Reasoning that
shows its work.
BonsAI-3 answers with citations, scores its own confidence, and refuses when the evidence doesn't hold. Built for teams that have to defend the answer — not just ship it.
- 2.4B
- tokens / day
- 99.98%
- uptime, 90d
- 14
- languages
- 0
- training on you
Four things that make a model shippable.
Accuracy demos are easy. BonsAI-3 is built around the parts that decide whether a model survives procurement, legal review, and a bad Tuesday in production.
Grounded by default
Every claim carries a source span you can click. If retrieval doesn't support the answer, the model says so instead of guessing.
0.3% unsupported-claim rate
Calibrated confidence
Per-sentence scores you can threshold in code. Route the uncertain 4% to a human before it ever reaches a customer.
ECE 0.021
An audit trail, not a vibe
Prompt, retrieval set, model version, and output diff are logged for 13 months. Export the whole thread for a review.
13-month retention
Boring to integrate
OpenAI-compatible endpoint, typed SDKs, streaming, and structured JSON output. 128k context across 14 languages.
p50 first token 310ms
Halden & Co: 12,400 contract memos, one reviewer.
A 40-lawyer firm replaced its first-pass research memo workflow with BonsAI-3 grounded on its own precedent library. Same reviewers, roughly seven times the throughput.
- time per memo
- 41 min
- unsupported claims
- 7.4%
- reviewer re-checks
- every paragraph
- time per memo
- 6 min
- unsupported claims
- 0.3%
- reviewer re-checks
- flagged spans only
“We didn't buy a model. We bought a reviewer's second pair of eyes that never gets tired.”Maya Okafor · Head of Legal Ops
Read the full write-up →Ninety seconds, no signup.
Four screens that cover what actually happens when you put this in front of a customer.
Point it at your sources.
Upload PDFs, connect a bucket, or let BonsAI read your Postgres read-replica. No fine-tuning, no 6-week onboarding.
The claim we hear most: “it told us it didn't know.”
Selected deployments. Numbers below come from customer-run evaluations, published in full on request.
“BonsAI cut contract review from 41 minutes to 6 — and every finding links back to a clause. Our reviewers stopped double-checking the model, which was the whole problem.”
“The citation layer is the reason we could ship at all. Compliance signed off in nine days because they could read the evidence themselves.”
“We benchmarked four models on 12,000 support threads. BonsAI was the only one that said 'I don't know' when it didn't — that's why it won.”
Priced per token. No seat tax.
Start free forever. Overage is billed transparently at the rates below — or set a hard cap and we stop.
Free
Kick the tires. No card, no sales call.
100k tokens / month
- ✓ BonsAI-3 Flash
- ✓ Playground + 3 saved prompts
- ✓ Citations on every answer
Builder
Most pickedFor solo builders shipping their first grounded product.
5M tokens / month included
- ✓ BonsAI-3 Flash + Pro
- ✓ 128k context, 14 languages
- ✓ OpenAI-compatible API + SDKs
Scale
Production workloads with budgets you can defend.
100M tokens / month included
- ✓ Everything in Builder
- ✓ Confidence thresholds + human routing
- ✓ 13-month audit trail + exports
Enterprise
Single-tenant VPC, on-prem, or regulated deployment.
Committed volume pricing
- ✓ Everything in Scale
- ✓ VPC peering or on-prem weights
- ✓ HIPAA BAA + custom DPAs
Estimate only — overage billed at $0.60 / 1M (Flash) and $1.80 / 1M (Pro).
The honest answers.
If something here reads like marketing, tell us — we'll rewrite it.
How accurate is BonsAI-3, really?
On our grounded QA benchmark it scores 94.2% exact-match with citations enabled, at a 0.3% unsupported-claim rate. Without grounding, accuracy drops about nine points and we stop calling it reliable.
Do you train on my data?
No — not on Free, Builder, Scale, or Enterprise. Prompts and outputs are retained for 30 days by default for abuse monitoring, then deleted. Zero-retention is available on Scale and Enterprise, and self-hosted weights are available on Enterprise.
What happens when I hit my monthly token limit?
You choose: soft overage billed at $0.60 per million tokens on Flash, or a hard cap that returns a 429 and escalates to your fallback.
Which systems can it read from?
S3 and GCS buckets, Postgres and Snowflake read-replicas, Notion, Confluence, Google Drive, SharePoint, and any REST endpoint you expose.
Can we self-host or run in our own VPC?
Yes. Enterprise supports VPC peering with a dedicated inference endpoint, or delivery of quantized weights for on-prem deployment on your own GPUs.
Is there a free tier?
100,000 tokens a month, forever, no credit card. It's enough to run several hundred real prompts.
Run your first grounded query in 60 seconds.
No credit card. No sales call. Paste a document, ask a hard question, and see whether the answer can prove itself.