August 2026 Summaries

Mixedbread uses PlanetScale for Postgres Metal as the control-plane database for its AI retrieval platform, while storing customer documents and indexes in S3-compatible object storage, including customer-managed buckets. The control plane manages access permissions, store lifecycles, distributed processing coordination, and usage accounting, requiring low latency, consistency, and reliability without slowing ingestion or search. After operating its own Postgres clusters, Mixedbread adopted PlanetScale following the service’s 2025 general availability to reduce operational responsibilities such as replication, backups, failover, tuning, and networking, while Metal’s local NVMe storage offered more predictable I/O performance. The company reports that its busiest control-plane queries run at sub-millisecond p99 latency, with all major query patterns below 1.5 milliseconds, and that managed capacity scaling and PgBouncer configuration have simplified operations without user-visible downtime. Mixedbread also connects PlanetScale metrics and database guidance to its agent-based observability system, helping engineers trace production issues from affected services to queries and code, and plans to focus further on retrieval accuracy, reliability, and scalability.

Mixedbread has launched Toast 1, a specialized agentic search model designed to decompose queries, retrieve and inspect evidence, and provide curated context to either operate independently or support frontier models as a subagent. The company reports that Toast 1 matches or exceeds frontier retrieval quality while reducing search cost, latency, token use, and agent turns, particularly when paired with Mixedbread Search, although it can also use existing retrieval backends. Reported evaluations include 70% correctness at roughly $1.15 per task on Databricks’ OfficeQA Pro V2 with GPT-5.6 Sol in Codex, and equivalent scores on a 33-task legal knowledge benchmark while using 3.5 times fewer tokens than a vanilla retrieval setup. On deep-search benchmarks, Mixedbread says standard Toast 1 runs cost about $0.016–$0.023 per query with roughly eight-second median latency, compared with 20 seconds to four minutes for comparable frontier-model agents. Toast 1 is available through the Mixedbread API at launch-discounted token pricing, integrates with Chat Completions, coding agents, and Mixedbread Stores, and includes an introductory API credit offer.