RuntimeWire is testing Google's callable model against the Fable 5.1 model, while Google's public catalog still stops at Gemini 3.7 Flash.
By Ryan Merket · Published
Primary source: X - @LuminaBench
Why it matters #
Gemini 3.8 Flash is entering Google's distribution layer only 20 days after 3.7. Its published limits appear unchanged, putting the focus on quality, latency and price once Google documents the release.
Google's Gemini 3.8 Flash is live and callable through the Gemini API, with RuntimeWire currently testing the model head to head against the Fable 5.1 model. Results will follow soon. Google's public documentation, however, has yet to acknowledge the release.
Lumina (@LuminaBench), an account that tracks unreleased AI models, posted a screenshot on September 2nd showing gemini-3.8-flash
in account-scoped metadata returned by Google's API. The entry listed a 1,048,576-token context window, a 65,536-token output limit and support for methods including generateContent
, batchGenerateContent
, countTokens
and createCachedContent
.
RuntimeWire's access moves the story beyond a metadata sighting: Gemini 3.8 Flash can be called and evaluated through Google's API. It does not establish general availability, production stability or access for every developer.
Google's public Gemini API model catalog still identified Gemini 3.7 Flash as its "latest and most capable Flash model" as of reporting. The catalog, last updated on August 27th, listed endpoints through gemini-3.7-flash
and contained no entry for Gemini 3.8 Flash.
The Gemini API release notes were similarly silent. Their latest entry, dated September 1st, covered agentic video understanding for existing Flash models. Google's most recent documented general-purpose Flash launch remained Gemini 3.7 Flash on August 13th.
The gap indicates a staged deployment: Google has made the model callable for at least some users before completing a broad API rollout, publishing pricing or issuing a model card. Developers without confirmed access should not assume that gemini-3.8-flash
is ready to replace a production endpoint.
A separate listing from LLM Gateway labeled Gemini 3.8 Flash as stable and available through Google Vertex AI on September 2nd. LLM Gateway advertised pricing starting at $0.75 per million input tokens and $3.75 per million output tokens. Google had not placed those terms in its own public Gemini API catalog or release notes at the time of reporting, leaving the price and broader distribution status without direct confirmation from Google.
Another early assessment was considerably less restrained. The @synthwavedd account wrote:
"Google have, dare I say... RELEASED A GOOD MODEL"
The account claimed Gemini 3.8 Flash offered roughly Claude Opus 5 performance at much lower cost and high speed. The comparison table supplied with the post listed the same $0.75 input and $3.75 output prices shown by LLM Gateway. Its benchmark results were mixed: Gemini 3.8 Flash led Claude Opus 5 on several listed evaluations and trailed it on others, making the broader performance claim an early third-party assessment rather than a conclusion supported by Google documentation.
The token limits have not moved
The specifications visible in Lumina's screenshot match the limits Google already publishes for Gemini 3.7 Flash: 1,048,576 input tokens and 65,536 output tokens. The metadata therefore provides no evidence that Gemini 3.8 Flash expands the amount of material developers can place into a prompt or receive in one response.
The consequential changes, if any, will have to come from model quality, latency, tool use, reliability or cost. RuntimeWire's head-to-head evaluation against the Fable 5.1 model is intended to test that performance directly. Google has not yet published a model card, benchmark results, supported input modalities, regional availability, rate limits or migration guidance for Gemini 3.8 Flash.
Google's release cadence explains why the endpoint is surfacing so quickly. Gemini 3.6 Flash reached general availability on July 21st, followed by Gemini 3.7 Flash on August 13th. Gemini 3.8 Flash became callable 20 days later, continuing a near-monthly cycle for the lower-cost Flash line.
That schedule turns model selection into an operational problem for teams building on Gemini. Google describes stable model IDs as versions that usually do not change, while aliases such as gemini-flash-latest
can be switched to newer releases. A new numbered endpoint gives developers a migration boundary, provided Google publishes the pricing, evaluation data and deprecation schedule needed to test it safely.
For now, gemini-3.8-flash
has crossed the threshold from API metadata to a working model. RuntimeWire is using it now and will publish comparative results soon; Google still needs to document what changed and specify who can call it.