Gemini 2.0 Flash is deprecated: what migration actually costs you

The full 2.0 family is going away

Google deprecated four models on February 18, 2026, all with a June 1 shutdown:

This is the last of the 2.0 generation. The experimental and preview variants were already gone - Flash Thinking shut down December 2025, Flash Live shortly after. June 1 is the final sweep.

Google says June 1 is "the earliest possible" date and will send advance notice before the actual cutoff. That said, planning for June 1 is the safe assumption.

What you are moving to

Here are the current Flash-tier options side by side. The faded rows are the deprecated models.

Text input pricing. Audio rates differ. Gemini 2.5 Pro input price is for prompts under 200K tokens. Source: ai.google.dev/gemini-api/docs/pricing

What migration costs per month

Take a typical production workload: 50M input tokens and 50M output tokens per month. Document Q&A, daily summaries, something in that range.

The last row is not a hypothetical worst case. If you swap the model string and leave everything else unchanged, thinking tokens will fire by default and your output costs will be much higher than the $2.50/1M headline. Google's own docs call this out explicitly.

The 2.5 Flash-Lite row deserves more attention. It matches 2.0 Flash pricing exactly and it is GA-stable. For most workloads that are doing extraction, classification, or summarization, it will be the correct landing point.

The thinking token billing problem

Gemini 2.5 Flash has a thinkingBudget setting. When thinking runs, those tokens bill at the full output rate - $2.50 per million. A 500-token response with 1,000 thinking tokens bills as 1,500 output tokens, not 500.

Google called this out in the 2.5 Flash launch post: "If you want to keep the lowest cost and latency while still improving performance over 2.0 Flash, set the thinking budget to 0." We tested this ourselves and it works as described - thinking-off 2.5 Flash is noticeably faster than thinking-on, and the output is identical for straightforward classification and extraction tasks.

const response = await genAI
  .getGenerativeModel({ model: "gemini-2.5-flash" })
  .generateContent({
    contents: [{ role: "user", parts: [{ text: prompt }] }],
    generationConfig: {
      thinkingConfig: { thinkingBudget: 0 },
    },
  });

Which path to take

The right migration depends on what 2.0 Flash was actually doing.

What else changes

The 8K output cap on 2.0 Flash was a real constraint for anything producing long code or documents. Both 2.5 models raise that to 65K. If your app was hitting truncation errors, the migration fixes that problem at the same cost or better.

Context caching can offset part of the cost increase

If your workload sends the same large system prompt or document repeatedly, context caching changes the math. The cache read rate is much lower than fresh input.

On 2.5 Flash, cache reads cost $0.03/1M - 10% of the fresh input rate. If you have a 100K token system prompt queried 1,000 times a day, caching reduces the effective input cost on those tokens from $30 to $3 per day. That can close most of the gap for heavy-context workloads.

FAQ

Sources

• Google Gemini API Deprecations
• Google Gemini API Pricing
• Gemini 2.5 Flash model page
• Gemini 2.5 Flash-Lite model page
• Start building with Gemini 2.5 Flash (Google Developers Blog)
• Google Gemini API Changelog