Google is quietly changing what "Flash" means.

Gemini 3.5 Flash is the first model in the 3.5 family, and Google has not been subtle about the framing. It is positioned as "frontier intelligence with action" — in practice, the strongest agentic and coding model Google has shipped, at output speeds around 4× other frontier-class models. It is already the default in the Gemini app and in AI Mode in Search, and it is the engine behind Antigravity, the Gemini Enterprise Agent Platform, and the upcoming personal agent ("Gemini Spark").

Two things make this release a real reset rather than a routine point update. First, the model card and independent measurements both show Flash beating the previous-generation Pro on the benchmarks that map to real builder workflows — Terminal-Bench 2.1, MCP Atlas, OSWorld-Verified, GDPval-AA, SWE-Bench Pro. Second, every product announcement around it talks about agents and long-horizon workflows, not chatbots. Google's own demos pitch 3.5 Flash as something that can refactor an entire legacy codebase to Next.js or stand up an OS inside Antigravity, with sub-agents fanning out underneath it.

If you were used to thinking of Flash as the cheap, fast option you reach for when you do not need Pro, the rest of this post is mostly about how aggressively that mental model is now wrong.

Strip out the marketing and there are four substantive shifts in this release.

If you have been tracking the Gemini line, the surprise is less "is this model better" and more "Google reshuffled which tier does what." A Flash that wins on agentic benchmarks and ships as the consumer default is a signal about Google's roadmap, not just a model card.

The headline numbers are clustered in agentic and tool-use evals. Pulling the four most-cited rows together makes the pattern obvious: 3.5 Flash beats 3 Flash by a large margin, and it beats 3.1 Pro by a smaller but consistent one on these axes.

Three things to read out of that table. The agentic gap between Flash and Pro is large and consistent — these are not noisy wins. The multimodal numbers are strong but incremental rather than transformative. And on long-context retrieval at 128k, Flash still trails Pro, which is exactly the seam that a future 3.5 Pro is presumably being shipped to close.

3.5 Flash was co-developed with Antigravity, Google's agent-first IDE. That co-design shows up in the kinds of demos Google leads with: agents iterating an OS inside Antigravity, refactoring entire legacy codebases to Next.js, spawning sub-agents to generate, test, and fix code in a loop. Benchmarks like SWE-Bench Pro back it up — gains over 3 Flash, parity or slight wins over 3.1 Pro.

If you are building anything in the "coding agent with tools" shape — a Cursor-style IDE companion, a CI-fixing bot, a repo-wide migration runner — 3.5 Flash is now the most capable Gemini-family option and one of the most cost-competitive frontier choices in absolute terms, even after the price hike.

Text, image, audio, video, and PDF inputs are all in. The context window is 1M in / 64k out. CharXiv Reasoning sits at 84.2% (slightly ahead of 3.1 Pro) and MMMU-Pro at 83.6% (ahead of both 3 Flash and 3.1 Pro). For everyday multi-doc, multi-modal agent work this is "frontier-tier enough" and probably the boring right choice.

The caveat is long-context retrieval: at 128k MRCR v2, Gemini 3.1 Pro still leads. If your stack stuffs a single very large document into context and depends on accurate retrieval inside it, 3.5 Flash is not unambiguously an upgrade.

A lot of "frontier" announcements stay locked inside a Labs preview or a single SKU. 3.5 Flash launched simultaneously into the Gemini app, Search AI Mode, AI Studio, the Gemini API, Android Studio, Gemini Enterprise, and Antigravity. That breadth is itself a statement: Google is treating this model, not Pro, as the new product floor.

The Flash brand used to mean "cheap and fast." Half of that is still true. Artificial Analysis reports list pricing around 1.50 USD / 9.00 USD per million input/output tokens, versus 0.50 / 3.00 for Gemini 3 Flash — about a 3× jump on the rate card. On their Intelligence Index workloads, where output tokens dominate, real runs come in over 5× more expensive than 3 Flash and roughly 75% more expensive than Gemini 3.1 Pro.

The honest read is that the agentic tier has its own price band now, and Google is calling it "Flash" because it inherits the latency profile, not the cost profile. If your workload is verbose or tool-heavy — i.e. the workload the model is optimized for — you should re-cost it before flipping the model name in your config.

3.5 Flash is explicitly tuned for long-horizon, multi-step, tool-using sessions, with sub-agents and pauses for human confirmation at decision points. That is fantastic if you are building Antigravity-style systems. It is largely irrelevant if you are building a documentation Q&A bot, a customer support assistant, or a small refactor helper that calls the model once and returns a string.

For those single-pass workloads, you are paying a premium for headroom you do not use. The right move there is often "stay on 3.1 Pro or 3 Flash, wait for 3.5 Lite or 3.5 Pro, then re-evaluate" — not "ship the newest model because it is the newest."

The model card itself acknowledges that 3.5 Flash does not meaningfully surpass Gemini 3.1 Pro on frontier-safety capability evals and trails Pro on some hard reasoning and long-context benchmarks. Independent write-ups specifically call out 128k MRCR v2 as a place to stay on 3.1 Pro. For deep legal, scientific, or multi-book single-pass workloads, "newer" is not the same as "better" here.

