Google released Gemini 3.6 Flash on July 21, 2026. It is the successor to Gemini 3.5 Flash and the new default workhorse model in the Gemini family. Same 1M context window, same multimodal input, but faster, cheaper per output token, and better at coding and agentic tasks.
The headline numbers: 304 tokens per second output speed, $7.50 per 1M output tokens (down from $9.00), and 17% fewer output tokens needed to accomplish the same tasks. Computer use is now a built-in tool. The knowledge cutoff jumps from January 2025 to March 2026.
Here is everything you need to know: specs, benchmarks, pricing, API access, and where it fits against Claude Sonnet 5 and GPT-5.6.
Key Specs
| Spec | Value |
|---|---|
| Release date | July 21, 2026 |
| Model family | Gemini 3.6 |
| Input modalities | Text, Image, Video, Audio, PDF |
| Output modalities | Text only |
| Context window | 1,000,000 tokens |
| Max output tokens | 65,536 |
| Knowledge cutoff | March 2026 |
| Output speed | 304 tokens/sec |
| Input pricing (global) | $1.50 per 1M tokens |
| Output pricing (global) | $7.50 per 1M tokens |
| Cached input pricing | $0.15 per 1M tokens |
| Reasoning/thinking | Yes |
| Computer use | Built-in (new) |
| Tool use | Function calling, Structured output, Search as a tool, Code execution |
The 1M context window and 65K max output carry over from 3.5 Flash. No changes there. The improvements are in efficiency, speed, and capability.
What Changed From 3.5 Flash
3.6 Flash is not a ground-up redesign. It builds on 3.5 Flash with targeted improvements based on developer feedback. Here is what is different:
Cheaper output tokens. Output price dropped from $9.00 to $7.50 per 1M tokens. Input stays at $1.50. For agent workflows that generate a lot of output, this is a meaningful cost reduction.
Fewer tokens per task. Google reports 3.6 Flash uses 17% fewer output tokens than 3.5 Flash to accomplish the same tasks. Combined with the lower per-token price, the effective cost per task drops significantly.
Computer use. 3.5 Flash had no computer use capability. 3.6 Flash ships with built-in computer use via the Gemini API and Gemini Enterprise. This is a big deal for agent builders who need browser or desktop automation.
Knowledge cutoff. January 2025 to March 2026. The model now knows about events through early 2026 without needing Search as a tool.
Speed. 304 tok/s vs 289 tok/s. Marginal improvement, but 3.6 Flash was already the fastest frontier-tier model. It is now faster.
Benchmarks
Google’s official numbers comparing 3.6 Flash to 3.5 Flash:
| Benchmark | 3.6 Flash | 3.5 Flash | Change |
|---|---|---|---|
| DeepSWE (Datacurve) | 49% | 37% | +12 points |
| MLE Bench | 63.9% | 49.7% | +14.2 points |
| OSWorld-Verified | 83.0% | 78.4% | +4.6 points |
| GDPval-AA v2 | 1421 | 1349 | +72 Elo |
The DeepSWE and MLE Bench jumps are significant. DeepSWE measures real-world software engineering tasks, and a 12-point improvement means 3.6 Flash generates fewer unwanted code edits and needs fewer execution loops. MLE Bench tests machine learning research tasks, where the 14-point jump puts 3.6 Flash well ahead of where 3.5 Flash was.
OSWorld-Verified tests computer use capability, and the 4.6-point improvement aligns with the new built-in computer use feature.
Google also claims 3.6 Flash is more token-efficient on agentic workflows, taking fewer reasoning steps and tool calls to accomplish multi-step tasks. This matches the 17% token reduction claim.
