Google announced the retirement of Gemini 2.5 Flash, and developers who built on it are asking the company to keep the model online longer. A discussion on Google AI’s official forum gained immediate consensus (120 points on Hacker News, 80 comments): users running 2.5 Flash in production report that newer models in the family don’t measure up on their use cases.

Developers’ internal benchmarks show that Gemini 3 Flash doesn’t maintain the same performance, even after adapting prompts to the new guidelines. For some workflows the successor model costs three times more, for others it arrives too late (600-700ms versus 300-400ms for 2.5 Flash in Australia, where 3.5 Flash has no local deployment). Low latency isn’t a luxury: below 400ms a voice agent works, above it breaks.

Why this matters to you. If you’re building on third-party APIs, this story is the reminder you need: the models you use today can disappear tomorrow, and the official replacement might not do your job. The community is asking for continuity, but the decision isn’t theirs. If a critical piece of your system depends on a specific model, you need a tested plan B ready, not hope that the vendor changes course.

In detail

What came before

Gemini 2.5 Flash was for months the reference model for anyone seeking a balance between quality, latency, and cost: capable enough for complex tasks, fast enough for voice agents, economical enough for high volumes. Not the most powerful in the family, but the most reliable for workloads demanding consistency rather than peaks of intelligence.

Many teams chose it for exactly this reason: a model that does one thing well and does it the same way every time is easier to orchestrate than a brilliant but unpredictable model. They put it in production, calibrated their prompts, built pipelines that depend on its latency and cost.

What changes

Google announced the retirement without offering a replacement that maintains the same quality/latency/cost ratio. Gemini 3 Flash doesn’t keep pace in user tests, and 3.5 Flash costs three times as much. For teams working in regions where local deployment matters (Australia in particular), latency doubles or triples, making the model unusable for real-time use cases.

The community is asking Google to extend the model’s life, at least until a true replacement is available. The request is reasonable, but there’s no guarantee it will be accepted: model providers’ roadmaps follow internal logic that rarely aligns with production timelines of users.

Implications for builders

This story is a textbook case of a risk many underestimate: model stability isn’t guaranteed. An API can disappear, a model can be retired, a price can change. If your system depends on a specific model without tested alternatives, you have an undocumented critical dependency.

As we discussed when Zuckerberg admitted agents run slower than expected, provider development pace doesn’t always align with builders of real applications. The pressure to ship new models can lead to retiring old ones before replacements are ready for all use cases.

What to do

If you use Gemini 2.5 Flash in production, you have two paths:

1. Test alternatives immediately. Don’t trust public benchmarks: take your real use cases, the ones the model must handle every day, and verify the proposed replacement can handle them. If it can’t, the problem is yours to solve, not Google’s.

2. Have a plan B with a different provider. Not for ideology, but because if a single model is a critical point of failure, you need to know where to go if it disappears. Testing an alternative on another provider (OpenAI, Anthropic, a self-hosted open-weight model) gives you a real option, not just wishful thinking.

The broader lesson: if you build on third-party models, evaluate without benchmarks your specific use cases, and track which models actually handle them. There’s no substitute for testing on your data.