Skip to main content
Gemini 3.8 Flash (gemini-3.8-flash) is the multimodal text model Google put out on September 2, 2026, taking text, image, video and audio input. APIYI has both the Gemini native and OpenAI-compatible endpoints open, and ran 150 paired test cases across both protocols before launch, using gemini-3.7-flash as the reference.
Gemini 3.8 Flash is live on APIYI: model name gemini-3.8-flash, available in the default and svip groups. Deep thinking is on by default and thinking tokens bill as output, so lower the thinking tier or turn it off for latency- and cost-sensitive paths (see “Thinking control” below).
Official specs are not published yet: as of this page, neither Google’s Gemini API model list nor the DeepMind model card carries 3.8 Flash, and no launch blog post is up. So context window, max output, knowledge cutoff, and official benchmark scores are all marked “not published” here — we neither relay nor guess at them. Everything marked as measured comes from APIYI’s own testing; official material will be folded in once it lands.

Why it stands out

Half the price of 3.6 Flash

$0.75 in / $3.75 out per 1M tokens — half of 3.6 Flash ($1.50 / $7.50), and line-for-line identical to 3.7 Flash, so migrating from 3.7 costs nothing.

150 pre-launch test cases

Fired simultaneously against 3.7 Flash across both protocols: core capability, reasoning, parallel tool calling and image/video understanding all line up, with no regression unique to 3.8.

Both endpoints, no friction

Gemini native format (official SDK, no Google API key needed) and OpenAI-compatible format (just change base_url) are both open.

Near-zero migration cost

From 3.7 Flash it is a one-line model name change: request shape, parameters and response fields are unchanged, with one field-set difference and zero type changes measured.

Model information

Measured capability matrix

Results from APIYI testing on September 2, 2026 — 150 case logs and 198 calls, with every case fired simultaneously against gemini-3.7-flash to rule out time-of-day effects:
Search grounding is unconfirmed: passing tools: [{"googleSearch": {}}] returns 200 with a correct answer, but the response carries no groundingMetadata, meaning the answer came from the model’s own knowledge rather than a live search. gemini-3.7-flash behaved identically in the same run, which points at route-level enablement rather than a model capability difference. If you depend on live retrieval, verify on a slice of traffic first rather than designing around grounding being active.

Pricing

Line-for-line identical to gemini-3.7-flash, so migrating from 3.7 changes nothing about your costs; against 3.6 Flash ($1.50 / $7.50) it is a straight halving.
On pricing: thinking tokens bill as output — that is the most direct reason to manage the thinking tier. Google has not published official pricing for 3.8 Flash; for reference, the $0.75 / $3.75 currently in effect for 3.7 Flash is Google’s own limited-time promotional rate, stated as valid through December 31, 2026. APIYI prices match the provider line for line, with discounts coming through top-up bonuses — see top-up promotions.

Thinking control

Deep thinking is on by default: even a “1+1” prompt burns a few hundred thinking tokens first. Measured across the three tiers (same question, native endpoint):
The minimal tier is gone (3.6 Flash had it, 3.8 does not): the native endpoint returns 400 Thinking level is unsupported: THINKING_LEVEL_MINIMAL for thinkingLevel: "minimal". Use thinkingConfig: {"thinkingBudget": 0} to turn thinking off entirely.The OpenAI-compatible endpoint needs more care: reasoning_effort: "minimal" does not error — it returns 200, yet measurably still burned 76 thinking tokens that billed as output. The parameter is silently ignored and thinking happens anyway. Code carried over from 3.6 Flash with minimal still set will get no error telling you to change it.
To inspect the reasoning, pass thinkingConfig: {"includeThoughts": true} — the response then carries thought parts flagged thought: true, with usage in usageMetadata.thoughtsTokenCount. On the OpenAI-compatible endpoint, use reasoning_effort (low / medium / high) and read usage.completion_tokens_details.reasoning_tokens.

Examples

OpenAI-compatible format (no changes to existing code)

Migration guide

Change the model name, nothing else. Request shape, parameters and response fields tested unchanged; response field sets differ in one group with zero type changes, so clients need no adaptation. Pricing is identical, so your existing budget carries over.
The price is cut in half (input $1.50 → $0.75, output $7.50 → $3.75), but one thing must change: thinkingLevel: "minimal" is no longer supported and the native endpoint returns 400 — switch to thinkingConfig: {"thinkingBudget": 0}. On the OpenAI-compatible endpoint, reasoning_effort: "minimal" will not error but is silently ignored while thinking still bills, so fix that too.
Not published by Google, and we will not infer it from 3.7’s number. A long-context lookup task with a 14.5K-character prefix passed in testing. If you lean heavily on long context, ramp on a slice of traffic first, confirm the boundary, then cut over. This page will be updated once official specs land.
No. Put your APIYI key (starting with sk-) straight into the x-goog-api-key header; with the official google-genai SDK just point base_url at https://api.apiyi.com.
Code execution, URL context, safetySettings, and strict handling of stopSequences and seed are available only on the Gemini native endpoint. The OpenAI-compatible endpoint covers the usual ground (chat / streaming / function calling / JSON Schema / vision), but stop and seed tested as having no effect there.