On June 17, 2025, Google took the preview label off Gemini 2.5 Pro. The stable alias gemini-2.5-pro shipped that day, closing out more than three months of iteration since the experimental build debuted in late March. For production workloads, the message was simple: pin the stable name and stop chasing preview identifiers through your configs.
Developers already running the newest preview felt almost nothing. Google confirmed the stable build is identical to the June 5 release. The real changes landed elsewhere: Gemini 2.5 Flash went stable with a new price sheet, and a new Flash-Lite preview appeared aimed at high-volume workloads. Taken together, the three moves signaled that Google considered the 2.5 generation ready to underwrite.
Stable is the 06-05 build
The road to GA was unusually long. 2.5 Pro debuted experimentally as gemini-2.5-pro-exp-03-25 in March 2025, reached free users days later, picked up a coding-focused update in early May, and received its final preview on June 5 with adaptive thinking. The June 17 changelog entry simply reads: released gemini-2.5-pro, the stable version — in practice, the 06-05 build promoted.
Ars Technica quoted Google saying Gemini 2.5 is stable and ready for developers to build on. Nothing changed in the Gemini app, which had already been serving the final preview; the point of the day was to give API users a fixed target. Write gemini-2.5-pro and you get the 06-05 build, with no silent behavior swaps hiding behind the alias.
Flash repriced: input up, output down
Gemini 2.5 Flash went stable the same day, with the changelog releasing the first stable gemini-2.5-flash. Its changes were the most concrete. Per 9to5Google’s summary of the new pricing, input tokens rose from $0.15 to $0.30 per million, while output tokens fell from $3.50 to $2.50 per million.
The structural shift mattered more than the individual numbers. Google collapsed the separate thinking and non-thinking rates into one unified price and removed the old tiering by input size. Applications that generate a lot of output — long-form generation, agentic workflows — came out ahead. Pipelines that stuff large retrieved context into the prompt paid more, and teams that saved money by disabling thinking lost that lever entirely. Existing cost models needed rework, not just a rate check.
Flash-Lite preview: built for high volume
The new gemini-2.5-flash-lite-preview-06-17 has a narrow brief: translation, classification, and other high-throughput, latency-sensitive tasks. Google claimed lower latency than both 2.0 Flash-Lite and 2.0 Flash, better quality than 2.0 Flash-Lite, a 1 million-token context window, multimodal input, and native tools including Search grounding, code execution, URL context, and function calling. Thinking is off by default, with budgets available when a task needs it.
Price is the pitch. Ars Technica noted Flash-Lite costs about a third of Flash for text, image, and video inputs, and less than a sixth of Flash for output tokens. Ars also pointed out that every 2.5 model supports adjustable thinking budgets, giving cost control a concrete knob. For metered, high-volume services, those numbers redraw the cost curve outright.
App tiers and Search integration
Consumer access settled into three tiers the same day: limited free usage, 100 prompts per day for AI Pro subscribers, and the highest limits for AI Ultra. The model picker’s division of labor was equally clear — 2.5 Pro positioned for reasoning, math, and code, 2.5 Flash for fast, general-purpose help.
There was an internal angle as well. Ars reported that custom versions of Flash and Flash-Lite already powered AI Overviews and AI Mode in Search, with Google routing queries to different models based on complexity. Stabilizing the family locked in the foundation for that tiered routing, which is why the stability mattered inside Google too, not just for external developers.
What it meant for developers
Translated into engineering work, the announcements left three action items. First, production configs could move to the gemini-2.5-pro stable alias, simplifying configuration management. Second, Flash billing had changed structure, so services with different input-output mixes had to re-estimate costs rather than assume a uniform discount. Third, the changelog scheduled the 2.5 Flash 04-17 preview for deprecation on July 15, 2025 — anything still calling that preview needed a migration date on the calendar.
Looking back from 2026, June 17 reads as the line where the 2.5 generation shifted from constant iteration to dependable infrastructure: fixed version names, fixed prices, and a published deprecation schedule. For teams weighing model vendors mid-2025, that certainty mattered more than any benchmark delta — production bills and on-call maintenance were riding on it.
Sources
- Gemini API Changelog - Google
- Gemini 2.5 Pro launch - 9to5Google
- Google’s Gemini AI family updated with stable 2.5 Pro - Ars Technica
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.