Trio-Flash: Give your agent a next move

One current state. A few allowed moves. One inspectable answer. Trio-Flash is a hosted decision API for developers building GUI agents, games, and other interactive systems.

It is also a first look at where we are taking Trio. Our longer-term work is about intelligence for physical operations. Trio-Flash brings one focused part of that idea to developers now: turn a task, a current state, and a short list of possible actions into a decision your software can inspect.

Start with a problem you can see

We are preparing Trio Falling Blocks as a compact demo. For each falling piece, the program enumerates legal placements and gives the model a short list of candidates. The board can then show the choice and its consequence. We will publish a current-model recording after the end-to-end run passes release checks.

The demo is designed to be easy to audit: board state, candidate moves, probabilities, chosen action, and resulting board are visible together. Fixed seeds let us replay the same piece sequence across runs. One request controls one piece, so the connection between a decision and its result stays visible.

Falling Blocks is a compact test, not the destination. Our first users are developers training and evaluating GUI agents. They can send a goal and a structured description of the current interface, then ask Trio-Flash to choose among allowed controls. The same API shape also fits games and workflow routing. In every case, the application owns the state, choices, safety rules, and execution. The preview accepts text or JSON state, not screenshots or live video.

How we will measure it

Jev set a useful bar for hosted decision models. We will compare paired decisions on the same frozen tasks and report actual latency and cost per task. A matching price per million tokens alone does not mean two services cost the same per decision: token counts can differ. We are also interested in whether probabilities help an application know when to act and when to ask for review.

Before publishing a comparison, we will run both services on a frozen evaluation, record paired decisions, and measure live latency on the launch hardware. We will publish the protocol, date, and results if that gate passes. Until then, this post makes no accuracy or latency comparison.

A cloud API, built for experiments

Send a natural-language task, a text or JSON state, and two to eight choices with IDs your application understands. The response contains a selected choice, probabilities, a decision status, and usage. Your code can execute the choice, request human review, or wait for more information.

Trio-Flash will launch as a closed, hosted service. We are not releasing its weights, model size, training method, or internal implementation during the preview. We will show the parts developers need to evaluate it: the API contract, inputs and outputs, reproducible demos, measured latency, benchmark protocol, service limits, and billing behavior.

The planned limited-preview price is $0.042 per million billed input tokens, with free output and a $5minimum top-up. It is pay as you go, with no required subscription. This is a limited-preview rate and may change before general availability. The current rate card and service availability will be shown in the Playground before a paid request.

What Trio-Flash does and does not do

Trio-Flash makes a bounded decision from the state and choices you provide. It does not watch a camera feed, operate a robot, move money, or run a business on its own. Those systems must observe the world, enforce their own constraints, and decide whether to execute a model choice.

The limited preview is for developers who want to test small, repeated decisions against outcomes they can measure. Bring a GUI-agent evaluation, a game, or a real workflow with a safe review path. The useful feedback is concrete: where did the model help, where did it hesitate, and would you keep calling it?

Try your own next move

The playground and API keys will open after the live service, billing flow, and launch measurements pass validation. Until then, see the Trio-Flash product preview and the example API request.

Trio-Flash is not affiliated with or endorsed by Jev. Preview capabilities vary by task; evaluate decisions on your own examples before relying on them.