A more token-efficient workhorse model
Built on 3.5 Flash: same 1M context and adjustable thinking, with the knowledge cutoff moved to March 2026.
Overview
Gemini 3.6 Flash is Google DeepMind's workhorse model, released on July 21, 2026. It is built directly on Gemini 3.5 Flash and belongs to the Gemini 3 family of natively multimodal reasoning models. It keeps the 1M context window and adjustable thinking levels of 3.5 Flash, moves the knowledge cutoff forward to March 2026, and improves across coding, knowledge work, and multimodal understanding.
The pitch for this generation is not "bigger" but "cheaper to run". On the Artificial Analysis Index, 3.6 Flash consumes 17% fewer output tokens than 3.5 Flash for comparable tasks, and up to 65% fewer on long-horizon software engineering evals such as Datacurve's DeepSWE. It also takes fewer reasoning steps and tool calls to finish a multi-step workflow. CrossModel exposes it as gemini/gemini-3.6-flash.
Key capabilities
| Dimension | Detail |
|---|---|
| Context window | 1,048,576 tokens (about 1M) |
| Max output | 65,536 tokens (about 64K) |
| Knowledge cutoff | March 2026 |
| Input modalities | Text, image (Google's native model also supports audio, video, and PDF) |
| Output modalities | Text |
| Tools | function calling, structured outputs, streaming, adjustable thinking levels, context caching, code execution, computer use (preview) |
Google publishes a single input/output rate for 3.6 Flash with no long-context tier, which suits frequent, cost-sensitive workloads at scale. See live rates in the model catalog.
Token efficiency
Same job, fewer tokens spent
Google's framing: fewer reasoning steps and tool calls, which lowers the total cost of a single agentic task.
For teams building agents, the real bill is not "price per million tokens" — it is "how many tokens does one task burn". That is exactly what 3.6 Flash optimizes: fewer output tokens, fewer reasoning steps, fewer repeated tool calls, and fewer wasted code edits and execution loops. −17% is the cross-task average; −65% is the extreme observed on DeepSWE — and the workloads with the biggest gap tend to be the long-horizon tasks where 3.5 Flash was most prone to wandering.
Benchmarks
Three axes dominate the evaluation story: long-horizon coding and terminal work, computer use, and long-context retrieval. The official model card reports all of these alongside 3.5 Flash, 3.1 Pro, GPT-5.6 Luna, Grok 4.5, and Claude Sonnet 5.
Coding, terminal, and computer use
Coding and agent scores move up across the board
Values in parentheses are Gemini 3.5 Flash on the same benchmark, from the official model card (July 2026).
The clearest gains are in long-horizon engineering work: DeepSWE v1.1 rises from 37% to 49% and MLE-Bench (machine learning engineering) from 49.7% to 63.9% — the latter lands near Claude Sonnet 5's 66.9%, an unusual position for a Flash-tier model. OSWorld-Verified 83.0% is the one row where 3.6 Flash leads every model in the comparison, and Google has made computer use a built-in client-side tool in the Gemini API and Gemini Enterprise. Terminal-bench 2.1 78.0% and SWE-Bench Pro 58.7% are steady progress but still trail pricier flagships like Luna, Grok 4.5, and Sonnet 5.
Knowledge work and long context
Long context is the single biggest jump this generation
MRCR v2 at the full 1M window goes from 26.6% to 54.0%, moving full-window retrieval into usable territory.
GDPval-AA v2 climbs from 1349 to 1421 Elo, a measure of real knowledge-work output. Customers such as Hebbia and Harvey report it is particularly capable at multimodal tasks — document parsing, chart and data analysis, report drafting — which the chart-reasoning result of CharXiv 85.2% (89.4% with tools) backs up.
The real leap is long context: MRCR v2 @128K goes from 77.3% to 91.8%, and the full 1M window doubles from 26.6% to 54.0%. The previous generation's conclusion — that precise retrieval across a filled million-token window remains hard industry-wide — is meaningfully rewritten here. 54% is still not high-fidelity, but it moves full-window retrieval from "basically unusable" into "usable with verification".
Safety
3.6 Flash ships with enhanced Frontier Safety safeguards in the CBRN (chemical, biological, radiological, nuclear) and cyber-offense domains, making it substantially more resistant to jailbreaks while being trained to minimize refusals for beneficial uses. The model card reports a 5.45 percentage-point improvement in multilingual safety and 1.35 points in text safety over 3.5 Flash, with unjustified refusals up only 0.25 points. Under the Frontier Safety Framework it reaches no Critical Capability Levels.
When to use it
- Scaled agent workflows: frequent multi-step tool calls, computer use, and terminal automation, where token efficiency directly sets the cost per task.
- Long-horizon coding and migrations: DeepSWE- and MLE-Bench-shaped work that needs sustained edit, verify, and correct loops.
- Very long document understanding: cross-document retrieval over contracts, codebases, filings, and meeting transcripts — near-lossless within 128K and usable at the full window.
- Multimodal knowledge work: document parsing, chart and data analysis, and report drafting with structured outputs.
If you are on Gemini 3.5 Flash today, 3.6 Flash is close to a no-tradeoff swap: same context and interface shape, better scores, fewer tokens burned. Reach for a Pro tier or another vendor's flagship only when you need top-end single-shot coding results.
CrossModel exposes Gemini 3.6 Flash through an OpenAI-compatible /v1/chat/completions API. Current pricing is available in the model catalog.