Google just did something odd with its naming. The model it calls "Flash" — the cheap, fast tier meant for throughput — now beats its own Pro tier on the benchmarks that describe real work. Gemini 3.5 Flash shipped generally available on May 19, 2026 at Google I/O, and it quietly collapsed the two-tier mental model the whole industry had been running on.
That model was simple: Pro for hard problems, Flash for volume. 3.5 Flash breaks it. On coding and agentic tasks, the Flash model outscores Gemini 3.1 Pro while running roughly 4x faster and often at less than half the cost. Google's own framing says it plainly — this is frontier intelligence at Flash latency.
The benchmarks that matter
Google frames 3.5 Flash directly against Gemini 3.1 Pro, and the gap runs the "wrong" way on the tasks agents actually do:
| Benchmark | 3.5 Flash | 3.1 Pro | Δ |
|---|---|---|---|
| Terminal-Bench 2.1 (coding) | 76.2% | 70.3% | +5.9 |
| MCP Atlas (tool use) | 83.6% | 78.2% | +5.4 |
| Finance Agent v2 | 57.9% | 43.0% | +14.9 |
| GDPval-AA (Elo) | 1656 | 1314 | +342 |
| CharXiv Reasoning (multimodal) | 84.2% | 83.3% | +0.9 |
| SWE-Bench Pro (Public) | 55.1% | 54.2% | +0.9 |
The +14.9 point jump on Finance Agent v2 is the largest single delta in the suite, and it's not an accident. It lines up exactly with where Google is deploying the model — Macquarie Bank is piloting 3.5 Flash to reason over 100-plus-page financial documents during customer onboarding.
There is a real trade here, and Google doesn't hide it. 3.5 Flash trails 3.1 Pro on the benchmarks that reward raw parametric knowledge: Humanity's Last Exam (40.2% vs 44.4%) and ARC-AGI-2 (72.1% vs 77.1%). Long-context recall is mixed too — at a dense 128k window, 3.5 Flash gives back 7.6 points on MRCR v2.
The honest read: if your work is a hard question answered in one shot, stay on Pro. If your work is an agent that has to plan, call tools, and finish a task, 3.5 Flash is the new default.
Speed and price are the real story
Frontier intelligence at 4x the speed of comparable frontier models, often at less than half the cost.
That's the pitch, and the pricing backs it up. 3.5 Flash runs $1.50 per million input tokens and $9.00 per million output tokens, with cached input at just $0.15. That's about 40% cheaper than Gemini 3.1 Pro (which sits at $2.50 / $15) while beating it on agentic work.
The catch — and it's a small one — is that 3.5 Flash is roughly 3x more expensive than the outgoing Gemini 3 Flash ($0.50 / $3). Google raised the floor of the Flash tier because the floor is now genuinely capable. For any agent harness that reuses a system prompt across dozens of tool calls, the 90% cache discount makes cached context, not per-request input, the dominant cost lever.
The specs round out the picture: a 1,048,576-token (1M) input window, up to 65,536 output tokens, text-image-audio-video input, dynamic thinking on by default, and a January 2026 knowledge cutoff. The API model ID is gemini-3.5-flash, no preview suffix — it was GA on day one.
Built for agents, not chats
3.5 Flash didn't ship alone. Google released it alongside Antigravity 2.0, a standalone desktop app for agent work with parallel subagent execution and scheduled background tasks. The Antigravity harness is co-optimized with 3.5 Flash specifically, which is how one run can spin up multiple subagents working in parallel on a single problem.
The model is also the engine behind Gemini Spark, Google's new personal AI agent that runs 24/7 and takes action under a user's direction. And it's now the default model for the Gemini app and AI Mode in Search globally — meaning hundreds of millions of everyday queries already run on it.
The enterprise rollout reads like a who's-who of long-horizon workflows: Shopify running parallel subagents for merchant growth forecasts, Salesforce Agentforce automating multi-turn enterprise tasks, Ramp doing multimodal OCR on messy invoices, Xero autonomously handling multi-week 1099 tax prep, and Databricks diagnosing issues across massive datasets. The common thread is work that crosses many tool calls — exactly where a 4x speedup compounds into finishing during a coffee break instead of over an afternoon.
What's next
Google confirmed Gemini 3.5 Pro is already in internal use and slated to ship the following month. If 3.5 Flash beats 3.1 Pro today, the interesting question is what 3.5 Pro does to the frontier when it lands. For now, Google has moved the frontier line down to the tier most people can actually afford to run at scale.
The Bottom Line
Gemini 3.5 Flash is the most consequential Flash release Google has shipped, and the modest version bump undersells it. It beats Gemini 3.1 Pro on coding and agents, runs 4x faster, and costs 40% less than Pro-class models — while giving up ground only on academic reasoning and dense long-context recall. If you're building agents that plan and execute, this is the model to reach for today. If you're asking one hard research question, wait for 3.5 Pro. Either way, the old "Pro for smart, Flash for cheap" rule of thumb just stopped being true.
