Skip to main content
Gemini 3.6 Flash (gemini-3.6-flash) is Google’s multimodal text model updated in July 2026 (stable release), accepting text/image/video/audio/PDF input with a 1M context window and 64K output. APIYI has completed a full dual-endpoint test pass (26+5 cases): both the native Gemini format and the OpenAI-compatible format work out of the box, and native tools — Search grounding, code execution, URL context — are verified working.
Available on APIYI now: model name gemini-3.6-flash, in the default / svip groups. Thinking is ON by default (thinking tokens bill as output) — lower the thinking tier for latency- or cost-sensitive workloads (see “Controlling thinking” below). For lightweight workloads, consider its cheaper sibling Gemini 3.5 Flash-Lite.

Highlights

Full native tool suite

Google Search grounding, Maps grounding, URL context, code execution, and Computer Use (Preview) all verified working on the native endpoint — no Google API Key required.

Full multimodal understanding

Image, PDF, and audio inputs verified accurate (video shares the same pipeline). The 1M context fits an entire book or codebase.

Four thinking tiers

thinkingLevel minimal/low/medium/high measured at 0/403/487/837 thinking tokens — monotonic, so you can budget thinking precisely per task.

Two endpoints, zero friction

Native Gemini format (official SDK, just change base_url) and OpenAI-compatible format, both live at official pricing.

Model details

Verified capability matrix

APIYI test results from July 22, 2026 (official claims vs. measured behavior):
Code execution note: our tests confirm code really runs upstream (a non-memorizable sha256 task returned the correct digest), but the executableCode / codeExecutionResult parts are currently not echoed in the response — the code and its result appear in the text body instead. Apps that render those two fields separately should take note.

Pricing

Pricing note: thinking tokens bill as output — the main reason to manage thinking tiers. APIYI matches official pricing; the discount comes from top-up bonuses: +10% on $100, up to +20% (≈17% off). See top-up promotions.

Controlling thinking

Thinking is ON by default: even “1+1” produces ~200 thinking tokens first. Measured tiers:
Pass thinkingConfig: {"includeThoughts": true} to get thought parts back (marked thought: true); spend shows in usageMetadata.thoughtsTokenCount. On the OpenAI-compatible endpoint use reasoning_effort (low/medium/high) and read usage.completion_tokens_details.reasoning_tokens.

Quick start

OpenAI-compatible format (drop-in for existing code)

FAQ

No. Put your APIYI token (the sk- key) in the x-goog-api-key header. With the official google-genai SDK, just set base_url to https://api.apiyi.com.
No. google_search, url_context, codeExecution, Maps grounding, and Computer Use are native-format only. The OpenAI-compatible endpoint covers the standard set: chat, streaming, function calling, JSON Schema, and vision.
A repeated long prefix may hit on the second request (measured: 8,176 of a 15.9K-token prefix), but hits are probabilistic — don’t build cost models on them. The explicit cache API (cachedContents) is not yet enabled on the platform.
Pick 3.6 Flash for tools, deep thinking, and stronger reasoning. Pick 3.5 Flash-Lite ($0.30 in / $2.50 out, no thinking by default, roughly twice as fast) for high-frequency, latency- and cost-sensitive workloads.