Gemini 3.6 Flash

Canonical version: Gemini 3.6 Flash.

Gemini 3.6 Flash is Google's workhorse Gemini model, released on 21 July 2026 alongside Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber. Its pitch is not intelligence, it is cost per completed task: cheaper than the model it replaces, meaningfully better at coding and tool use, and 17% more token-efficient.

That combination matters more than it sounds. Most agentic spend is Flash-tier spend.

The three models

Gemini 3.6 Flash at $1.50 / $7.50 per million tokens, in and out. Better than 3.5 Flash on coding (DeepSWE 49% vs 37%), on ML research tasks (63.9% vs 49.7%), and on agentic tool use. Google's framing is precision: fewer unwanted code edits and fewer execution loops, which is exactly the failure mode that makes cheap models expensive in agent workloads.

Gemini 3.5 Flash-Lite at $0.30 / $2.50, running at about 350 output tokens per second per Artificial Analysis. It beats Gemini 3 Flash on some benchmarks including SWE-Bench Pro, which is a reasonable summary of how fast this tier is moving.

Gemini 3.5 Flash Cyber, a specialist for finding, validating, and patching security vulnerabilities at scale. Not generally available: a limited pilot for governments and trusted partners, delivered through CodeMender. Worth noting as a category. Purpose-built security models are becoming a distinct product line rather than a benchmark row.

Why the numbers deserve suspicion

The announcement quotes DeepSWE improvement as "up to 65%" in one paragraph and 49% versus 37% in another. Those are not the same claim, and nothing in the post reconciles them.

More telling: Google benchmarks these models only against previous Gemini versions. No comparison to Claude Fable 5, to GPT-5.6, to Kimi K3, to GLM. For a release whose entire argument is cost-effectiveness, the absence of any competitive cost comparison is the loudest thing in the document.

The strategic question underneath

Google spent its earnings call being asked, by four banks in a row, what it plans to do about not having a state-of-the-art model. Sundar Pichai's answer each time was some version of "everyone uses Flash anyway, including us internally, and Gemini 4 pre-training has started".

Both halves of that argument are defensible and neither is comfortable.

The case for Google. Flash is the model that actually runs the products. Cloud revenue was up 82% year over year. If the business is serving intelligence at volume inside an existing ecosystem, a cheap fast good-enough model is a better asset than a benchmark crown. Pareto frontier beats leaderboard position.

The case against. Google's last uncontested best-model moment was Gemini 3 in November 2025. Since then it has been passed not only by Anthropic and OpenAI but by open-weight models. When a competent team can self-host something that matches your flagship, "we sell intelligence as a service" needs a second sentence, and compliance is usually that sentence.

The reason the frontier still matters commercially. It is the proof of capability that lets you sell everything below it. That is closer to why Wall Street keeps asking than any technical argument about model quality.

Caveats

• Knowledge cutoff was listed as March 2026 at launch, then changed to "unknown". Testers reported the model unaware of well-documented 2025 events. Search grounding compensates less than you would hope, because a model with stale priors searches with stale keywords
• 3.6 Flash is more expensive than GLM 5.2 and arguably weaker, and the announcement offers nothing to argue otherwise
• "17% more token-efficient" is the Artificial Analysis index figure. The "up to 65%" is a single favorable benchmark. Use the first number
• Flash Cyber is not something you can evaluate. Treat it as a signal about where security tooling is heading, not as an available option

References

• Official announcement — https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/
• Model Garden listing — https://console.cloud.google.com/agent-platform/publishers/google/model-garden/gemini-3.6-flash
• Hacker News discussion (760 points) — https://news.ycombinator.com/item?id=48993414
• Intelligence vs cost comparison — https://artificialanalysis.ai/#intelligence-comparison-tabs

Related

• Gemini
• Google
• Google DeepMind
• Large Language Models (LLMs)
• AI Frontier Model
• AI Foundation Models
• AI Agents
• AI Tool Use
• Claude Fable 5
• GPT-5.6
• Kimi K3
• Qwen 3.8
• Artificial Analysis
• SWE-Bench
• OpenAI
• Anthropic

About Sébastien

Ready to get to the next level?