PRIVATE, ACCOMPANIED PILOT

Your first workflow.
Measured together.

Start with one repeatable text task: summarization, translation, classification, drafting or structured extraction. Batch processing does not require JSON answers. Your operator agrees the request allowance and assigns pilot credit before you begin.

Plan your pilot

This is an invitation-only, accompanied pilot for 2–3 participants. Prepare your task, preferred output, approximate request volume and 20–50 authorized examples. Agree the allowance and review criteria with the person who invited you before submitting. Signing in does not grant credits or an invitation.

Download the pilot brief ↓ · Already invited? Open your console ↗

Choose a model

• Ministral 3 8B — start here for classification and structured extraction. $0.12 input / $0.12 output per million tokens.
• Mistral Small 4 — compare on the same documents when the task needs more capability. $0.12 input / $0.42 output per million tokens.
• Qwen 3.5 9B is experimental. It is outside client extraction pilot allowances. Reasoning can exhaust its generation budget before a final answer.

These are pilot rates. Quality and costs must be checked on your own authorized examples before a commercial commitment.

On our 16 September 2026 36-document synthetic extraction challenge, Ministral returned 32 exact extractions and Mistral Small 33. Errors included amounts and support priority despite valid JSON. This small test is a reason to keep human review, not a production accuracy estimate.

With revised extraction guidance, both models returned 12/12 exact results on a separate, newly written synthetic challenge that day. Different documents and a small sample mean this is not a measured production accuracy gain. The original 36-document results remain unchanged.

What you pay for

1. A quote gives a maximum charge. Requesting a quote does not reserve credit.
2. Submitting reserves that maximum in your workspace.
3. When the batch is settled, complete responses that pass the requested format checks are charged for actual input and output tokens. Reasoning is included in output usage.
4. Failed, refused or incomplete responses have no customer charge. The unused reserve is released.

Format validation does not guarantee factual accuracy. An accepted response with incorrect facts, including valid JSON with a wrong amount or classification, remains billable. Your Python integration and human reviewer must check fields before posting invoices, issuing refunds or changing customer records.

There are no automatic inference retries. A retry needs a new quote and an explicit submission. All admitted requests, including retries, failures and cancelled requests, count towards your pilot request allowance. Cancelling queued work releases its credit reservation.

Live inference uses operator-funded pilot credit. Stripe checkout is test-only, adds fictitious test credit and cannot fund live inference. No real customer payment is collected in this pilot.

Our separate synthetic OCR challenge produced 1/8 entirely exact results and 5/8 correct amount-and-currency results. OCR remains experimental and requires checking the original image. These extraction measurements do not evaluate summaries, translations or code.

Run your first batch

1. Use your approved account to sign in, then verify the workspace name and live balance.
2. Create an application key for your own integration. Keep it on your server.
3. Choose a free-text, Markdown, code or JSON example, or start with one non-sensitive document. Review the quote and submit once.
4. Download both the results and errors. Match each response using custom_id, then compare with independently checked expected values.

The completion target is 24 hours, best effort. Submitted work cannot be recalled. If a batch needs operator review, keep its ID and contact your pilot operator; do not resubmit an uncertain purchase.

Evaluate a real workflow

Begin with 20–50 authorized, anonymized documents for one task. Agree the expected fields before running the batch, count every attempt, and record correctness, critical amount errors, human review time, delay and cost per correct result. Your operator reviews the findings with you before expanding the allowance.

For PDFs and images, the Python example extracts text or runs OCR locally and submits that text. Native vision, handwriting and medical extraction are outside this offer. Low-quality scans can corrupt a number before the model sees it.

Results are retained for seven days. Download what you need and request deletion when finished. Review data handling and provider limitations before uploading source documents.

Download the pilot review template ↓