Toast 1

Toast 1 is a hosted text model from Mixedbread that is specialized for multistep search and evidence synthesis. Released on August 13, 2026, it can run as a standalone retrieval AI agent or as a search subagent for a more general model. Mixedbread describes its job as decomposing a question into subqueries, gathering and inspecting evidence, and returning a curated context package for downstream reasoning.^[1][2][11]

Toast 1 is accessed through Mixedbread's API under the model ID toast-1. It is distinct from Mixedbread Search, the company's managed indexing and information retrieval service, and from toast-harness, a public implementation of a compatible search-agent loop. Mixedbread says Toast 1 is most deeply integrated with its own search service but can use another retrieval backend exposed through function calling.^[1][2][5]

Scroll sideways for more →

Release and role

Mixedbread introduced Toast 1 as its first specialized search agent. The release article positions specialization as a way to assign evidence gathering to a smaller, purpose-built component while reserving a frontier model's context and computation for analysis and final-answer production. A launch demonstration shows one employment-comparison query using 16 tool calls in three rounds over 5.33 seconds, but this is a single vendor-selected trace rather than a latency benchmark.^[1]

The intended workflow has four broad stages. Toast 1 first interprets the initial question and produces narrower searches. It then retrieves candidate material, inspects the sources, and revises its search when necessary. Finally, it selects and organizes the evidence returned to the caller. This is a product-level description of behavior. Mixedbread has not published enough architectural detail to determine how planning, routing, retrieval policy, or evidence selection is implemented inside the model.^[1][2]

The model can also answer directly. Mixedbread's Chat Completions documentation describes a single request in which Toast 1 plans Store searches, inspects returned evidence, and generates a grounded response. A caller that wants a general model to do the final reasoning can instead expose Toast 1 as a retrieval subagent and pass its evidence package onward.^[1][3]

API and retrieval interface

Toast 1 is served through an OpenAI Chat Completions-compatible endpoint at https://api.mixedbread.com/v1. The model accepts text and returns text. Its 131,000-token context window counts messages, tool definitions, tool results, and generated output together, and the maximum generated output is 4,000 tokens. Mixedbread's Chat Completions service defaults to toast-1 if a request omits the model field.^[2][3]

The documented tool surface includes ordinary function calling and five hosted Store tools:^[2][3]

• store_search performs semantic search across one or more Stores.
• store_grep performs exact-text or regular-expression matching.
• store_list_chunks supports filtered, sorted chunk inspection.
• store_metadata_facets exposes metadata fields and representative values.
• list_stores discovers Stores available to the API key.

Hosted retrieval is opt-in. A completion that declares no Store tool performs no Store retrieval. Applications can require a particular tool with tool_choice, cap hosted retrieval operations with max_tool_calls, and request selected chunk payloads through include. Responses can expose hosted tool-call records and planning narration in Mixedbread-specific fields.^[3]

Mixedbread also documents previous_completion_id for continuing a stored completion. Setting store to false enables zero data retention for completion content and prevents creation of a retrievable completion, although Mixedbread says it still records operational model and token metadata. This distinction matters for applications searching confidential document collections.^[3]

Mixedbread says any retrieval backend can be connected through function calling. The launch evaluations, however, primarily exercise Toast 1 with Mixedbread Search or within Mixedbread's own integration. The public launch material does not provide a set of disaggregated third-party-backend results that would establish equivalent quality across different indexes and search engines.^[1][2]

Public harness and reproducibility

Mixedbread released toast-harness as an Apache-2.0 Python package. It provides an agent loop and tools for corpus search, exact matching, chunk retrieval, document reading, filtering, context pruning, and final ranking. It also records token use and rounds executed. The harness can run synchronously or asynchronously, and its asynchronous implementation supports parallel tool calls and terminal completion, failure, and cancellation events.^[5]

The repository does not contain Toast 1 model weights. Instead, the caller injects a generation function that returns a Chat Completions-shaped response. Retrieval is also injected: it can use a Mixedbread Store through an API key or another client that implements the expected interface. The Apache-2.0 license therefore applies to the harness, not to an undisclosed Toast 1 checkpoint.^[5]

