- Gemini 3.5 Flash scores 83.6% on MCP Atlas, beating GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro on multi-step workflows.
- Gemini 3.5 Flash introduces thought preservation, automatically carrying reasoning context across turns with no API changes required.
- The default thinking effort level drops from high to medium — teams migrating without testing could see silent quality regressions.
- Input pricing sits at $1.50 per million tokens, with batch inference cutting that in half, making scaled agentic deployment genuinely affordable.
Google’s Gemini 3.5 Flash Lands With a Bold Claim
Gemini 3.5 Flash is now generally available, and Google isn’t being subtle about what it wants developers to believe: that the traditional tradeoff between speed and intelligence in AI models is no longer a given. Flash-tier models have always been the budget option — fast, cheap, and quietly disappointing on anything requiring real reasoning depth. Google’s argument with this release is that the gap has meaningfully closed, at least for the workloads that are actually dominating developer conversations right now.
Those workloads are agentic. Multi-step pipelines, iterative code cycles, tool-chaining via the Model Context Protocol — this is where AI investment is flowing in 2025 and 2026, and it’s also exactly where Gemini 3.5 Flash is designed to compete. The benchmarks Google leads with aren’t the standard academic ones. They’re agentic-native, and the numbers are hard to dismiss.
The MCP Atlas Score That Should Get More Attention
On MCP Atlas — a multi-step workflows benchmark specifically designed around MCP tool chains — Gemini 3.5 Flash scores 83.6%. That puts it ahead of Gemini 3.1 Pro at 78.2%, Claude Opus 4.7 at 79.1%, and GPT-5.5 at 75.3%. Read that again: a Flash-tier model is outscoring frontier-class models from Anthropic and OpenAI on the benchmark most directly relevant to agentic development.
That result matters beyond the headline number. The infrastructure around MCP is expanding fast — tools like Glama.ai and a growing ecosystem of agentic middleware are being built on the assumption that MCP becomes the standard way AI agents talk to external systems. If Gemini 3.5 Flash genuinely leads on MCP orchestration quality while keeping costs at Flash-tier pricing, the economics of deploying these systems shifts considerably. You don’t have to burn Pro-tier budget to get Pro-tier agentic performance. That’s the real implication here.
The Finance Agent v2 results reinforce this pattern. Gemini 3.5 Flash scores 57.9% on financial analysis and decision-making tasks, ahead of Claude Sonnet 4.6 at 51.0%, Claude Opus 4.7 at 51.5%, and GPT-5.5 at 51.8%. For enterprises running financial workflows through AI agents — structuring reports, analysing data, chaining decisions — these aren’t abstract improvements.
Thought Preservation: The Architectural Change Worth Watching
Beyond the benchmark story, the most technically significant addition in Gemini 3.5 Flash is thought preservation. The model now automatically maintains intermediate reasoning across multi-turn conversations. When thought signatures are present in the conversation history, reasoning context carries forward — and the SDKs handle this automatically with no changes to your API calls.
That might sound like a subtle quality-of-life feature, but consider what it replaces. Most teams building multi-turn agentic sessions today are doing this manually: reconstructing context between turns, summarising prior steps, managing memory through external scaffolding code. It’s boilerplate that adds complexity, maintenance burden, and failure points. Thought preservation moves that logic inside the model itself.
The practical wins are clearest in iterative debugging and code refactoring — exactly the long-horizon tasks where maintaining reasoning continuity matters most. JetBrains reports that Gemini 3.5 Flash delivers coding and reasoning quality close to Gemini Pro levels while preserving the speed and cost profile that makes Flash viable for real-time developer workflows. Low-reasoning coding performance has improved by 10–20% compared to the previous Flash generation, which is a meaningful jump for an incremental release.
Enterprise validation backs this up from a different angle. Box benchmarked Gemini 3.5 Flash against Gemini 3 Flash on their own evaluation set — designed specifically to reflect the multi-step tasks their customers perform daily — and found a 19.6% improvement. For life sciences use cases, accuracy on data extraction and calculation tasks is reportedly 96.4% higher. Financial services firms building structured reports from data see 46.7% greater accuracy. Those are Box’s numbers, not Google’s, which gives them some independence.
Thinking Levels: What Actually Changed and Why It Matters for Migration
Gemini 3.5 Flash ships with a revised thinking system that every team migrating from Gemini 3 Flash Preview needs to understand before they touch production. The numeric thinking_budget parameter is gone. In its place is a string enum called thinking_level with three values: low, medium, and high.
More importantly, the default has changed. In Gemini 3 Flash Preview, the default thinking effort was high. In Gemini 3.5 Flash, it’s medium. Google’s reasoning is that medium delivers strong results across most tasks while being faster and more cost-efficient — which is likely true. But if your team migrates without explicitly testing for quality regressions, you could end up running production workloads at a lower reasoning level than you had before, and the degradation might not be immediately obvious.
Google’s own recommended framework is sensible: start at medium as your baseline, drop to low for speed-sensitive agentic loops where latency matters more than depth, and escalate to high only for hard reasoning problems or complex mathematics. For most agentic tasks, medium should be the right default — but verify that with your specific workloads before committing.
One more migration note that’s easy to miss: thought preservation is on by default. That’s generally a good thing, but it does increase token usage across multi-turn sessions. Teams tracking inference costs closely should account for that when modelling expenses post-migration.
Gemini 3.5 Flash Pricing and Availability
On pricing, Gemini 3.5 Flash costs $1.50 per million input tokens and $9.00 per million output tokens — thinking tokens included in that output rate. Context caching runs $0.15 per million tokens, with storage at $1.00 per million tokens per hour. Batch inference halves all those rates, which makes high-volume offline workloads significantly more affordable. A free tier is available through Google AI Studio for experimentation.
The model is live now under the ID gemini-3.5-flash (last updated May 2026) and accessible through the Gemini API, Google AI Studio, the Gemini App, Google Antigravity, the Gemini Enterprise Agent Platform, and Android Studio. It supports the full toolset: function calling, structured output, search grounding, Google Maps grounding, URL context, file search, code execution, and thinking — all combinable in a single request.
One limitation worth flagging: Computer Use is not supported in Gemini 3.5 Flash. Teams running computer-use workloads should stay on Gemini 3 Flash Preview for now. Google hasn’t indicated a timeline for adding this capability.
What This Means for the Broader AI Model Market
Google has been playing catch-up narrative for most of the past two years, but Gemini 3.5 Flash is a case where the numbers tell a genuinely competitive story. Beating Claude Opus 4.7 and GPT-5.5 on an agentic benchmark at Flash pricing isn’t a minor footnote — it’s a direct challenge to the assumption that serious agentic work requires frontier-model spend.
The timing is also notable. As MCP adoption accelerates and more production agentic systems get deployed at scale, cost efficiency at the inference layer becomes a real competitive variable. A model that leads on MCP orchestration benchmarks while charging less than a tenth of what frontier models cost per token changes the conversation about what’s economically viable to build. Developers who dismissed the Flash tier as a prototyping tool should probably revisit that assumption.
