Apps · Research & analysis
Deep Research Kit
Multi-hop research agent over the open web.
Tokens / 7d
36B
Routing directive
auto:reasoning
On our metal
16%
Licence
Open source
Research agents fetch dozens of pages, then synthesize. The read step is long-context and cheap; the synthesis step is short and expensive. Splitting the two across routes roughly halves cost.
How it connects
- Read phase on auto:long-context, synthesis on auto:reasoning.
- Extended thinking enabled only for the final synthesis call.
- Citations enforced through structured output.
Traffic pattern
Two-phase: 90% of tokens in the read phase, 90% of cost in the synthesis phase without split routing.
gemini-3.1-progpt-5.5o3
Point it at the gateway
const notes = await client.chat.completions.create({
model: "auto:long-context", messages: readPhase });
const report = await client.chat.completions.create({
model: "auto:reasoning", reasoning: { effort: "high" },
messages: synthesize(notes) });