Pricing Comparison
| Model | Input (per 1M) | Output (per 1M) | Speed |
|---|---|---|---|
| Gemini 3.6 Flash | $1.50 | $7.50 | 304 tok/s |
| Gemini 3.5 Flash | $1.50 | $9.00 | 289 tok/s |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | 350 tok/s |
| Claude Sonnet 5 | $2.00 | $10.00 | ~180 tok/s |
| GPT-5.6 Sol | $5.00 | $30.00 | ~150 tok/s |
| DeepSeek V4 Pro | $2.19 | $8.76 | ~120 tok/s |
3.6 Flash is competitive on output price. At $7.50/1M output, it undercuts Claude Sonnet 5 ($10.00) and DeepSeek V4 Pro ($8.76). GPT-5.6 Luna ($6.00) is cheaper on output, but Gemini is 3x faster. Combined with the 304 tok/s speed, the cost-per-quality ratio is hard to beat.
For high-volume use cases where even $7.50/1M is too much, Gemini 3.5 Flash-Lite launched at $0.30/$2.50 with 350 tok/s. It is less capable but significantly cheaper.
Cached input at $0.15/1M remains a 90% discount. For agent loops that re-send system prompts and tool definitions, caching is essential.
API Access and Setup
3.6 Flash is available on:
- Gemini API via Google AI Studio
- Vertex AI for enterprise
- Gemini App (consumer)
- Antigravity CLI
- Android Studio
- GitHub Copilot (announced today)
- OpenRouter (expected soon)
Quick API Example
import google.genai as genai
client = genai.Client(api_key="YOUR_API_KEY")
response = client.models.generate_content(
model="gemini-3.6-flash",
contents="Write a Python function that implements a thread-safe LRU cache.",
config={
"temperature": 0.7,
"max_output_tokens": 1024,
}
)
print(response.text)With Thinking Mode
response = client.models.generate_content(
model="gemini-3.6-flash",
contents="Find the bug in this code and explain the fix.",
config={
"thinking_config": {"thinking_budget": 8192}
}
)
for part in response.candidates[0].content.parts:
if part.thought:
print("Thinking:", part.text)
else:
print("Response:", part.text)For a full setup walkthrough including streaming, function calling, and structured output, see our Gemini 3.5 Flash API Setup Guide (the setup process is identical, just change the model name).
Computer Use
This is new in 3.6 Flash. Computer use lets the model interact with a browser or desktop environment through screenshots and mouse/keyboard actions. It is available as a built-in client-side tool via the Gemini API.
Google’s benchmark numbers show 83.0% on OSWorld-Verified, up from 78.4% on 3.5 Flash (which did not have computer use as a built-in tool). This puts 3.6 Flash in competitive territory with Claude Opus for browser automation tasks.
For agent builders, computer use opens up workflows that were previously impossible with Gemini models: filling out forms, navigating web apps, interacting with legacy systems that only have a GUI.
Also Released: 3.5 Flash-Lite and 3.5 Flash Cyber
Google shipped two additional models alongside 3.6 Flash:
Gemini 3.5 Flash-Lite is the fastest 3.5-class model at 350 tok/s, priced at $0.30/$2.50. It is designed for high-volume, low-latency tasks like agentic search and document processing. It outperforms the older 3.1 Flash-Lite significantly and even beats Gemini 3 Flash on some agentic benchmarks. See our Gemini 3.5 Flash-Lite complete guide for details.
Gemini 3.5 Flash Cyber is a cybersecurity-specialized model fine-tuned for finding and fixing security vulnerabilities. It is available exclusively through Google’s CodeMender agent for governments and trusted partners. Not publicly available.
What This Means for Developers
If you are using 3.5 Flash: Switch to 3.6 Flash. The API is identical (just change the model name), it is cheaper on output, faster, and better at coding. There is no reason to stay on 3.5 Flash unless you have specific compatibility constraints.
If you are choosing between models: 3.6 Flash is now the best price-performance option for coding and agentic workloads. It undercuts Claude Sonnet 5 and GPT-5.6 on price while matching or beating them on most benchmarks. The 304 tok/s speed is 2-4x faster than competitors.
If you are building agents: The built-in computer use and 83% OSWorld score make 3.6 Flash viable for browser automation. Combined with the 1M context window and MCP Atlas-leading tool use, this is a strong agent backbone.
