Mixedbread Toast 1 specialized search agent launch

Toast 1 delivers frontier‑level search at a fraction of the cost

Toast 1, Mixedbread’s first specialized search agent, achieves the same or better answer quality than top‑tier models such as Claude Opus 5 and GPT‑5.6 Sol, while costing up to 10× less and completing queries 12× faster. The agent can be used as a standalone retrieval system or as a sub‑agent for any frontier model, taking full control of the search loop – decomposing queries, gathering evidence, inspecting sources, and curating context before handing the result to the reasoning model.

Cost‑performance breakthrough on enterprise benchmarks

OfficeQA Pro V2

  • 70 % answer correctness at ≈ $1.15 per task – the highest score reported in Databricks’ OfficeQA Pro V2 benchmark.
  • By contrast, Claude Fable 5 on Databricks Genie achieved 60 % correctness at ≈ $4 per task, and GPT‑5.6 Sol without Toast 1 reached only 33 % correctness.
  • The improvement stems from Toast 1’s ability to produce token‑efficient evidence packages, freeing frontier models to focus compute on reasoning.

“GPT‑5.6 Sol with Toast 1 made available as a sub‑agent within Codex reaches 70 % answer correctness at approximately $1.20 per task, above the previous Pareto frontier from the Databricks evaluation.” – Mixedbread blog

Harvey LAB Law Firm Knowledge benchmark

  • Identical task score of 55 across three configurations, but token usage dropped from 80.6 M (vanilla agent) → 47 M (with Mixedbread Search) → 23 M (with Toast 1).
  • Turns per task fell from 21.7 → 14.6 → 11.2, a 3.5× reduction in token consumption and > 60 % cost savings.
  • The benchmark demonstrates that higher‑quality search does not compromise answer quality; it merely makes the process far more efficient.

Performance on deep‑search benchmarks

  • Median latency: 8 seconds per query (standard configuration) and 11 seconds for the highest‑quality fusion mode.
  • Cost per query: $0.016–$0.023 (standard) and $0.05–$0.07 (fusion).
  • Compared to frontier‑model retrieval agents that take 20 seconds to 4 minutes, Toast 1 is 7–11× cheaper and considerably faster.
  • On benchmarks such as BrowseComp Plus, OfficeQA Pro, and LongSeal, Toast 1 matches or approaches the best frontier‑model sweeps while remaining in a lower cost‑latency tier.

Integration flexibility

Backend‑agnostic design

  • Toast 1 is co‑designed with Mixedbread Search but works with any existing retrieval index. No data migration is required.
  • It can be added to existing agentic workflows via the Mixedbread Chat Completions API or the provided golden harness (GitHub: mixedbread-ai/toast-harness).

Coding‑agent support

  • Integration with coding agents is a single command: npx skills add mixedbread-ai/skills.
  • OpenCode integration enables direct use as a sub‑agent in code‑centric pipelines.

Example Python usage with Mixedbread Stores

from mixedbread import Mixedbread
client = Mixedbread()
results = client.stores.search(
    store_identifiers=["legal-documents"],
    query="does the MSA allow assignment on a change of control?",
    search_options={"agentic": True},  # enables Toast 1
)

Pricing and availability

  • Launch pricing (discounted):
    • $0.30 per million input tokens
    • $0.036 per million cached input tokens (cache writes are free)
    • $0.72 per million output tokens
  • Mixedbread Search calls made by Toast 1 are billed at a special rate (see Mixedbread pricing page).
  • An API key with $5 in free credits is available for immediate trial.

Community reaction on Hacker News

  • Users praised the concept of a dedicated search LLM, noting the potential to replace multi‑step Google searches with a single LLM‑driven loop.
  • Several commenters asked for comparisons with existing cloud search agents (Perplexity, Gemini with search, Parallel AI) and expressed interest in open‑weight alternatives.
  • Questions about on‑prem deployment and data ownership were raised, highlighting demand for flexible deployment options.
  • The overall sentiment is positive, with many seeing Toast 1 as a practical step toward cheaper, faster, and higher‑quality agentic retrieval.

What makes Toast 1 distinct?

  1. Specialized architecture – trained specifically for deep search, allowing it to generate concise, high‑quality evidence packages.
  2. Full search loop automation – from query decomposition to evidence curation, the agent removes the need for external orchestration.
  3. Pareto‑frontier improvement – simultaneous gains in cost, latency, and token efficiency across multiple real‑world benchmarks.
  4. Plug‑and‑play compatibility – works with Mixedbread Search for optimal performance but remains agnostic to other retrieval backends.

Future outlook

Toast 1 joins a growing ecosystem of specialized search agents, including SID‑1 and Chroma’s Context‑1. As more enterprises adopt agentic workflows, the demand for cost‑effective, high‑quality retrieval will drive further innovations in dedicated models, tighter integration with retrieval primitives, and broader support for on‑prem and hybrid deployments.

Sources

Related