Gemini 3.5 is the model that rewrote what a frontier benchmark looks like: agent scores first, language scores second, and four-times-faster output as the headline number.
On 19 May 2026, at Google I/O, the company unveiled Gemini 3.5 — the next iteration of its frontier model family, explicitly framed as "frontier intelligence with action". Gemini 3.5 Flash is generally available today; Gemini 3.5 Pro rolls out the following month. The benchmark table Google led with is the part that reveals the strategy: 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas, 1656 Elo on GDPval-AA, 84.2% on CharXiv Reasoning, and a claim of four-times-faster output tokens per second than other frontier models. The pitch is no longer "smarter than GPT" or "more thoughtful than Claude". The pitch is "more useful as an agent than anything else available".
What changed in 3.5
Two model variants ship in the 3.5 family. Gemini 3.5 Flash is the cost-effective workhorse: optimised for agentic long-horizon tasks, coding, and high-volume API workloads. Google's headline claim is that 3.5 Flash outperforms Gemini 3.1 Pro on coding and agentic benchmarks while maintaining the speed advantage. Gemini 3.5 Pro is the deeper-reasoning sibling, coming next month, and is the model Google expects enterprises and Antigravity Pro users to default to for complex multi-step work. Both variants ship inside Gemini Enterprise, the Gemini API, Google AI Studio, Android Studio, and Google's consumer Gemini app.
Three capability shifts matter most. First, agentic execution. Gemini 3.5 is tuned to handle long-horizon workflows — multi-tool, multi-step, multi-hour tasks where the model has to plan, execute, observe, replan, and keep going without a human in the loop on every cycle. Second, multimodal graphics and UI generation. Where Gemini 3 could read images well, Gemini 3.5 generates richer, more usable UI surfaces from prompts — the work that turns "design me a settings page" into a render the developer can actually inspect. Third, raw speed. Four-times-faster output tokens per second is the kind of headline number that matters most for agentic workloads, where end-to-end latency compounds across many model calls.
3.5 Flash delivers frontier performance for agents and coding, excelling at complex long-horizon tasks that deliver real-world utility.
— Google, Gemini 3.5 launch (https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/)
Why agent benchmarks lead the table
Look at the order of the Gemini 3.5 benchmark table. Terminal-Bench 2.1 — a measurement of how well a model can complete real terminal tasks — is at the top. MCP Atlas — a measurement of how well a model can use Model Context Protocol tools — is next. GDPval-AA and CharXiv reasoning follow. The order is the message. The dominant model-evaluation question has shifted from "how well does the model converse" to "how well does the model act". Gemini 3.5 is the first major frontier release whose top-line numbers all describe action, not language.
The shift is consistent across the field. Anthropic's Claude evaluations now lead with SWE-bench Verified and tool-use scores. OpenAI's recent releases foreground SWE-bench, GAIA, and tool-use benchmarks. The labs have collectively decided that the next year of revenue comes from agentic workloads, and the benchmarks have re-sorted themselves to match. Whether the benchmarks measure what users actually care about is a separate, still-open question — but the field's answer for the next twelve months is "agent benchmarks first".
💡The headline benchmark is the strategy
When the headline numbers on a frontier release describe how well the model acts, not how well it speaks, the customer the lab has in mind has changed. The new customer is a developer building an agent, not a user typing into a chatbox. Every product decision downstream flows from that customer choice.
How 3.5 lands inside Google's stack
Gemini 3.5 does not arrive alone. The model is the engine; Antigravity 2.0 is the chassis; Google AI Studio, Android Studio, and Firebase are the steering wheel and the road. The model is also threaded into the consumer Gemini app, AI Mode in Google Search, and the Gemini Enterprise Agent Platform. The integration is the strategic move; the model is the licence to make the integration possible.
That packaging matters because of who Google is competing against on each surface. On the consumer side, ChatGPT remains the default for hundreds of millions of users. On the enterprise side, Anthropic's Claude is the preferred frontier model for an increasing slice of regulated industries, and the recent Stainless acquisition tightened Anthropic's developer story. On the search side, Perplexity, Brave Search, and the open-source RAG stacks all chip at Google's default position. Gemini 3.5 has to be defensible on every one of those surfaces, and the agentic-action framing is Google's answer for why a single model family can be.
The cost picture is the other half of the story. Google priced an AI Ultra plan at $100 per month with 5x higher usage limits in Antigravity. That price point sets a competitive floor: any agent platform charging more than $100 for comparable usage has to justify the premium with substantively better behaviour, not just better branding.
