Introducing Toast 1

Mixedbread Team | August 13, 2026 | ⏱️ 6 min read

We are excited to announce the release of Toast 1, our premier specialized search agent.

Toast 1 delivers frontier-grade search capabilities, matching or even surpassing the performance of Claude Opus 5 and GPT-5.6 Sol. Despite this power, it is designed for extreme efficiency: it is up to $\text{10}\times$ more affordable and $\text{12}\times$ faster than its generalist counterparts.

While it is optimized for use with Mixedbread Search, it remains flexible enough to integrate with any search backend of your choice.

---

The Shift Toward Specialized Agency

Modern frontier models have evolved to handle genuine knowledge work—they can reason through complexities and analyze vast document sets. However, as high-level intelligence becomes more expensive (metered), there is a critical need for specialized agents that provide the same quality at a fraction of the cost.

Relying solely on generalist models for every step of a workflow $\rightarrow$ Using specialized sub-agents for specific tasks.

Toast 1 can operate in two ways:

1. As a standalone retrieval agent.
2. As a sub-agent managed by a larger frontier model.

The Search Loop

Toast 1 completely manages the retrieval process. Instead of a simple query, it executes a sophisticated loop:

By offloading this "grunt work," your primary model can reserve its compute and context window for what it does best: reasoning, acting, and synthesizing the final response.

Example Query: "How did the employment rate change in retail compares to the healthcare sector?"

Performance Trace:

• Tool Calls: 16
• Rounds: 3
• Latency: 5.33s

[
  {"step": 1, "action": "decompose", "sub_queries": ["retail employment trends", "healthcare employment trends"]},
  {"step": 2, "action": "grep", "pattern": "employment_rate_2024"},
  {"step": 3, "action": "curate", "status": "success"}
]

---

Pareto Optimal Search

Specializing agentic labor doesn't just lower costs; it often improves the final outcome. Toast 1 establishes a new Pareto frontier for agentic workloads, optimizing the trade-off between cost per task and speed per task.

Financial Analysis: OfficeQA Pro V2

Using the OfficeQA Pro V2 benchmark from Databricks (which tests 90 complex enterprise financial questions), we saw a massive leap in efficiency.

When GPT-5.6 Sol utilized Toast 1 as a sub-agent via Codex, it achieved:

• Correctness: 70%
• Cost: \approx \1.15$ per task

This is the highest score reported in the OfficeQA v2 release. For comparison:

• Claude Fable 5 (via Databricks Genie): 60% correctness at \approx \4.00$ per task.
• GPT-5.6 Sol (without Toast 1): 33% correctness.

The secret lies in the economics of evidence. Toast 1 creates high-quality, token-efficient "evidence packages," ensuring the reasoning model has exactly what it needs without wasting resources.

---

Legal Agentic Benchmark - Firm Knowledge

The Harvey LAB Law Firm Knowledge benchmark tests an agent's ability to navigate massive, noisy institutional legal data. In this domain, precision is everything—small details distinguish one precedent from another.

We tested three configurations on a subset of 33 tasks. The results showed that while the final score remained identical, the token efficiency varied wildly.

Token Consumption Comparison

Key Takeaway: By implementing Toast 1, we achieved the same performance using $\approx 3.5\times$ fewer tokens.

How Toast 1 Optimizes Legal Search:

• Reduces false positives in noisy datasets.
• Minimizes redundant agent loop iterations.
• Streamlines the evidence gathering process.
• Maintains high-fidelity retrieval of specific legal precedents.