3.5 was developed under Google's Frontier Safety Framework, and the published claim is strengthened cyber and CBRN safeguards, plus better calibration on sensitive questions (less over-refusal). Independent of how good the model itself is, broad agent availability — including personal agents like Gemini Spark that run continuously — increases the blast radius for misaligned behavior. Google is already in litigation tied to prior Gemini conduct. Enterprises shipping this need their own guardrails, audit layers, and incident playbooks; the model card does not absorb that risk for you.

The 3.5 Flash launch is paired with explicit deprecation timelines in the Gemini API docs and a parallel notice for Vertex AI. The earliest shutdown date for the 2.5 family is October 16, 2026, with at least six months of notice once the Gemini 3 migration path is fully GA.

The direction of travel is unambiguous: you have runway, but you are expected to be on 3.x or 3.5 in production within the next year. Two specific traps to plan around:

Google has said the 3.5 family will fill out — Pro is being used internally and is the next public drop, and there is no 3.5 Lite model card yet. We can say plausible things about each without inventing numbers.

The framing Google has used publicly is that Pro orchestrates and Flash executes: Pro holds the long-horizon plan, Flash carries out brute-force tool calls and parallel work as sub-agents underneath it. Combine that with the fact that 3.5 Flash slightly regresses against 3.1 Pro on long-context and hard reasoning, and the implication is clear: 3.5 Pro is being built to restore that lead and to act as the top of an explicitly hierarchical agent stack.

Expect: highest-reasoning, deepest long-context, more expensive, and pitched as the orchestrator in multi-agent topologies — not as a chat default.

3.5 Lite (or 3.5 Flash-Lite, the naming is not pinned yet) has not shipped, but the prior at 3.1 Flash-Lite is "scalable thinking model for high-volume tasks at low cost and latency." Read forward: a tier optimized for tagging, summarization, classification, and the kinds of simple agents where you fire millions of small calls and care about per-token cost above all else.

It will almost certainly absorb the migration target currently pointed at 3.1 Flash-Lite, finishing the deprecation table.

One underdiscussed shift between 2.5 and 3.5: Google has effectively split "live" into two product lines. The streaming, voice-and-video flash-live endpoints are being routed to gemini-3.1-flash-live-preview — that remains an explicitly low-latency, conversational mode. Separately, 3.5 Flash is designed for multi-hour autonomous sessions, with pauses for human confirmation at decision points, and is positioned as the core engine for Antigravity-style harnesses and the Gemini Enterprise Agent Platform.

In 2.5, long-running and live felt like preview features grafted onto a chat model. In 3.5, the long-running case is a first-class product surface — that is the actual productization step here, not just "we added streaming."

The honest version of this story is that "agent > chat" is not obviously true for everyone. A large share of real workloads are still fundamentally single-pass: RAG-style Q&A over docs, customer support, simple summarization, unit tests, small refactors, lightweight code suggestions. For those, the model gets called once, returns a string or JSON, and the workflow moves on.

Agentic stacks add failure modes that simply do not exist in a single-pass call: tool misconfiguration, environment drift, brittle DOMs or terminals, race conditions between sub-agents, and a much larger surface for prompt-injection. Buying a model that is optimized for those workflows when you do not run them means paying for headroom you cannot use, and inheriting risks you do not have to take.

The right framing is not "should I move to 3.5 Flash?" It is "is my workload shaped like an agent?" If yes, 3.5 Flash is probably the strongest Google option in front of you right now. If no, 3.1 Pro / 3 Flash / future 3.5 Lite are all defensible choices, and "default in the Gemini app" is not a reason to migrate.

One question I keep getting from people on Gemini 2.5 Flash: the recommended migration target is 3.5 Flash, but the new model is 3× the per-token cost, optimized for agents you do not run, and verbose in ways that inflate real bills further. What does the cross-vendor landscape look like if you treat the 2.5 Flash price/feature profile as the actual constraint, not "stay on Google"?

The anchor: 2.5 Flash sits at $0.30 in / $2.50 out per million tokens, 1M context, 65K output, multimodal (text + image in → text out), prompt caching, GA on Vertex AI. Below is the full landscape — every serious cross-vendor option in the same price-and-capability neighborhood — with 2.5 Flash at the top and 3.5 Flash at the bottom as bookends. Each row is tagged with the single biggest reason it fits or does not.

Reading note: every row below supports text + image input and prompt caching — those columns are dropped to keep the table readable. The PDF column captures the one modality that actually varies (native, via Converse API, or partial). All rows output text.

Reading the table:

The honest framing: "stay on Google" is one valid migration, but it is not the only one, and given how aggressively Google has repriced Flash, it is worth at least putting two or three of these on a real eval bench. The 2.5 → 3.5 hop is not free — and the cross-vendor hop, evaluated this quarter, may be the cheaper version of the same work.

If you are on 2.5 Flash today and the migration target is 3.5 Flash, do not change the model string in your config and call it done. The deprecation is real, but the new model has a different cost profile, different output verbosity, and a different design target. Treat the cutover as a real re-eval.