Pricing
Govern every AI response.
Kenshiki Labs pricing is usage-based across three deployment tiers — Workshop, Refinery, Clean Room — with the same governance contract running everywhere. Workshop is hosted, usage-only, suitable for board-meeting-ready governed AI within a quarter. Refinery adds a private-deployment platform fee for customers whose data cannot leave their perimeter. Clean Room moves to a software-license model with unlimited in-boundary usage for the highest-assurance deployments. The page also publishes the Kadai model lineup with infrastructure costs and per-token rates so buyers can model total cost of ownership without an internal margin worksheet.
Pick your environment
Pick your environment
Same API contract everywhere. Workshop is usage-only, Refinery adds private-deployment fees, and Clean Room moves to software licensing with unlimited in-boundary usage.
Start here
Workshop
Usage only
No platform fee. Start under $100/month.
Models: Kadai 16B, 42B, 14B + BYOK
Full governance pipeline on shared infrastructure. Bring your own model key or use Kadai.
- Full bounded-synthesis pipeline
- Kadai models on shared L4 GPUs
- BYOK option (GPT, Claude, OpenRouter)
- L1–L3 claim evaluation
- Same API contract as Refinery and Clean Room
Private deployment
Refinery
From $3,500/mo
Platform fee + standard usage rates. Annual commitment.
Models: Kadai 70B Instruct, 235B
Same governance pipeline, private infrastructure. Managed by Kenshiki Labs on your VPC or ours.
- Everything in Workshop, plus:
- Private inference — no public model API
- L4 evaluation (hidden-state probes)
- Chain of custody
- Managed deployment and monitoring
- 99.9% SLA (Kenshiki Labs-managed)
Air-gapped
Clean Room
From $50K/year
Software license + hardware BOM published. Contract.
Models: Kadai 400B, 670B
Disconnected operation on customer premises. Hardware root of trust. Signed attestation on every output.
- Everything in Refinery, plus:
- Fully disconnected — no external dependency
- Hardware root of trust
- Signed attestation chain
- Independently verifiable by third parties
- Unlimited usage within deployment
Compare the assurance boundary
Compare the assurance boundary
The easiest way to buy Kenshiki Labs is to choose the environment first. The API contract stays the same. What changes is where inference runs, where the evidence boundary lives, and how much runtime evidence and attestation you get about what happened inside the system.
| Dimension | Workshop | Refinery | Clean Room |
|---|---|---|---|
| Where model runs | Shared Kadai or your existing public-model provider | Private inference in your VPC or Kenshiki Labs-managed private deployment | Customer premises, disconnected if required |
| Where data lives | Kenshiki Labs-managed cloud boundary with governed retrieval and exportability | Private deployment boundary with governed data sources under local control | Customer premises only |
| Audit / assurance boundary | Full pipeline audit with L1-L3 evaluation | Private inference, local telemetry, L4 evaluation, and chain of custody | Signed attestation chain anchored to hardware trust |
| Best for | Pilots, teams evaluating governed synthesis, and BYOK onramps | Enterprise production workloads that need private deployment and managed operations | High-assurance or disconnected programs that cannot depend on external connectivity |
| Starting price | Usage only, under $100/month to start | From $3,500/month plus usage | From $50K/year software license |
Two products. One contract.
Two products. One contract.
Use them together or independently. Workshop and Refinery meter Kura and Kadai directly. Clean Room keeps the same contract but moves to software licensing rather than API metering.
Reasoning API
Kadai
Governed inference. Question in, bounded answer out with claim-level evaluation.
Includes up to 8K governed input tokens after compilation. Additional governed context: $0.01 per extra 8K. Plus token pricing per Kadai model. BYOK: $0.015/request.
Every governed request includes
- Prompt Compiler — CFPO rewrite, 5 passes
- Claim Ledger — L1–L3 evaluation (L1–L4 in Refinery/Clean Room)
- Boundary Gate — classification and output state assignment
- Per-claim audit trail
- Telemetry headers
Governed request boundary
- One end-to-end Kadai call through Compiler, retrieval, Ledger evaluation, and Boundary Gate
- Base governance fee includes up to 8K governed input tokens after compilation
- Governed input means the compiled prompt plus admissible inline and retrieved evidence
- Model output tokens are billed under model inference, not governance
BYOK support boundary
- Kenshiki Labs operates the governance pipeline around BYOK requests: compilation, retrieval scoping, Ledger evaluation, Gate decisions, and audit records
- Kenshiki Labs does not control the uptime, latency, output stability, or semantic behavior of the upstream provider
- Hosted Kadai is the primary path for stronger service commitments and more stable evaluation behavior
Evidence Store
Kura
Governed RAG. Ingest sources, preserve provenance, retrieve with access control.
| Component | Unit | Rate |
|---|---|---|
| Ingestion | per billable page | $0.05 |
| Storage | per GB/mo | $0.50 |
| Retrieval | per query | $0.005 |
| Embedding refresh | per billable page | $0.05 |
Includes
- SIRE tagging and Crosswalk mapping
- OpenFGA/ReBAC per-caller evidence scoping
- SHA-256 hashing, HMAC-SHA-256 watermarks
- REST API access
- Data export access
Working billing definitions
- Billable page
- One normalized page-equivalent processed by the ingestion pipeline. PDFs bill by rendered page, slide decks by slide, and structurally complex files normalize into page-equivalents before billing.
