Toast 1: a drop-in search agent that cuts retrieval tokens 3.5x
Mixedbread's API-only model runs the entire search loop for you; plus Mole, an Apache-2.0 terminal research agent with a hard budget and verified quotes.
Copy markdown
Toast takes over the whole search loop
Give it a query and Toast 1 decomposes it into subqueries, gathers evidence, inspects sources, and hands back curated context — standalone or as a subagent inside a frontier model like GPT-5.6 Sol or Claude Opus 5. It's the retrieval half of an agent, spun out as its own specialized model.
The token math: 3.5x fewer, same answers
On Harvey's LAB firm-knowledge benchmark it matched vanilla agents while burning 3.5x fewer tokens, and hit 70% answer correctness on OfficeQA Pro V2 at about $1.15 a task. If your RAG pipeline is bleeding context window on retrieval, this is the lever to pull.
Pricing, and how to wire it in
API-only, no open weights: $0.30 per million input tokens, $0.72 per million output, cached input at $0.036. Standard queries run $0.016–$0.023 at ~8s; a higher-quality fusion mode is $0.05–$0.07. Call it via Mixedbread's API, the Chat Completions endpoint, OpenCode, or store search with agentic=True.
Elsewhere: Mole does deep research in your terminal, on a leash
Mole is a new Apache-2.0 CLI research agent that reserves every API call against a ledger (measured overshoot: 0%), throws out any claim whose quote isn't verbatim in the source, and keeps local CSVs on your machine — the model only sees aggregates. Install via Homebrew or curl; works with Claude or any OpenAI-compatible endpoint.