Mixedbread just officially released Toast 1, their first dedicated AI search agent.
Mixedbread Toast 1 is a dedicated AI search agent engineered for rapid, cost-effective evidence retrieval from vast document stores. It boasts superior efficiency, being 12x faster and 10x cheaper than GPT-5.6 Sol for similar workloads. Toast 1 provides accurate, structured results in about 8 seconds, delivering substantial token savings and improved accuracy, with tasks starting at just
.15.
It isn't another general-purpose LLM trying to write marketing copy or ship boilerplate code. It does exactly one job: surface rock-solid evidence across massive, messy document stores.
Frontier Quality, Fractional Cost
When it comes to retrieval quality, Toast 1 matches or outperforms frontier baselines like Claude Opus 5 and GPT-5.6 Sol. The difference is efficiency. Toast 1 completes the same retrieval workload at 1/10th the cost and 1/12th the runtime of GPT-5.6 Sol.
While most frontier retrieval agents take anywhere from 20 seconds to 4 minutes to parse complex requests, Toast 1 returns structured results in roughly 8 seconds.
Toast 1 manages the entire search loop autonomously: feed it a raw question, and it breaks it down into sub-queries, hits search APIs, verifies source reliability line by line, strips out noise, and packages a dense, clean context block for your primary model. No fragile prompt engineering. No trial-and-error loops.
On the Databricks OfficeQA Pro V2 benchmark (90 enterprise financial scenarios):
- GPT-5.6 Sol + Toast 1 (as a sub-agent on the Codex framework) scored 70% accuracy at $1.15 per task, setting a new high-water mark for both accuracy and cost efficiency.
- The previous top performer, Claude Fable 5 on Databricks Genie, managed 60% at $4.00 per task.
- Standalone GPT-5.6 Sol without Toast 1 only hit 33%.
Real-World Token Savings
In the Harvey LAB legal knowledge benchmark across 33 random tasks, all evaluated setups ultimately reached the same quality score of 55. But the token consumption told a completely different story:
- Basic agent using filesystem search: 80.6 million tokens
- Mixedbread Search: 47 million tokens
- Toast 1 as a search sub-agent: 23 million tokens
Task rounds dropped from 21.7 down to 11.2 per task (a 42% reduction).
Cutting token bloat isn't just about shaving server bills. It frees up the primary model's context window so it can focus on deep reasoning, case analogies, nuance detection, and multi-step argument synthesis—rather than drowning in pages of irrelevant clauses.
True Plug-and-Play Integration
You don't need to rebuild your infrastructure or migrate indexes to use it. Toast 1 is agnostic and works on top of your existing Elasticsearch, Pinecone, or custom vector backends.
Enabling it takes a single line:
“`python
search_options = {"agentic": True}
“`
It ships with two deployment tiers:
- Standard Tier: $0.016 to $0.023 per query | 8s median latency
- High-Fusion Tier: $0.050 to $0.070 per query | 11s median latency
Both tiers match frontier search quality while dramatically cutting latency and operational overhead.
Toast 1 reflects a broader shift toward vertical co-design. Its training data, API protocols, and scheduling logic are natively aligned with Mixedbread’s embedding stack and Silo infrastructure. Alongside systems like SID-1 and Chroma’s Context-1, Toast 1 proves that specialized search agents—not bloated monolithic models—are the real unlock for production-grade AI.
—
*Adapted from the Hacker News post "Introducing Toast 1".*
