Levanto launches Sage2 for streaming decisions and dirt-cheap extraction

If you build voice agents or customer bots, you know the pain of someone changing their mind mid-sentence. One second they demand a refund, the next they just want to swap shoe sizes.

Traditionally, your backend has to repackage the entire chat transcript and throw it back into a massive language model just to figure out what changed. That kills latency and balloons your token bill.

Levanto rolled out Sage2 to solve exactly that problem. It gives developers two distinct building blocks: structured extraction and streaming decisions.

Real-time decisions as context grows

The standout trick in Sage2 is its Streaming API, running over a WebSocket connection. Each session handles up to 32 parallel questions and 16,384 tokens of combined context and instructions.

Instead of resending the full transcript on every turn, your client only pipes down whatever new words just arrived. The server keeps the running context and evaluates your questions continuously.

If you ask whether the customer has formally agreed to a quote, an objection keeps the flag at false. The second they say yes to a revised proposal, the state flips.

Levanto claims this streaming architecture evaluates ongoing decisions three times faster than standard patterns because it processes text as it grows.

Slashing extraction costs against frontier giants

Alongside live decisions, Sage2 handles typed field extraction. If a buyer says they want a midsize SUV for 450 dollars a month, the API pulls the car type and monthly budget while returning null for unspecified loan terms.

Once the customer adds a 60-month duration, it converts the text into standard ISO 8601 formats with exact character positions and confidence scores.

Levanto claims Sage2 extraction beats Cloudflare's Clairic and costs 166 times less than Claude Opus 5.5. The company is also rolling out beta support for reasoning, web grounding, and vision, with 20 million free tokens available to developers.

Why it matters

We think Sage2 is a smart push toward modularizing agent architecture away from one-size-fits-all model calls.

Using flagship models like Opus 5.5 to run routing logic and JSON extraction is pure overkill. It wastes money and introduces painful roundtrip delays that ruin conversational software.

By turning decision-making into an incremental stream, Sage2 gives agent builders a lightweight, dirt-cheap traffic cop that sits in front of heavy reasoning models. For realtime support bots, checkout flows, and dynamic tool selectors, this is an immediate win on both speed and infrastructure costs.

That said, you will still need to plan your states carefully. Pre-defining binary checks works wonders when workflows are strictly bounded, but open-ended conversations can easily push against that 32-question cap. Whether Sage2's low-cost extraction holds up across messier real-world slang remains the main thing builders need to test.