What to Do Before Gemini 2.5 Pro and 2.5 Flash Are Retired on December 1, 2026

Shutdown: Tuesday, Dec 1, 2026. Google's Gemini API retires gemini-2.5-pro and gemini-2.5-flash. From that date, requests to gemini-2.5-pro are automatically routed to gemini-3.8-flash and requests to gemini-2.5-flash to gemini-3.5-flash-lite, billed at the new model's rates. Your calls keep answering, but with a different model and different parameter rules.

Free AI model shutdown alerts: get an email only when a model you might use is being retired. No digest, no spam. Subscribe at aideprecations.com/subscribe (double opt-in).

Affected model IDs (from Google)

Model ID	Released	Shutdown	Recommended replacement	After Dec 1, requests route to
`gemini-2.5-pro`	Jun 17, 2025	Dec 1, 2026	`gemini-3.8-flash`	`gemini-3.8-flash`, billed at Gemini 3.8 Flash rates
`gemini-2.5-flash`	Jun 17, 2025	Dec 1, 2026	`gemini-3.5-flash-lite`	`gemini-3.5-flash-lite`, billed at Gemini 3.5 Flash-Lite rates

Source: Gemini API deprecations (read Nov 10, 2026).

Both models are deprecated and no longer selectable in Google AI Studio. API access to gemini-2.5-pro is limited to existing users who have actively used it.
Not affected: gemini-2.5-flash-lite "is not deprecated and remains available through the API until further notice."
Vertex AI is different: on Google Cloud (Gemini Enterprise Agent Platform), these IDs already reached their retirement date on Oct 20, 2026 per Google Cloud model versions. If you use Vertex, check that page.

Why "auto-routing" is still a breaking change

Cost: routed requests are billed at the new model's rates (pricing). Google lists introductory pricing for Gemini 3.8 Flash through Dec 31, 2026.
Sampling parameters: temperature, top_p and top_k are deprecated and ignored on Gemini 3.6 Flash / 3.5 Flash-Lite and later; Google says future generations will return HTTP 400 (API changes).
Prefilled model turns: a request whose last non-empty turn is a model turn returns HTTP 400.
Thinking: thinking_budget is replaced by thinking_level. gemini-3.8-flash supports low, medium (default) and high, not minimal. gemini-3.5-flash-lite defaults to minimal.
Function calling: audit it; with generateContent every FunctionResponse must include call_id and name.

Step 1: Find every call site

rg -n --hidden -g '!node_modules' -e 'gemini-2\.5-(pro|flash)\b' . | rg -v 'flash-lite' rg -n --hidden -g '*.{yml,yaml,json,toml,tf,tfvars}' -g '.env*' -e 'gemini-2\.5-(pro|flash)' .

Step 2: Pick the replacement explicitly

gemini-2.5-pro → gemini-3.8-flash (Google checklist)
gemini-2.5-flash → gemini-3.5-flash-lite (Google checklist)

Switch on your schedule instead of letting the Dec 1 routing do it, so you can test first.

Step 3: Update the code (google-genai SDK)

import os from google import genai from google.genai import types client = genai.Client() # reads GEMINI_API_KEY # was: "gemini-2.5-pro" (gemini-2.5-flash -> "gemini-3.5-flash-lite") MODEL = os.environ.get("GEMINI_MODEL", "gemini-3.8-flash") resp = client.models.generate_content( model=MODEL, contents="Summarize this ticket in 3 bullets: ...", config=types.GenerateContentConfig( system_instruction="Output only the bullets.", # removed: temperature=0.2, top_p=0.9, top_k=40 (ignored on Gemini 3.x) # was: thinking_config=types.ThinkingConfig(thinking_budget=1024) thinking_config=types.ThinkingConfig(thinking_level="low"), ), ) print(resp.text, resp.model_version)

Interactions API (the style Google's current docs use):

interaction = client.interactions.create( model="gemini-3.8-flash", input="Summarize this ticket in 3 bullets: ...", system_instruction="Output only the bullets.", generation_config={"thinking_level": "low"}, ) print(interaction.output_text)

Upgrade first: pip install -U google-genai.

Step 4: Remove prefilled model turns

If you ended contents with a model turn (e.g. "Translation:" or "{" to force JSON), that now returns HTTP 400. Use system_instruction or structured outputs instead.

Step 5: Testing checklist

Frozen prompt set on old vs new: quality, refusals, length, latency.
Compare thinking_level low / medium / high; for gemini-3.5-flash-lite agents with tools, Google suggests medium or high.
With temperature ignored, re-check tests that relied on low-temperature output.
Function calling: validate against schemas, include call_id + name, watch for Malformed_Function_Call.
Parse 100% of structured output with your production parser.
Preserve thought signatures across multi-turn calls.
Price real token counts (including thought tokens) at the new rates.
Log model_version and alert if it isn't the model you set: after Dec 1 that's how silent routing shows up.

Step 6: Roll out before Dec 1

Canary 5–10%, then 100%, with the model ID in config.
After cutover, fail CI if gemini-2.5-pro or gemini-2.5-flash (not -lite) appears in the repo.