Specialised Search Agents: How a Small Model Cuts Your Company's AI Bill

Specialised Search Agents: How a Small Model Cuts Your Company's AI Bill

The short answer first: on 13 August 2026 Mixedbread launched a model called Toast 1. It is not a general model you chat with — it is a specialised search agent whose only job is to dig through your documents and come back with evidence. What deserves a business owner's attention is not the product name but what it exposed: a large share of what you pay for AI today goes on the act of searching rather than on thinking, and separating the two cuts cost substantially while improving answer quality at the same time.

What was actually announced

According to the company's official announcement, Toast 1 is a search agent trained for one task: it takes a question, breaks it into subqueries, gathers evidence from your sources, inspects it, then hands over a curated package of the relevant material. It does not write your final answer — it prepares the raw material the large model then reasons over.

The company states it matches or outperforms frontier models on search quality specifically, at up to ten times lower cost and twelve times faster. Those are vendor figures and should be read as such, but the more important claim is architectural: it does not require you to change your setup, and runs over the retrieval index you already have without migrating your data somewhere new.

Announced launch pricing is $0.30 per million input tokens, $0.036 per million cached input tokens, and $0.72 per million output tokens. In practice the company puts a single query at roughly $0.016 to $0.023 with an eight-second median latency, rising in its highest-quality configuration to roughly $0.05 to $0.07 with an eleven-second median.

Why this matters to a business owner: the economics of search

When you build an assistant on top of your company's documents, you assume most of the cost sits in composing the answer. The opposite is true. The model reads dozens of documents to find three useful lines, and you pay for everything it read, not for what it used. That is why the bill grows as your archive grows, why answers slow down, and why the right answer increasingly gets lost in the noise.

What a specialised search agent does is move that heavy lifting to a cheaper, faster model trained precisely for it, then pass a condensed result to the large model. The large model stays in the picture, but it spends its capability on what nothing else does well: understanding, inference, and framing the decision.

The practical rule: do not pay frontier-model prices for work that resembles an archivist's job. Buy thinking from the expensive model and searching from the cheap one.

The numbers from two independent benchmarks

Marketing claims are arguable; published benchmarks are clearer. On OfficeQA Pro v2, released by Databricks to measure answer correctness across ninety questions in complex, realistic enterprise financial situations, GPT-5.6 Sol with Toast 1 available as a subagent reached 70% correctness at roughly $1.15 per task. By comparison the previous best performer reached 60% at roughly $4 per task, while the same model without a search agent managed only 33%. Adding the search layer roughly doubled accuracy and cut cost at once.

On Harvey's law firm knowledge benchmark, across a subset of thirty-three tasks, answer quality stayed constant while token consumption fell from 80.6 million to 47 million by improving search alone, then to 23 million once the specialised search agent was added — finishing in half the number of turns, with a cost reduction of over 60%.

Notice the shared lesson: the model's ability to reason did not change. The quality of the material reaching it did. That is exactly what we see in client projects — the company assistants that fail are rarely short on intelligence; they are short on feeding.

Where this shows up in your Saudi business

• Contracts and agreements. A question like "which of our contracts allows termination on thirty days' notice?" is a search question, not a reasoning question, and today it consumes hours of a legal employee's time.
• Tender documents and RFPs. Extracting qualification requirements, penalties and delivery dates from hundreds of pages before you decide whether to bid at all.
• Invoices and progress claims. Tracing a single cost line across years of documents to find where profit is leaking.
• The HR archive. Leave policies, allowances and residency expiry dates scattered across files and disconnected systems.
• Customer support. An assistant that answers from product manuals and meeting minutes instead of inventing an answer.

In every one of these cases, the real asset you own is not the model — it is the archive. A model can be swapped out in a week; a well-organised archive is built only once.

Three questions to ask before you buy

First: what does a single question cost me, not what does the monthly subscription cost? If your vendor has no numeric answer, you do not know what your bill becomes at scale.

Second: is the search layer separate from the model? Separation means you can change the large model tomorrow without rebuilding everything, and it is the single strongest hedge against provider price swings.

Third: where do my documents live, and do they leave? That is a governance question before it is a technical one, and it bears directly on your obligations under the Personal Data Protection Law if your archive holds customer or employee data.

How we approach it at Origami

We do not sell a particular model, and we do not advise a client to tie their operation to one provider because it made the news this week. What we build is the architecture that keeps the model a replaceable part: an organised knowledge source that you own, an independent search layer, clear permissions defining who sees what, and cost measured per query rather than per monthly invoice.

If your company is sitting on a large archive it has never benefited from, start with the narrowest repetitive and expensive question you have, and measure time and cost before and after. A small, measured step is worth more than a broad AI programme with no indicator attached to it.

Sources

• Official Toast 1 announcement from Mixedbread, 13 August 2026: mixedbread.com
• Official pricing page: mixedbread.com/pricing
• OfficeQA Pro v2 benchmark from Databricks: databricks.com
• Law firm knowledge benchmark from Harvey: harvey.ai