What Google actually shipped
Three models went out on the same day. Gemini 3.6 Flash is the one most people will use, described by Google as its workhorse tier. Gemini 3.5 Flash-Lite targets high-throughput work like agentic search and document processing at $0.30 per million input tokens and $2.50 per million output. The third, Gemini 3.5 Flash Cyber, is tuned for finding and patching security vulnerabilities and is not generally available: access is limited to governments and trusted partners under a pilot programme, and it powers Google's CodeMender tool.
Flash models have always been the cheap, fast tier rather than the clever one. That framing matters when reading the benchmark jumps below, because the comparison is against last quarter's Flash, not against a frontier model.
The numbers Google published
Every figure in this table comes from Google's own launch materials. None of it has been independently verified, and we have not run our own testing yet.
| Benchmark | Gemini 3.5 Flash | Gemini 3.6 Flash |
|---|---|---|
| DeepSWE (production-ready code) | 37% | 49% |
| MLE Bench (ML research) | 49.7% | 63.9% |
| OSWorld-Verified (computer use) | 78.4% | 83.0% |
| GDPval-AA (knowledge work) | 1349 | 1421 |
| Output tokens used | baseline | 17% fewer |
| Knowledge cutoff | January 2025 | March 2026 |
The DeepSWE and MLE Bench jumps are the eye-catching ones, but the knowledge cutoff is probably the change you will feel first. Moving from January 2025 to March 2026 closes a fourteen-month gap. If you have been fighting a Flash model that had never heard of a library released last summer, that stops now.
The token efficiency claim is measured against the Artificial Analysis Index rather than an internal metric, which is worth something. Google frames it as the model taking fewer reasoning steps and fewer tool calls to finish multi-step work, so the saving is not just terser prose.
What the pricing change is worth
Input stays at $1.50 per million tokens. Output drops from $9.00 to $7.50 per million, a cut of about 17%. Stack that on top of generating 17% fewer output tokens for the same task and an output-heavy workload gets meaningfully cheaper, not marginally.
Two caveats before you budget around it. The token reduction is an average across a benchmark suite, so your own prompts may not track it. And output pricing only dominates your bill if you are actually output-heavy; a retrieval workload that stuffs long documents into context and gets short answers back is paying mostly for input, where nothing changed.
The Pro model that did not arrive
Google's Pro tier was last updated in February 2026. At I/O in May, Google said the Pro version was already in internal use and pointed at a rollout the following month. That did not happen. On July 16, Bloomberg reported the launch had slipped because the model was not meeting internal performance targets, and this release went out five days later with no Pro in it.
Google DeepMind product lead Logan Kilpatrick said Pro is testing with partners and that the team hopes it will "land soon." He also said DeepMind has begun its most ambitious pre-training run so far, for Gemini 4.
Read that how you like. Shipping three efficient Flash variants while the flagship slips is a reasonable thing to do if your customers are building agents at scale and care about cost per task. It is a harder look if you are comparing release cadences with the other labs.
How it sits against the competition
The gap Google is trying to close is visible in the release calendar. Since February, OpenAI has shipped GPT-5.5 and started rolling out the GPT-5.6 family, Anthropic has released Opus 4.8 and Sonnet 5, and it opened access to Fable 5. Chinese open-weight labs have kept a similar pace.
For a like-for-like sense of where 3.6 Flash lands, these are the models it will be measured against on price and on agentic coding:
- Claude Sonnet 5, Anthropic's cheaper agentic model, which scores 63.2% on SWE-bench Pro and competes directly on cost per agent run.
- GPT-5.6 Sol, OpenAI's most capable model of the cycle, though a restricted rollout means very few people can actually use it.
- Claude Fable 5, Anthropic's frontier model, pulled offline and then restored in July with a new cybersecurity classifier.
- Kimi K3, Moonshot's 2.8-trillion-parameter open-weight model, the option if you would rather not rent a frontier API at all.