- Billing preview
- The dashboard should show the billable page-equivalent count before ingestion is finalized, so scans and oversized tables do not surprise the buyer.
- Retrieval query
- One logical governed retrieval call with bounded fan-out. Internal retries, reranking, and Ledger-side checks are not billed as separate queries.
What a real month looks like
What a real month looks like
Illustrative examples — your costs depend on corpus size, request volume, and model choice.
Pilot
Workshop$99/mo
Mid-size team evaluating governed synthesis. 500 source pages ingested, 5K retrieval queries, 2K governed requests using Kadai 14B Instruct, average 2K input / 500 output tokens per request.
| Component | Usage | Cost |
|---|---|---|
| Kura ingestion | 500 pages | $25 |
| Kura storage | ~2 GB | $1 |
| Kura retrieval | 5,000 queries | $25 |
| Governance fee | 2,000 requests | $40 |
| Kadai 14B tokens (in) | 4M tokens | $4 |
| Kadai 14B tokens (out) | 1M tokens | $4 |
| Total | $99/mo | |
Illustrative Workshop workload that stays inside the included 8K governed-context boundary.
Enterprise
Refinery$6,150/mo
Enterprise compliance team on private infrastructure. 5,000 source pages ingested, 50K retrieval queries, 20K governed requests using Kadai 70B Instruct.
| Component | Usage | Cost |
|---|---|---|
| Platform fee (annual) | — | $5,000 |
| Kura ingestion | 5,000 pages | $250 |
| Kura storage | ~20 GB | $10 |
| Kura retrieval | 50,000 queries | $250 |
| Governance fee | 20,000 requests | $400 |
| Kadai 70B tokens (in) | 40M tokens | $120 |
| Kadai 70B tokens (out) | 10M tokens | $120 |
| Total | $6,150/mo | |
Private deployment with platform fee plus standard Kura and Kadai usage. Clean Room shifts to software licensing and unlimited in-boundary usage instead of metered API billing.
Under the hood
Under the hood
Every Kadai model is fine-tuned by Kenshiki Labs for bounded synthesis. We show the parameters, the hardware, and the infrastructure cost. Base architectures are disclosed in technical documentation.
| Model | Params | Active | Arch | Hardware | Infra/mo | Input/1K | Output/1K | Environment |
|---|---|---|---|---|---|---|---|---|
| Kadai 16B Instruct | 16B | 2.4B | MoE | 1×L4 | ~$584 | $0.0005 | $0.002 | Workshop |
| Kadai 42B | 42B | 6.6B | MoE | 1×L4 | ~$584 | $0.0008 | $0.003 | Workshop |
| Kadai 14B Instruct | 14B | 14B | Dense | 1×L40S | ~$1,635 | $0.001 | $0.004 | Workshop |
| Kadai 70B Instruct | 70B | 70B | Dense | 4×L4 | ~$3,358 | $0.003 | $0.012 | Refinery |
| Kadai 235B | 235B | 21B | MoE | 1×H100 | ~$8,972 | $0.004 | $0.016 | Refinery |
| Kadai 400B | 400B | 17B | MoE | 8×H100 | ~$40,150 | $0.008 | $0.032 | Clean Room |
| Kadai 670B | 670B | 37B | MoE | 8×H100 | ~$40,150 | $0.010 | $0.040 | Clean Room |
Why show infrastructure cost? Active parameters — not total parameters — drive inference compute. A 235B MoE model activates 21B parameters per pass and runs on a single H100. A dense 70B model fires all 70B and needs four L4s. The 235B is not the cheaper absolute line item here, but it shows why headline parameter count alone is a bad proxy for cost or runtime shape. We show runtime physics, not internal margin worksheets.
Questions
Questions
Can I use Kura without Kadai?
Can I use my own model?
What base models are behind Kadai?
What counts as a page, query, and governed request?
What happens on long-context governed requests?
Why compare a 235B MoE to a 70B dense model?
What SLA applies to BYOK?
Is there a free tier?
What happens if I exceed my spending cap?
Can I export my data?
What's the commitment?
Next Step
Pick the assurance boundary that matches the consequence.
Start in Workshop if you want the fastest path to a governed response. Move to Refinery or Clean Room when the environment itself becomes part of what you have to document and defend.