Gemini 3.6 Flash

What is Gemini 3.6 Flash?#

Gemini 3.6 Flash is Google's July 21, 2026 Flash release, the generation before Gemini 3.7 Flash, built for computer use, long-context retrieval, and everyday agentic work. It takes text, image, audio, video, and PDF input, outputs up to 65k tokens of text, and carries a 1M-token context window. It runs on Unifically as google/gemini-3.6-flash, with thinking on a tunable level and output measured at 234.6 tokens per second.

Key features of Gemini 3.6 Flash#

The strongest computer-use score in its table

83.0% on OSWorld-Verified in Google's July 2026 evaluation, ahead of Claude Sonnet 5 at 81.2%, Gemini 3.5 Flash at 78.4%, and GPT 5.6 Luna at 72.6%. Desktop automation agents are where this model earns its keep.

Long-context retrieval that holds up

91.8% on GDM-MRCR v2 at 128k tokens with eight needles, well ahead of Grok 4.5 at 81.4% and Claude Sonnet 5 at 71.6%. On the 1M pointwise variant it scores 54.0%, roughly double Gemini 3.5 Flash.

Fast output, solid independent score

Artificial Analysis measures 234.6 output tokens per second, the second-fastest of 188 tracked models, with an Intelligence Index of 52, above Gemini 3.1 Pro Preview at 48 while running about twice as fast.

Verified coding and security patching

77.45% on Vals AI's independent SWE-bench Verified re-run, plus first place on the CyberBench Patch track at 84.75%. It also cut output tokens about 17% against Gemini 3.5 Flash on the same work.

Best for#

Use cases#

Use Gemini 3.6 Flash for computer-use agents that fill forms, drive internal tools, and run browser workflows; its OSWorld-Verified score leads the models around it. It fits retrieval pipelines over large document sets, where 128k-deep recall keeps citations honest. Data teams can run it for machine-learning workflow automation and bulk extraction at 234 tokens per second. It also works as the stable Flash choice for pipelines that were tuned before Gemini 3.7 Flash arrived and should not change behavior mid-quarter.

Limitations#

Time to first token is high at high thinking, near 17 seconds in Artificial Analysis measurements, even though output speed after that is near the top of the board.

It is not a frontier coding model: Google's own table puts it well behind GPT 5.6 Luna on DeepSWE v1.1 (49% against 67%) and behind Claude Sonnet 5 on SWE-Bench Pro.

Gemini 3.7 Flash replaced it as the newest Flash three weeks after release, so it no longer gets the "latest" label. Output is text only.

Gemini 3.6 Flash vs Gemini 3.5 Flash#

The 3.6 release was a big step over 3.5 Flash: DeepSWE v1.1 rose from 37% to 49%, MLE-Bench from 49.7% to 63.9%, OSWorld-Verified from 78.4% to 83.0%, and 1M-token MRCR retrieval roughly doubled from 26.6% to 54.0%. It also finishes the same work in about 17% fewer output tokens. If anything still runs on Gemini 3.5 Flash, moving up is a clear win.

When to use Gemini 3.6 Flash#

Use Gemini 3.6 Flash when the job is computer use, long-context retrieval, or high-volume multimodal processing and you want known, regression-tested behavior. For new builds that just want the strongest Flash model, start with Gemini 3.7 Flash instead and keep 3.6 Flash as the fallback your pipeline already understands.

API examples#

Call Gemini 3.6 Flash from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/google/gemini-3.6-flash.

curl -X POST https://api.unifically.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "google/gemini-3.6-flash",
    "messages": [
      { "role": "user", "content": "Extract every invoice line item from this PDF into JSON with amounts and dates." }
    ]
  }'

The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.