Two new Google models are selectable for any Dawn agent: Gemini 3.5 Flash for fast multimodal work, and Gemini 3.1 Pro Preview for deeper reasoning on complex tasks. Both are available on every Dawn plan with no separate enable step.
Dawn Team
Product Engineering
Two Google models join the Dawn model catalogue this week: Gemini 3.5 Flash for fast multimodal work, and Gemini 3.1 Pro Preview for deeper reasoning on complex tasks. Both are selectable wherever agents pick a model in Dawn, and both are available on every plan with no separate enable step.
The two new entries appear in the Google section of the model picker alongside the rest of the catalogue. Pick the one you want on the work that benefits, save, and the next message that agent answers uses the new model.
Gemini 3.5 Flash launched at Google I/O in May 2026 as the first model in the Gemini 3.5 series. It is aimed squarely at agentic work and coding, the kind of run where latency and tool-use reliability matter as much as raw reasoning. Google’s published numbers put its output speed at roughly four times other frontier models, which is what makes it a good fit for high-volume agent runs that have to call tools, read responses, and decide what to do next without dragging out each turn.
Inputs cover the full multimodal range: text, images, audio, and video, with a 1M token input context window. Output is text with a 64K token output capacity.
On the public benchmarks Google released with the model, Gemini 3.5 Flash actually outperforms the larger Gemini 3.1 Pro on several challenging tasks, including 76.2% on Terminal-Bench 2.1 (a coding benchmark), 1656 Elo on GDPval-AA (real-world agentic tasks), 83.6% on MCP Atlas (scaled tool-use reliability), and 84.2% on CharXiv Reasoning (multimodal understanding). In practice that means Flash is a strong default for most Dawn workloads where Gemini is the right fit, especially anything involving images, audio, or video as part of the input.
Gemini 3.1 Pro Preview is Google’s deeper reasoning model. The headline figure Google has shared is its 77.1% on ARC-AGI-2, a benchmark for novel logic patterns, which is more than double what Gemini 3 Pro achieved. The model is built for the work that benefits from extended thinking before answering: architectural decisions, complex code reviews, multi-step plans with tradeoffs to weigh, and long-horizon agent runs where each turn’s quality compounds into the next.
Inputs cover the same multimodal range as Flash (text, images, audio, video, and code), with the same 1M token input context. The model can handle up to 900 images per prompt, up to 8.4 hours of audio, or up to one hour of video in a single request. Output is text with the same 64K token output capacity.
A “thinking level” parameter controls how much reasoning the model applies before answering, which gives agents using this model a knob for trading off cost and speed against depth on a per-request basis.
Both models are good. The question is which fits the work an agent is actually doing.
Flash is the right default for fast tool-using runs, high-volume conversation, multimodal input (a screenshot to triage, a voice note to act on, a short video to summarise), or coding tasks where the agent benefits from latency staying low across many tool calls. It is also a good fit for agents that handle a wide mix of requests, including the short ones where deeper reasoning would be wasted.
Pro Preview earns its keep on the agents whose work most benefits from extended thinking: planning agents, review agents, long explore runs with several connected sources to weigh, and anything where the answer needs to hold a complex problem in mind across many turns. Pro Preview costs more per response than Flash, which is the usual tradeoff for the deeper reasoning. The logic from the GPT-5.5 release applies: extra reasoning is worth paying for when the answer has to land right the first time.
Both models appear in the Google section of the model catalogue. Open the agent you want to upgrade, change the model, save. The agent’s system prompt, skills, memory pages, and connected sources all carry over unchanged.
A workspace can run several agents on different models at the same time. A triage agent answering quick questions in Slack might sit on Gemini 3.5 Flash for the multimodal speed. An engineering-review agent that operates on pull requests might move to Gemini 3.1 Pro Preview for the reasoning depth. An agent handling customer comms might sit on something else entirely. Each agent picks its own model and the cost reporting follows that choice.
Both models are available across every Dawn plan with no separate enable step or upgrade gate. The picker shows them on free workspaces, paid workspaces, and BYOK workspaces alike. Workspaces using Dawn-managed credentials see the new entries appear automatically. Workspaces using bring-your-own-key can point Dawn at a workspace Google API key for the same routing.
The release is exactly what the title says: two Google models in the catalogue, multimodal across the board, available everywhere with no gating. Existing agents keep their existing models. Existing connected sources, skills, and memory pages keep working with the new models the same way they worked with the old. There is no breaking change to migrate around and no feature flag to flip.
If your team has been waiting to evaluate either of these in real day-to-day use, the path is the usual one: open the agent you want to try it on, change the model, and use it on work where the new model is likely to earn its keep.