Gemini 3.7 Flash launched with a time-limited introductory price. A useful budget separates that period from the scheduled rate, and a current article also acknowledges the newer 3.8 release.
Keep the version history clear
Google introduced 3.7 Flash on August 13. It attributes improvements to developer feedback and algorithmic work. The short release interval does not prove a particular training history, so the earlier assertion about what could not have been pretrained is removed.
Google then announced 3.8 Flash on September 2. This page preserves the 3.7 event rather than silently rewriting it as a 3.8 launch. A newer model is a reason to evaluate, not evidence that an existing integration has already changed.
Two dates in the price table
The current Gemini API tariff lists standard 3.7 input at $0.75 and output, including thinking, at $3.75 per million tokens through December 31. The scheduled January 1 rates are $1.50 and $7.50. Cache storage, cached input, grounding and service tiers have separate lines.
For an illustrative workload with two million uncached input tokens and half a million total billed output tokens, the current token charge is $3.375. At the scheduled standard rate, it is $6.75. That example excludes caching and tool charges and assumes no change in usage. Future actual billing still depends on the tariff then in force.
Count the work the API bills
Half a million visible answer tokens is not necessarily half a million billed output tokens if thinking is additional. Record the response's usage breakdown rather than estimating solely from displayed text. Repeated attempts and longer reasoning can outweigh a small price difference between models.
The 3.7 model documentation also distinguishes input and output limits. A large input window does not guarantee a single response can reproduce the entire document. Structure a task around the intended output and inspect truncation or finish conditions.
For a migration to 3.8, keep the same task set, tool permissions and success criteria. Compare completed-task cost, latency and errors, not just benchmark rank or headline token price. Record the model identifier and review any alias behavior; a version-looking name alone is not proof that every serving detail is immutable.
Preserve August 3.7 launch while adding September 3.8; verify introductory and scheduled rates, count thinking and remove unsupported pretraining claims.