Note the asymmetry. Those are mostly frontier or near-frontier models. Gemini 3.6 Flash is not trying to beat them on capability, it is trying to make the boring 80% of production traffic cheaper. Comparing it to Sonnet 5 on a hard reasoning task misses what it is for.
Where you can use it
Gemini 3.6 Flash went live in the Gemini app on launch day. Developers reach it through Google AI Studio, Android Studio, and Google Antigravity, the agent-first coding platform. Gemini 3.5 Flash-Lite is also rolling into Search.
The case for and against
| Reasons to switch | Reasons to wait |
|---|---|
| Output price down 17% and output token count down 17%, which compounds on generation-heavy work | Every benchmark here is Google's own, published at launch, with no independent verification yet |
| Knowledge cutoff moves forward fourteen months to March 2026 | Flash is still the cheap tier; if a task needs frontier reasoning this is the wrong model |
| Large coding gains on Google's evals, DeepSWE 37% to 49% | The efficiency figure is an average across a suite and may not hold for your prompts |
| Available immediately in the app, AI Studio, Android Studio and Antigravity | No Gemini 3.5 Pro to pair it with, and no committed date for one |
Who should use it
Worth moving to if you:
- Run production traffic on 3.5 Flash already. The migration is cheap and the pricing goes down, so the burden of proof is on staying put.
- Build agents that make many tool calls, where fewer reasoning steps translate directly into lower cost and lower latency.
- Kept hitting the old January 2025 knowledge cutoff when asking about recent tooling.
Probably not the right fit if you:
- Need top-end reasoning. That is what the Pro tier is for, and it is not here.
- Are mostly paying for input tokens. Input pricing did not move.
- Want independent benchmark confirmation before you migrate. Give it a few weeks.
Our take
This is a competent, unglamorous release, and unglamorous is fine. Most production AI spend is not frontier reasoning, it is thousands of small calls where a 17% cut on both price and token count shows up on the invoice at the end of the month. On Google's own numbers the coding improvement is real enough to matter.
What we would not do is read it as evidence that Google has caught up. A Flash release does not answer the question the Pro delay raised, and Google has now missed its own stated timeline once. We will revisit this once independent benchmarks land and once we have run it ourselves.
In line with our editorial policy, this article is based on launch-week specifications and Google's published benchmarks rather than extended hands-on testing, and it will be updated when that changes.
Frequently Asked Questions
When did Gemini 3.6 Flash launch?
July 21, 2026. Google DeepMind announced it alongside Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber, roughly two months after 3.5 Flash debuted at I/O 2026 in May.
How much does Gemini 3.6 Flash cost?
$1.50 per million input tokens and $7.50 per million output tokens. Input is unchanged from 3.5 Flash; output dropped from $9.00. Combined with the reported 17% cut in output tokens, an output-heavy workload should land noticeably below what the same job cost on 3.5 Flash.
Is Gemini 3.6 Flash better than Gemini 3.5 Flash?
On Google's own published evaluations, yes, and not by a small margin on coding: DeepSWE goes from 37% to 49% and MLE Bench from 49.7% to 63.9%. Computer use rises from 78.4% to 83.0% on OSWorld-Verified. These are Google's numbers, not independent testing.
What is the knowledge cutoff for Gemini 3.6 Flash?
March 2026, up from January 2025 on the previous model. That is a fourteen-month jump and the most practical change for anyone asking the model about recent libraries, APIs or events.
Where is Gemini 3.5 Pro?
Still not released. Google last updated its Pro tier in February 2026 and said in May that Pro was already in internal use. Bloomberg reported on July 16, 2026 that the launch had slipped because the model was falling short of internal targets. Google says it is testing with partners.
Sources
- Google, Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber announcement, July 21, 2026.
- 9to5Google, benchmark and pricing breakdown, July 21, 2026.
- TechCrunch, on the absent Gemini 3.5 Pro, July 21, 2026.