HN
Today

Introducing Toast 1

Mixedbread AI introduces Toast 1, a specialized AI search agent boasting frontier-level search quality at a fraction of the cost and speed of top-tier models like GPT-5.6 Sol and Claude Opus 5. This new agent redefines the economics of evidence gathering for complex tasks, freeing up more expensive frontier models for pure reasoning. The Hacker News community is abuzz with the potential for democratizing high-performance AI search and driving down operational costs in agentic workflows.

10
Score
0
Comments
#3
Highest Rank
3h
on Front Page
First Seen
Aug 14, 3:00 PM
Last Seen
Aug 14, 5:00 PM
Rank Over Time
735

The Lowdown

Mixedbread AI has launched Toast 1, a specialized search agent designed to deliver 'frontier search quality' while being significantly more cost-effective and faster than leading generalist large language models such as Claude Opus 5 and GPT-5.6 Sol. This agent aims to address the rising costs of advanced AI by offering a highly optimized solution for complex knowledge retrieval tasks, allowing more expensive models to focus solely on reasoning.

  • Agentic Search: Toast 1 operates as an autonomous retrieval agent, capable of decomposing initial queries into subqueries, gathering evidence, inspecting sources, and curating relevant context. This specialization allows it to fully manage the search loop, enabling larger frontier models to concentrate their resources on high-level reasoning and task completion.
  • Performance Benchmarks: It establishes a new Pareto frontier in agentic workloads. On Databricks' OfficeQA Pro V2 benchmark, GPT-5.6 Sol with Toast 1 achieved 70% answer correctness at approximately $1.15 per task, outperforming previous state-of-the-art systems. For legal tasks using Harvey LAB's Firm Knowledge benchmark, Toast 1 maintained answer quality while reducing token consumption by 3.5 times and cutting agent turns by half.
  • Cost and Speed Efficiency: Toast 1 is touted as being up to 10x cheaper and 12x faster than comparable frontier models for deep search, with a standard run costing approximately $0.016-$0.023 per query and an eight-second median latency. This represents a 7-11x cost reduction compared to other high-performance retrieval agents.
  • Integration and Availability: Available through the Mixedbread API, Toast 1 is designed to be backend-agnostic, seamlessly integrating with existing retrieval indexes, coding agents, and Mixedbread Stores, offering flexibility to users.

Toast 1 represents a significant step towards more efficient and accessible AI-powered knowledge work, specializing retrieval to enhance the overall performance and cost-effectiveness of AI agentic systems in diverse applications.