From January, though, its input rate is 20% higher than 2.5 Pro's short-prompt rate ($1.50 against $1.25), so input-heavy work such as retrieval over long documents can end up dearer per call. Our Gemini 3.8 Flash January price guide covers the 1 January step in detail.
| Rate, per 1M tokens | gemini-2.5-flash today | gemini-3.5-flash-lite |
|---|---|---|
| Input, text, image, video | $0.30 | $0.30 |
| Input, audio | $1.00 | $0.30 |
| Output incl. thinking | $2.50 | $2.50 |
| Batch input and output | $0.15 and $1.25 | $0.15 and $1.25 |
| Cached input, text | $0.03 | $0.03 |
| Cache storage per hour | $1.00 | $1.00 |
Token rates for 2.5 Flash traffic are flat except for audio input, which drops by 70%. The real cost change comes from token counts: if Flash-Lite thinks less at its minimal default, output spend falls, and if you raise thinking to recover quality, it rises again.
A worked example
Take a workload of 500 million input tokens and 100 million output tokens a month, all prompts under 200,000 tokens. Rupees use ECB reference rates of 19 November 2026 (₹102.55 and $1.0992 per euro), crossed to ₹93.30 per US dollar and rounded to two decimals.
| Scenario | Monthly cost, USD | Monthly cost, INR |
|---|---|---|
| gemini-2.5-pro today | $1,625 | ₹1,51,613 |
| Routed to 3.8 Flash in December | $750 | ₹69,975 |
| 3.8 Flash from January, same tokens | $1,500 | ₹1,39,950 |
| 3.8 Flash from January, 30% more output tokens | $1,725 | ₹1,60,943 |
| gemini-2.5-flash today, or 3.5 Flash-Lite at the same tokens | $400 | ₹37,320 |
The workings: 500 × $1.25 + 100 × $10 = $1,625. In December, 500 × $0.75 + 100 × $3.75 = $750. From January, 500 × $1.50 + 100 × $7.50 = $1,500, and with 130 million output tokens, $750 + 130 × $7.50 = $1,725. For Flash, 500 × $0.30 + 100 × $2.50 = $400 on either model. Each dollar figure is multiplied by 93.30 for the rupee column. The 30% line is a scenario, not a forecast; Google only says 3.8 Flash can use more tokens on complex tasks. Gemini 4 Hydrogen lists the same $1.50 and $7.50, so from January the Hydrogen upgrade costs the same per token as the route.
Grounding is billed differently
This is the cost line that surprises people. On 2.5 models, Grounding with Google Search gives 1,500 free grounded prompts a day, then charges $35 per 1,000 grounded prompts. On Gemini 3 and newer it gives 5,000 free search requests a month across all those models, then charges $14 per 1,000 search queries, and one prompt can trigger several queries. Google also applies a default cap of 1,500 grounding requests a day per project for Gemini 3 and newer on the Gemini Developer API.
Take 1,000 grounded prompts a day for 30 days. On 2.5 Flash that sits inside the free daily allowance, which 2.5 Flash shares with 2.5 Flash-Lite: $0. Routed to Flash-Lite, at one query per prompt, it is 30,000 queries, of which 25,000 are billable: 25 × $14 = $350 a month. At an average of 1.5 queries per prompt it is 45,000 queries, 40 × $14 = $560, and 1,500 queries a day, at the level of the default daily cap. Both figures assume no other Gemini 3 or newer traffic is using the shared monthly allowance. Count queries per prompt in your logs before 1 December.
How to test before 1 December
You have 11 days, and until the route starts you can still call both models side by side. After that, the old model is gone for good.
Step 1: find every call
Search code, configuration, infrastructure templates, prompt registries and gateway routing tables for the two strings. The pattern below matches gemini-2.5-pro and gemini-2.5-flash but not gemini-2.5-flash-lite or the 2.5 TTS and image models.
grep -rnE "gemini-2\.5-(pro|flash)([^-a-z0-9.]|$)" \
--include="*.py" --include="*.ts" --include="*.js" --include="*.go" \
--include="*.java" --include="*.yaml" --include="*.yml" --include="*.json" \
--include="*.env" --include="*.tf" .