The harness attempts exact token accounting from a policy tokenizer and fails by default if that tokenizer cannot be loaded. An optional setting permits an approximate character-based counter. This distinction is important because prompt truncation, cost, and the number of usable retrieval rounds can all depend on the selected counter.^[5]

A supplied BrowseComp-Plus runner accepts query text and relevant-document identifiers, records retrieved rankings, and scores nDCG@10 and recall@10. These tools make parts of the retrieval loop inspectable, but they do not by themselves reproduce the complete launch evaluation. Reproduction also requires the evaluated model endpoint, exact corpus and queries, retrieval configuration, sampling settings, prices, and any model-pairing logic.^[5][8]

Evaluation

Mixedbread evaluated Toast 1 on enterprise document reasoning, legal institutional search, and dedicated retrieval benchmarks. The company explicitly labels the Toast 1 OfficeQA run as its own, separate from figures published by Databricks. The results below are therefore vendor-reported unless otherwise stated.^[1]

OfficeQA Pro V2

OfficeQA Pro V2 is a Databricks benchmark of 90 questions grounded in about 1,400 US Treasury PDFs comprising roughly 120,000 pages. Tasks require evidence retrieval, document parsing, and analytical reasoning, and many require information from several source documents. Databricks released the benchmark, parsed corpus, and evaluation code on August 6, 2026.^[6]

In Databricks' provider-harness evaluation, GPT-5.6 Sol running in Codex scored 33.3 percent at about $4.70 per rollout. Databricks reported a strongest Genie configuration of 60 percent. Mixedbread later ran GPT-5.6 Sol in Codex with Toast 1 available as a subagent and reported 70 percent correctness at approximately $1.15 per task. The chart caption rounds the Toast configuration to about $1.20. That Toast result was not part of Databricks' original comparison and had not been independently replicated in the sources available at launch.^[1][6]

The benchmark's exact-match scoring and specialized historical financial corpus constrain what the number means. Databricks identifies document parsing, changing accounting conventions, temporal reconciliation, and entity scope as persistent failure modes. A score on this benchmark does not by itself establish accuracy on unrelated enterprise collections.^[6]

Harvey LAB Law Firm Knowledge

Harvey's Law Firm Knowledge environment is a synthetic law firm with 9,288 files, about 108 million tokens, 266 matters, and 250 search-and-reasoning tasks. Each task runs against the persistent corpus and is graded by an LLM judge using task-specific rubric criteria. Harvey describes incomplete and non-exhaustive search as a central failure mode in its own baselines.^[7]

Mixedbread selected 33 of the 250 tasks for a controlled comparison. It reported that a vanilla filesystem agent, an agent using Mixedbread Search, and an agent using both Mixedbread Search and Toast 1 all received the same task score of 55. Across the selected tasks, the vanilla setup used 80.6 million tokens and averaged 21.7 turns per task. Mixedbread Search used 47 million tokens and 14.6 turns, while the Toast 1 configuration used 23 million tokens and 11.2 turns.^[1]

From those totals, Mixedbread described the Toast configuration as using 3.5 times fewer tokens than the vanilla agent and reducing cost by more than 60 percent without changing the task score. The result concerns a randomly selected 33-task subset and a vendor-run stack. It should not be read as a full-benchmark result or as evidence that every legal search workload will preserve answer quality at the same reduction.^[1][7]

Retrieval sweeps

Mixedbread also reports experiments on BrowseComp-Plus, OfficeQA Pro, and LongSeal. For the plotted sweeps, it gives a standard Toast 1 configuration a median latency of eight seconds and an estimated cost of $0.016 to $0.023 per query. Its highest-quality fusion configuration is reported at an eleven-second median and $0.05 to $0.07 per query. Mixedbread further claims that systems reaching similar quality in its comparison cost 7 to 11 times more and took from 20 seconds to four minutes. The launch page does not expose a full text table of model scores and every configuration, so these comparisons remain vendor claims rather than independently auditable rankings.^[1]

BrowseComp-Plus is designed to make retrieval comparisons more controlled by replacing the live web with a fixed corpus of about 100,000 human-verified documents. Its evaluation separates end-to-end answer accuracy from evidence recall, search-call count, calibration, and retriever-only ranking metrics.^[8] LongSeal, part of the SealQA benchmark, places a useful document among many distracting or misleading documents to test long-context evidence identification.^[9] These tasks probe search behavior under particular corpora and scoring rules; they do not measure all dimensions of factuality, safety, or production reliability.