If cost is the primary concern: Use 3.5 Flash-Lite ($0.30/$2.50) for high-volume tasks and 3.6 Flash for complex ones. The two-tier approach gives you the best cost-quality tradeoff.
My Take
Gemini 3.6 Flash is the best value in the Flash lineup right now. At $1.50/$7.50 with 304 tok/s and a 1M context window, it undercuts Claude Sonnet 5 and GPT-5.6 on price while matching or beating them on most benchmarks.
The built-in computer use is the real story here. Google went from “no computer use” to “83% OSWorld” in one release. That is not incremental, that is a new capability. If you are building agents that need to interact with web applications, this changes what is possible with Gemini.
The 17% token efficiency improvement is the kind of thing that does not sound exciting but saves real money at scale. Fewer tokens per task means lower cost AND lower latency. It is a win on both dimensions.
My one concern: Google has not published Terminal-Bench or SWE-bench scores specifically for 3.6 Flash. The benchmarks they did publish (DeepSWE, MLE Bench, OSWorld, GDPval-AA) are strong, but the absence of the standard coding benchmarks is notable. I would like to see those numbers before declaring 3.6 Flash the coding champion.
For now, 3.6 Flash is the default choice for cost-sensitive, speed-sensitive workloads. If you need the absolute best coding quality, Claude Sonnet 5 or GPT-5.6 Luna may still have an edge. But for most developers, 3.6 Flash is the right call.
Coming Soon
Google confirmed two things in the 3.6 Flash announcement:
Gemini 3.5 Pro is testing with partners and will be broadly available soon. This will be the full frontier model in the 3.5 family, expected to compete with Claude Opus and GPT-5.6 on all benchmarks.
Gemini 4 pre-training has started. Google called it their “most ambitious pre-training run yet.” No timeline given.
Limitations
- Text-only output. Input is multimodal (text, image, video, audio, PDF), but output is text only. No image generation, no audio synthesis.
- No open weights. 3.6 Flash is proprietary. You cannot download or self-host it.
- Computer use is new. The 83% OSWorld score is good but not leading. Claude Opus has more mature computer use tooling.
- Long context retrieval. No specific MRCR v2 benchmark published for 3.6 Flash yet. 3.5 Flash scored 77.3% on MRCR v2 (128k), which trailed Gemini 3.1 Pro (84.9%). If needle-in-haystack retrieval over very long documents is your primary use case, this may still be a limitation.
FAQ
Is Gemini 3.6 Flash free?
Not directly. The free tier still uses older Gemini models. However, 3.6 Flash is included in Google AI Plus ($20/mo) and Pro ($20/mo) subscriptions. API usage is pay-per-token at $1.50/M input and $7.50/M output. For free alternatives, see our best free AI APIs guide.
How much faster is 3.6 Flash than 3.5 Flash?
Marginally faster on raw speed: 304 tok/s vs 289 tok/s. The bigger improvement is efficiency: 3.6 Flash uses 17% fewer output tokens to accomplish the same tasks, so real-world latency per task is lower.
Should I switch from 3.5 Flash to 3.6 Flash?
Yes. The API is identical (change the model name from gemini-3.5-flash to gemini-3.6-flash), it is cheaper on output tokens, faster, and better at coding and agentic tasks. There is no downside to switching unless you have specific compatibility constraints.
How does 3.6 Flash compare to Claude Sonnet 5?
3.6 Flash is cheaper ($1.50/$7.50 vs $2.00/$10.00) and faster (304 tok/s vs ~180 tok/s). Claude Sonnet 5 may still lead on certain coding benchmarks and has more mature tooling for some workflows. For a full breakdown, see our Gemini 3.6 Flash vs Claude Sonnet 5 comparison.
Does 3.6 Flash support computer use?
Yes. Computer use is now a built-in client-side tool via the Gemini API and Gemini Enterprise. The model scores 83.0% on OSWorld-Verified, up from 78.4% on 3.5 Flash.
What is the context window for 3.6 Flash?
1 million tokens input, 65,536 tokens max output. Same as 3.5 Flash.