Code search misses model names held in environment variables, feature flags, per-tenant database rows and no-code tools, so cross-check against your billing and usage reports by model. If a gateway such as LiteLLM or Cloudflare AI Gateway sits in front of Gemini, make the model a routing alias there; our AI gateway comparison covers the options.
Step 2: capture a 2.5 baseline now
Pick 100 to 300 real requests per use case, including long prompts, tool calls, JSON outputs, multimodal inputs and the awkward edge cases support has seen. Run them on the 2.5 model today and store the outputs, latency, input tokens, thinking tokens and output tokens. This baseline cannot be recreated after 1 December.
Step 3: run the same set on the target at each thinking level
Run the golden set on the target model at every supported thinking level, with temperature and other sampling parameters removed, and record usage per call. Google's thinking guide documents the usage fields used below.
from google import genai
client = genai.Client()
TARGET = "gemini-3.5-flash-lite" # or "gemini-3.8-flash"
LEVELS = ["minimal", "low", "medium", "high"] # 3.8 Flash: drop "minimal"
def run(prompt: str, level: str):
it = client.interactions.create(
model=TARGET,
input=prompt,
system_instruction="Return only valid JSON matching the agreed schema.",
generation_config={"thinking_level": level},
)
u = it.usage
return it.output_text, u.total_thought_tokens, u.total_output_tokens
for level in LEVELS:
text, thought, out = run("Extract the invoice number and total from: ...", level)
print(level, thought, out, text[:80])
Score each output against the baseline with the checks you already trust: exact match for extraction, schema validation for JSON, a rubric or a second model for open text, and a human pass on a sample. Then pick the lowest thinking level that meets your quality bar, as Google itself recommends. Our guide to agent evals in CI/CD shows how to turn this into a gate that runs on every change.
Step 4: fix the request shapes
- Delete temperature, top_p, top_k and candidate_count from every request to the new models.
- Replace prefilled model turns with structured outputs or a system instruction.
- Replace thinking_budget with thinking_level, and never send minimal to 3.8 Flash.
- On generateContent, add call_id and name to every FunctionResponse, and in stateless chats resend thought blocks unchanged.
- If you set max_output_tokens low, note that it caps thinking and output together; a request that hits it mid-thought returns incomplete output but still bills the thinking tokens.
Step 5: switch the string yourself
Set the model ID explicitly to your chosen target and ship behind a flag: 10% of traffic for 24 hours, then 100%, with spend and error alerts on. Aim to finish by Thursday 26 November, so you keep two working days (27 and 30 November) before the route starts on Tuesday 1 December. Then decide on January: either budget for 3.8 Flash at $1.50 and $7.50, or test Hydrogen against the same golden set.
Agent Platform runs on a different calendar
The route described above applies to the Gemini API used through Google AI Studio keys. Google Cloud's Gemini Enterprise Agent Platform keeps its own lifecycle table, and on 20 November 2026 it listed gemini-2.5-pro, gemini-2.5-flash and gemini-2.5-flash-lite with a retirement date of 20 October 2026, which has already passed. The replacements it names also differ: 3.8 Flash or 3.5 Flash for 2.5 Pro, and 3.8 Flash or 3.5 Flash-Lite for the two Flash models, with Gemma 4 as a further option for 2.5 Flash-Lite. If you call Gemini through a Google Cloud project, check that page and your project's logs, not the AI Studio deprecations page.
India-specific considerations
Google lists Gemini API prices in US dollars, so an Indian team's bill moves with the exchange rate as well as with token counts. At ₹93.30 per dollar, the $1,625 monthly 2.5 Pro workload above is ₹1,51,613, before any GST treatment your finance team applies to imported services. Budget the January line on measured 3.8 Flash tokens, not on December's introductory price.
A model change can also matter for governance. If your records under the Digital Personal Data Protection Act 2023, or your internal AI register, name the model that processes customer data, update them when the model changes, even though the model string in your code does not. Voice and call-centre teams that send audio to 2.5 Flash should note that audio input becomes 70% cheaper on Flash-Lite, which may change the case for transcription pipelines.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI and machine learning services team can inventory your Gemini 2.5 calls, run golden-set tests on 3.8 Flash, 3.5 Flash-Lite and Gemini 4 Hydrogen, and ship the model change behind a flag before 1 December. Talk to us at /contact-us/.
Last updated: 20 November 2026.