Pricing and availability

Toast 1 became available through the Mixedbread API at release. Mixedbread listed a 40 percent launch discount with the following token prices:^[1][4]

Scroll sideways for more →

Cache writes were free at launch. Retrieval is billed separately from model tokens. On August 14, 2026, the pricing page listed a special Toast 1 semantic-search rate of $1 per 1,000 queries without reranking and $1.50 per 1,000 with reranking. These are dated launch prices rather than permanent specifications.^[4]

Disclosure status and limitations

Mixedbread's launch article and model page do not disclose Toast 1's parameter count, architecture, base model, tokenizer identity, training hardware, training compute, optimization method, post-training recipe, or training-data composition. They also do not publish data-source counts, licenses, geographic or language distribution, or a training corpus manifest.^[1][2]

As of August 14, 2026, Toast 1 did not appear among the 12 public model repositories on Mixedbread's Hugging Face organization page.^[10] No Toast 1 checkpoint, formal model card, training code, arXiv paper, or standalone technical report accompanied the API release. The public harness is useful implementation evidence, but it cannot answer questions about the hosted model's internal architecture or reproduce its training.^[1][5][10]

The release materials also do not contain a Toast-specific system card, risk taxonomy, prompt-injection evaluation, citation-faithfulness score, privacy evaluation, red-team report, or misuse analysis. Retrieval agents can surface incomplete, stale, adversarial, or misleading material, and the upstream benchmarks document difficulty with distractors, ambiguous scope, historical formats, and exhaustive enumeration.^[6][7][9]

Practical limits include text-only input and output, a 4,000-token output ceiling, and a context budget shared by messages, tool definitions, tool results, and generated text. Applications must explicitly enable and scope Store tools, decide whether to retain completions, and validate performance on their own corpus. Because the published Toast results were run by Mixedbread and the model had been public for only one day at this cutoff, broad claims of universal frontier quality, a fixed speed multiplier, or a fixed cost advantage are not independently established.^[1][2][3]

References

1. ^Mixedbread Team. *Introducing Toast 1*. August 13, 2026. mixedbread.com/...toast-1
2. ^Mixedbread. *Models: Toast 1*. Accessed August 14, 2026. mixedbread.com/...models
3. ^Mixedbread. *Chat Completions*. Accessed August 14, 2026. mixedbread.com/...chat-completions
4. ^Mixedbread. *Pricing*. Accessed August 14, 2026. mixedbread.com/pricing
5. ^Mixedbread AI. *toast-harness*. GitHub repository, accessed August 14, 2026. github.com/...toast-harness
6. ^Databricks AI Research Team. *Introducing OfficeQA Pro V2: A New Benchmark for Enterprise Grounded-Reasoning*. August 6, 2026. databricks.com/...rk-enterprise-grounded-reasoning
7. ^Julio Pereyra, Niko Grupen, and Engram. *LAB: Law Firm Knowledge*. Harvey, August 7, 2026. harvey.ai/...legal-agent-bench-law-firm-knowledge
8. ^Zijian Chen et al. *BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agent*. arXiv:2508.06600. arxiv.org/...2508.06600
9. ^Thinh Pham et al. *SealQA: Raising the Bar for Reasoning in Search-Augmented Language Models*. arXiv:2506.01062, ICLR 2026. arxiv.org/...2506.01062
10. ^Mixedbread. *Public models*. Hugging Face organization page, accessed August 14, 2026. huggingface.co/...models
11. ^Mixedbread. *Changelog*. Accessed August 14, 2026. mixedbread.com/...changelog