On May 19, 2026, at its I/O developer conference, Google launched Gemini 3.5 Flash, the first model in its Gemini 3.5 series. Google called it its strongest agentic and coding model yet and said it outperforms the larger Gemini 3.1 Pro on key agent benchmarks while keeping the speed the Flash line is known for. The launch also reset expectations for the Flash tier's price: at $1.50 per million input tokens and $9 per million output tokens, Gemini 3.5 Flash costs several times more than earlier Flash models, even as Google positioned it as far cheaper than comparable frontier models.
Key Facts#
- Release date: May 19, 2026, at Google I/O. Model ID
gemini-3.5-flash. - Availability: the Gemini Enterprise Agent Platform, Google AI Studio, Google Antigravity and the Gemini Enterprise app.
- Vendor-reported results: 76.2% on Terminal-Bench 2.1, 1656 Elo on GDPval-AA, 83.6% on MCP Atlas and 84.2% on the CharXiv multimodal benchmark, which Google said beats Gemini 3.1 Pro on key benchmarks.
- Pricing per million tokens: $1.50 input, $9 output and $0.15 for cached input, compared with $0.30 and $2.50 for Gemini 2.5 Flash.
- Context: about one million input tokens and up to roughly 65,000 output tokens.
- Reasoning controls: four thinking levels (minimal, low, medium and high), with medium as the default.
- Announced alongside: Gemini Omni, a model that generates and edits video from mixed inputs, with Gemini Omni Flash due for developers in the following weeks, and Gemini 3.5 Pro, which Google said was in testing and coming the next month.
What Happened#
Google framed Gemini 3.5 as the start of an agentic era for Gemini, with the Flash model leading the way rather than following a Pro release. Its Cloud announcement said 3.5 Flash delivers intelligence that rivals large flagship models on multiple dimensions at Flash-class speeds, and that it is ideal for long-horizon agentic tasks, often at less than half the cost of comparable models.
The benchmark selection matched that pitch. Terminal-Bench measures command-line agent work, GDPval-AA measures economically valuable knowledge work, and MCP Atlas measures how well a model uses tools exposed through the Model Context Protocol. A Flash-tier model beating the previous Pro model on all three was the headline claim, although, as always, the numbers came from Google.
Developer-facing changes arrived with the model. Google's quickstart for Gemini 3.5 Flash introduces four thinking levels and sets the default to medium, replacing the high default used by earlier Gemini 3 models, to balance intelligence against cost and latency. It recommends a stateful Interactions API for multi-turn and sub-agent workflows, adds an ultra-high media resolution for individual images, and requires version 2.0 or later of the Python Gen AI SDK for the new features.
Google paired the model with new agent products. It introduced Gemini Spark, a personal agent for Gemini Enterprise customers that is built on Gemini 3.5 Flash and runs tasks in the background, and a Managed Agents API on its agent platform that spins up hosted agents with a single call. Antigravity gained a desktop app and a command-line interface.
Background#
Gemini 3.5 Flash arrived three months after Gemini 3.1 Pro, and in the same spring as OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 and Opus 4.8. All three companies were concentrating on agentic coding and tool use, and OpenAI and Google had both raised per-token prices for their newest models while arguing that efficiency lowers the cost of finished work.
Google's Flash line had long been its value option. Gemini 2.5 Flash cost $0.30 and $2.50 per million tokens, and the Gemini 3 Flash preview $0.50 and $3. Pricing Gemini 3.5 Flash at $1.50 and $9 suggested that Google now positions Flash as a frontier model in its own right, not just a cheaper sibling.
Why It Matters for Developers#
The price change is the first thing to plan around. If you route high-volume work, such as classification, extraction or chat, to "the Flash model" by default, Gemini 3.5 Flash may multiply your bill. Keep routing explicit: use Flash-Lite models for simple tasks, reserve 3.5 Flash for agentic and coding workloads where its gains matter, and measure cost per completed task, as described in our guide to observability and cost control for LLM apps.
The new default thinking level also changes behavior. Medium reasoning is cheaper and faster than the high default of earlier Gemini 3 models, but it may change quality on hard tasks. Set the thinking level explicitly in configuration, and re-run evaluations when you upgrade instead of relying on defaults.
For .NET teams, the Google.GenAI SDK implements IChatClient, so 3.5 Flash can slot into Microsoft.Extensions.AI pipelines next to other providers. Google's quickstart for the new features targets its Python SDK, so check the .NET SDK's release notes before designing around features such as the Interactions API. The MCP Atlas result is also a reason to invest in well-designed tools: a model that uses MCP tools well rewards clear tool names, tight schemas and small tool sets, as covered in our guide to AI agent architecture patterns.
What's Next#
Google kept iterating after I/O. By September 2026 its developer models included Gemini 3.5 Flash-Lite, added in July at $0.30 and $2.50 per million tokens, and newer Gemini 3.6, 3.7 and 3.8 Flash models offered at an introductory $0.75 and $3.75 per million tokens through the end of 2026. That rapid cadence means the "right" Flash model can change within weeks.
Gemini 3.5 Pro was the other open item from I/O, which Google said would follow a month later. For developers, the lesson of 2026 is to treat model IDs as configuration, keep a standing evaluation suite, and re-check pricing pages whenever you adopt a new model, because both capability and cost now move quickly.