TL;DR (May 2026): Google removed Gemini 2.5 Pro, 3 Pro, and 3.1 Pro from the free tier on April 1, 2026. Only Flash (15 RPM / 1,500 RPD) and Flash-Lite (30 RPM) are still free. If you were building on Pro and your bill just appeared, here's what to actually do instead of panicking.
April 1 was a real change, not a joke. Google flipped the switch and three families of Gemini Pro models went paid-only. The 3.x generation is paid even in preview. Flash and Flash-Lite stayed free, with reduced daily quotas.
I watched the GitHub issues fill up in real time. The pattern was the same one you've seen before with API pricing shifts: a wave of "where did the free tier go" posts, then a wave of "my agent costs $40 a day now" posts, then a slower wave of teams rewriting their model selection logic. Most teams ended up somewhere reasonable. Some teams over-corrected and broke their product. This is the field guide for the second group.
What's actually free now
The current shape of the Gemini free tier, as of late May 2026:
| Model | RPM (free) | RPD (free) | TPM (free) | Status |
|---|---|---|---|---|
| Gemini 2.5 Flash-Lite | 30 | 1,500 | 1,000,000 | Free |
| Gemini 2.5 Flash | 15 | 1,500 | 1,000,000 | Free |
| Gemini 2.5 Pro | 5 | 50 | 250,000 | Paid only since April 1 |
| Gemini 3 Pro / 3.1 Pro | - | - | - | Paid only (preview) |
Pro pricing on the paid tier is $2 per million input tokens, $10 per million output tokens for 3.1 Pro up to 200K context. Above 200K it steps to $4/$18.
The three replacements that actually work
The instinct when a free tier disappears is to find the next free thing. That instinct is wrong for production workloads, but it's the right instinct for hobby projects and prototypes. Here are the three replacements I've seen teams pick, in order from most-likely-to-work-for-you to least.
1. Just switch to Flash. Seriously.
This is the unsexy answer and it's the right answer for about 70% of use cases. Gemini 2.5 Flash and Flash-Lite are still free, they're still good models, and most of the workloads people were running on Pro didn't actually need Pro.
I keep seeing teams use Gemini Pro for tasks like "summarize this 2-paragraph email" or "extract the customer's intent from a 3-sentence message". Those tasks are Flash workloads. They were Flash workloads in 2025. Running them on Pro was never the right call, it just felt free so nobody audited it.
The Flash limits are 15 RPM and 1,500 RPD, with a million tokens per minute. Flash-Lite is 30 RPM. For most hobby projects and a meaningful fraction of indie SaaS, that's enough.
Where Flash falls down: complex reasoning chains, code generation longer than a few hundred lines, multi-step agent loops where each step needs to chain on the last one's output. Those genuinely need Pro-class capability.
2. Set up billing, eat the cost, instrument heavily
If you're already past prototype and you have real users hitting your Gemini calls, the cleanest answer is to enable billing and start treating it like the production line item it actually is.
The trap to avoid: many teams enable billing without instrumenting their per-customer cost first. Then the bill comes and they don't know which customer or which feature is driving 80% of the spend. Two weeks later they're rate-limiting everyone instead of fixing the one runaway use case.
The right move: before you enable billing, wire up per-customer token tracking. UsageBox does this with a thin wrapper around your model calls (or, if you'd rather wire your own, the Stripe Meter Events API is the long-route equivalent). When the bill arrives, you know exactly which 5% of customers are consuming 60% of the tokens, and you can decide whether to bill them, throttle them, or tier them up before the runaway becomes the new normal.
3. Multi-model routing
The advanced move: route easy requests to Flash (free), route hard requests to Pro (paid), and use a small classifier to decide which is which.
This is real work to set up. The classifier has to be cheap enough that it doesn't eat the savings (a small Flash call is fine), it has to be accurate enough that you don't constantly send Flash workloads to Pro by mistake, and you need to A/B test that the quality on routed-to-Flash requests isn't dropping below your threshold.
The payoff: teams I know who've done this well are running about 80% of their volume on Flash and 20% on Pro, with no perceptible quality drop on the user-facing output. Their bill dropped 70% from the all-Pro baseline.
The downside: this is engineering effort. You'll spend a week or two getting the routing right, and you'll spend ongoing effort tuning the classifier as your traffic mix evolves. If you don't have a couple of senior engineers to put on it, skip this and pick option 1 or 2.
The mistake to avoid
The biggest mistake I've seen since April 1: teams hard-cutting from Pro to Flash without measuring what breaks. Their agents get worse, their customer satisfaction dips, they don't connect the two events, and three weeks later they're rolling back to Pro at full bill without understanding why.
If you're going to downgrade models, A/B test it. Run 10% of traffic on Flash for a week, compare your quality metric (user thumbs, downstream completion rate, whatever you measure), and only flip the rest of your traffic once you have proof Flash holds up.
What about the other big LLM providers?
The follow-up question I keep getting: should I just switch to OpenAI or Anthropic? The honest answer is no, not on price alone. Even with the Pro tier going paid, Gemini is still the cheapest major model API for most workloads. The $2/$10 per million 3.1 Pro pricing is below GPT-5 and Claude Opus 4.7 by a wide margin.
The reasons to switch aren't price, they're capability fit. Claude Opus 4.7 is meaningfully better at code generation and long-context reasoning. GPT-5 is meaningfully better at tool use and structured output. If you're paying for Gemini Pro and not using the things that make Pro better than Flash, switch to Flash. If you're paying for Gemini Pro because you need the capability, you're probably going to keep paying for it; the question is just which Pro model.
Related reading
If you're auditing your AI spend after this change, see the platform comparison at Stripe Billing vs Chargebee vs Recurly vs Metronome, the cost models in Claude Sonnet vs Kimi vs DeepSeek: Real Pricing, and why the Gemini API free tier disappears once you enable billing (with the multi-project workaround that survives audits).