What the numbers actually mean
Benchmarks are messy. Terminal-Bench 2.1 at 76.2% sounds impressive on its own, but the question is how often the remaining 23.8% of tasks fail in ways that matter. MCP Atlas at 83.6% measures whether the model uses tools correctly, but does not measure whether the model picks the right tool in the first place. GDPval-AA at 1656 Elo is a relative score on a curated economic-value test set; it is not a ground-truth measurement of usefulness in the wild. CharXiv at 84.2% is real progress on multimodal reasoning, but the harder out-of-distribution multimodal cases remain.
What the numbers do tell us is direction. Gemini 3.5 represents a step-change improvement in agentic reliability over Gemini 3.1 on the metrics Google chose to measure. Independent evaluations from labs and academic groups will take a few weeks to land; until they do, the numbers should be read as "Google's strongest agentic frontier release yet", not as a settled ranking against Claude or GPT. The honest comparison is "Gemini 3.5 Flash beats Gemini 3.1 Pro on Google's own benchmark suite", which is a meaningful internal improvement, not a definitive cross-lab ordering.
What this means for engineering teams
For TK's readers running engineering teams, four practical shifts follow. First, treat Gemini 3.5 Flash as the new default for any greenfield agentic workload where Google Cloud is already in the stack — the four-times-faster output and the 76.2% Terminal-Bench score change the latency-and-reliability profile of agent workflows by enough to matter. Second, hold off on committing to Gemini 3.5 Pro until next month; the Pro release is where Google will actually try to dethrone Claude on regulated-enterprise work, and the cross-lab comparisons are the ones to wait for. Third, the multimodal UI-generation capability is worth a real evaluation for any team shipping a design-system or admin-shell product; the workflow from "prompt" to "rendered settings surface" is short enough to change how design and engineering partner. Fourth, the Antigravity integration changes the unit economics of running agents at scale; reassess any in-flight internal agent-platform projects against the Managed Agents specification before next quarter.
Frequently Asked Questions
These are the questions engineering leaders, AI researchers, and platform architects have been asking since the I/O announcement. Short answers follow, drawn from Google's launch and the broader I/O 2026 coverage.
What is Gemini 3.5?
In short, Gemini 3.5 is Google's frontier model family launched at I/O 2026, with a Flash variant available today and a Pro variant rolling out next month. The answer, simply put, is the next-iteration Gemini, tuned explicitly for agentic action — long-horizon tasks, tool use, code execution, and multimodal UI generation. The key is that the benchmark table leads with agent scores rather than language scores, which signals that the model is built for developers running agents, not users typing into a chatbox.
How does Gemini 3.5 compare with Gemini 3.1?
According to Google's launch, Gemini 3.5 Flash outperforms Gemini 3.1 Pro on coding and agentic benchmarks while maintaining four-times-faster output tokens per second. Research from the agentic-workload market shows latency compounding is the biggest cost in long-horizon agent tasks. Data from the benchmark table reveals 76.2% on Terminal-Bench 2.1 and 83.6% on MCP Atlas — both meaningful step-changes over the prior generation on the metrics Google publishes.
Why do agent benchmarks lead the announcement?
The shift from language benchmarks to agent benchmarks reflects where the labs expect the next year of revenue to come from. According to the cross-lab pattern visible across Google, Anthropic, and OpenAI announcements, every major frontier release in 2026 leads with tool-use and code-execution scores rather than conversational metrics. Analysis of customer adoption shows agentic workloads are the fastest-growing category of API consumption. Evidence from the benchmark choices reveals the labs have re-sorted what counts as a top-line number to match.
Who is Gemini 3.5 for?
Gemini 3.5 is for developers building agents on the Gemini API, engineering teams running agentic workflows inside Antigravity, regulated enterprises buying through Gemini Enterprise, and consumers using the Gemini app and AI Mode in Google Search. In other words, the model is targeted at four buyer profiles at once. The integrated stack — model plus Antigravity plus Studio plus Enterprise — is what lets Google address all four from the same release.
What are the real limits of Gemini 3.5?
Analysis of the launch reveals three durable limits. First, the independent-evaluation lag: cross-lab benchmark verification takes a few weeks, so the headline numbers should be read as Google's strongest internal release rather than a settled ranking versus Claude or GPT. Second, the benchmark-coverage gap: Terminal-Bench, MCP Atlas, and CharXiv measure useful capabilities but do not capture every dimension of agent reliability, especially under adversarial or out-of-distribution conditions. Third, the price-floor risk: at $100 per month for AI Ultra, the model is priced aggressively, and any margin-pressure cycle could change that. Each limit is structural, not cosmetic.
