Mixedbread AI
Listed at https://www.mixedbread.com
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | 按量计费，三块：索引（Fast $1.50/百万内容 token；High Quality 含 OCR/转写/摘要/多模态增强 $3/百万）、搜索（Grep $0.10/千次、Semantic $4/千次、Toast 1 $1/千次，重排 +$3.50 或 +$1.50/千次）、存储 $0.50/百万内容 token/月。Toast 1 按 LLM token：输入 $0.50/百万（发布价 $0.30）。 | per million tokens |
| Output | Toast 1 输出 $1.20/百万 LLM token（发布价 $0.72，40% 折扣）；缓存输入 $0.06/百万（发布价 $0.036），缓存写入免费。Semantic 检索重排加价 $3.50/千次。 | per million tokens |
| Cache Read | 缓存输入 token 计费远低于全新输入——Toast 1 缓存 $0.06 vs 全新 $0.50（约 -88%），缓存写入免费。 | Discounted |
🤖 Supported Models (5)
✨ Pros
- ✓Managed retrieval-as-a-service: upload PDFs/images/code/video and search semantically, no embedding model or vector DB to run, 100+ languages
- ✓Runs Modern ColBERT (late-interaction multi-vector retrieval) over 2.5B+ documents at 10k+ queries/sec, #1 on BrowseComp-Plus
- ✓Toast 1 search model (2026-08-13): matches/outperforms Claude Opus 5 and GPT-5.6 Sol on knowledge work, up to 10x cheaper and 12x faster
- ✓Open-source mxbai family with 50M+ downloads: mxbai-embed-large-v1 (1,024-dim) + mxbai-rerank-v3.1-listwise (GPT-5.6 Sol-level ranking, 25-54% lower latency)
- ✓End-to-end High Quality indexing with OCR, transcription, and multimodal enrichment ($3/1M content tokens)
- ✓Bring Your Own Bucket: source content stays in object storage you own; nothing retained on Mixedbread infrastructure
- ✓Starter $5 free credits, no card; Agentic search can stream live trace events
⚠️ Cons
- ×Retrieval is hosted-only — the full Search API (Modern ColBERT/Wholembed/multimodal) is a closed platform; you cannot self-host the whole retrieval stack
- ×No mainland-China direct endpoint; production from China needs a proxy/relay, 150-250ms cross-Pacific latency
- ×Search bills per query (Semantic $4/1K); high-frequency search (millions of queries/month) can outpace flat-rate embedding+vector-DB alternatives
- ×No raw embedding vector API — if you only need 1,024-dim vectors and self-managed indexing, the hosted product costs more; use the open weights or a dedicated embedding API
- ×Toast 1 launch pricing is a temporary 40% discount; list prices are $0.50/$1.20 per 1M (budget at list)
🎯 Best For
Teams that want managed, production-grade RAG: they prefer not to build and run an embedding + vector-database + rerank pipeline, need to ship semantic search over multimodal, real-world documents (PDFs/images/video/tables) quickly, and want 100+ language coverage and retrieval scaled to 2.5B+ documents.
💰 Pricing & Plans
| Component | Mode / Model | Price | Notes |
|---|---|---|---|
| Indexing | Fast | $1.50 / 1M content tokens | Content-aware tokenization |
| Indexing | High Quality | $3.00 / 1M content tokens | OCR, transcription, summaries, multimodal enrichment |
| Search | Grep | $0.10 / 1K queries | Exact & regex matching |
| Search | Semantic | $4.00 / 1K queries | Fine-grained semantic retrieval; +$3.50/K with rerank |
| Search | Toast 1 | $1.00 / 1K queries | Frontier search model; +$1.50/K with rerank |
| Storage | Per-token monthly | $0.50 / 1M content tokens / mo | Keeps indexed data ready for retrieval |
| LLM tokens (Toast 1) | Input | $0.50 / M (launch $0.30) | Specialized search model for knowledge tasks |
| LLM tokens (Toast 1) | Cached input | $0.06 / M (launch $0.036) | Cache writes free |
| LLM tokens (Toast 1) | Output | $1.20 / M (launch $0.72) | 40% off launch pricing |
🔧 API & Developer Experience
- •API Style: REST + OpenAI-compatible Chat Completions surface for the Toast 1 search model. A programmatic /question docs endpoint returns structured answers, and agentic search streams live server-sent trace events.
- •Model Types: One API covers retrieval end to end: Modern ColBERT semantic search, Grep (exact/regex), reranking, chunking, OCR/transcription via High Quality indexing, and the Toast 1 search model for knowledge tasks.
- •Multimodal: Upload PDFs, images, tables, audio, video, code, and documents; Mixedbread handles OCR, transcription, and multimodal enrichment internally across 100+ languages.
- •Storage & Privacy: Bring Your Own Bucket keeps source content in object storage you own; every indexing artifact is persisted to your bucket and nothing is retained on Mixedbread infrastructure.
- •Agentic Search: Agentic mode (streamable since 2026-08-19) lets an agent observe retrieval live, emit reasoning events, and converge on answers across multi-step searches.
- •Integrations: SDKs and APIs for building a Store, uploading files, searching, and reranking; dashboard, Discord, and GitHub community; open mxbai weights on Hugging Face for self-hosted embedding/rerank.
- •Plans & Limits: Starter: $5 credits, 3 users, 10 stores, 100 req/min. Scale: $20/mo + pay-as-you-go with higher limits and priority support. Enterprise: sales-quoted with custom models and self-hosting options.
⚡ Modern ColBERT & the Multi-Vector Era
Mixedbread is the production face of late-interaction, multi-vector retrieval. Its managed Search API runs Modern ColBERT over 2.5B+ documents at 10k+ queries per second, holding the #1 spot on BrowseComp-Plus, so teams get proven LTR quality without building the stack. This matters now because the open-source ecosystem is standardizing the same idea: Sentence Transformers v6.0 added MultiVectorEncoder, which loads PyLate, Stanford-NLP ColBERT, or colpali-engine checkpoints through the familiar API. Where a self-hosted dense setup indexes one vector per document and loses token-level meaning, multi-vector late interaction stores several vectors per document and matches them with a MaxSim score, improving retrieval on long, topic-dense documents. Mixedbread pairs this with a Wholembed v3 query engine (39% lower query-encoding p50) and open mxbai embed/rerank weights.
🌐 Regional Availability & Latency
Mixedbread is a US/Europe platform (legal entity mixedbread ai inc., with engineering in Berlin and San Francisco) serving a global API over Cloudflare-backed edge delivery. For North America and Europe, first-byte latency is low and the managed stack is fully hosted, with no regional self-host variant. From mainland China, direct access is unavailable and production traffic requires a proxy or relay, with typical cross-Pacific first-byte latency of 150-250ms; the index-once-read-many model still works well through a relay once data is ingested. Storage is region-agnostic because of Bring Your Own Bucket, so latency is dominated by where your object store and relay sit. The practical pattern for global RAG is to ingest data near your team or bucket and let the managed layer serve from a single global control plane.