Summary: Google shipped Gemini 3.1 Pro in preview on February 19, 2026, pitching doubled reasoning compared with Gemini 3 Pro at unchanged pricing through the gemini-3.1-pro-preview endpoint. The article says preview endpoints suit evaluation but not production commitments, because behavior and timelines can shift. It recommends shadow testing on real traffic and waiting for a stable release before migration.
On February 19, 2026, Google shipped Gemini 3.1 Pro as a preview. The pitch compresses into one line: reasoning doubles against Gemini 3 Pro, while pricing holds where it was. The endpoint name is just as plainspoken: gemini-3.1-pro-preview.
Same Price, Twice the Reasoning
Announcing doubled reasoning at an unchanged price is textbook price-performance positioning. With the gaps between frontier models increasingly hard to detect by feel, Google has chosen to speak in capability-per-dollar — and for teams whose API call volume runs high, that yardstick is more practical than any single benchmark.
The move also revives a cloud-era classic: push price-performance somewhere rivals cannot profitably follow, and let scale absorb the cost difference. The model market has inherited that strategy, and Google sits among the players best equipped to run it — infrastructure scale is the ground it plays best on.
For buyers, the arithmetic lands squarely in unit economics. If reasoning quality per dollar doubles at an unchanged list price, workloads that used to hover at break-even — long chains of thought, evaluation at high volume, exhaustive code review — move into affordable territory without a single negotiation. Price cuts grab headlines; capability-at-price decides what teams choose to build.
Reading the -preview Suffix
For anyone building against the API, the -preview suffix on the endpoint name is a deliberate signal:
- Right for evaluation and trial runs — send the new model through your own eval pipeline and quantify its effect on real workloads
- Wrong for production commitments, since preview behavior and timelines can still shift
- Production traffic belongs on stable endpoints, with migration handled as a proper project
Version lifecycle management has become a foundational skill for building on LLMs, and Google marking that boundary through naming alone ranks among the friendliest gestures a provider can make.
The disciplined adoption pattern: run the preview endpoint in shadow alongside your current model on real traffic samples, score both with your own rubric rather than public benchmarks, and time the production switch to a stable release. In effect, preview access is intelligence collected for free — the only price is the eval harness you were supposed to have anyway.
February’s Release Calendar, in Context
Set Gemini 3.1 Pro against the rest of the month: Claude Opus 4.6 and GPT-5.3-Codex landed on the same day, February 5; GLM-5 opened its weights on February 11; Claude Sonnet 4.6 took over the default slot on February 17. Four major camps showed their cards within two weeks. The pace compresses evaluation cycles as well: a team that spends three weeks benchmarking a release now files its verdict after the next release has already arrived.
Against a cadence that has blurred together, Google’s clearest differentiator is pricing backed by infrastructure scale. Double-at-the-same-price is not a limited-time promotion — it relocates the competitive axis from “who is strongest” to “who is strongest per dollar,” and that fight has barely begun. It also lifts the bar for rival pricing pages: once the market’s reference point is doubling at flat cost, a competitor shipping a modest improvement with a price increase has to win two arguments — capability and cost.