Duet Machines
For teams running one AI agent at volume

Move your highest-volume AI agent to a model you own.

About half the cost, proven on your own expert's standard and your business outcomes, improving every week.

Start freeRead the quickstart
platform.duetmachines.com/agents/claims-intakeSynthetic example

Migration

5 of 6 done
Next up · Step 6 of 6 · Under way

Rollout

Split 10%: the candidate serves a random tenth of conversations. Cost and the standard pass; the business outcome is still collecting.

Gates to leave split

2 of 3 passing
Cost per conversationPass

50% cheaper$0.023 against the current model's $0.046

On 236 priced conversations of the candidate and 2,140 of the current model (p < 0.001).

Defect rate (grader)Pass

0.5 pts lower3.8% against the current model's 4.3%

9 of 236 candidate graded conversations against 92 of 2,140 on the current model (p = 0.710).

Business outcome148 of 200 needed

0.7 pts higher33.1% against the current model's 32.4%

49 of 148 candidate conversations against 612 of 1,890 on the current model (p = 0.860). A pass needs 200 conversations per model.

Model spend a month

Same traffic
Today · GPT-5.4$18,400.00
Candidate · measured$9,200.00

$9,200.00 less a month, measured on the split, billed at 50% of GPT-5.4's price.

Review

Claims lead

I just need to know what I owe on the claim.

You owe $1,240 on claim 4471. Can you confirm your date of birth?

Read the balance before confirming identity.

The Overview of an example agent, drawn by the app's own components with synthetic data. Click a step.

How a migration runs

Four steps, each a page of the agent, each ending in evidence you can check.

  1. 1

    Connect

    Send production traffic and the business outcomes it leads to: an SDK, OpenTelemetry, a JSONL import or the gateway itself.

    …/api/gateway/v1Traffic connectedOutcomes mapped
  2. 2

    Define good

    Your expert reviews real conversations, pass or fail with a reason. Graders learn that standard and are measured against it on held-out reviews.

    Promised a payout date the policy does not allow.

  3. 3

    Train and test

    A cheaper open model learns from your own traffic and is checked on tests built from real conversations, confirmed failures and stress tests.

    GPT-5.4✓✓✕✓✓✓✕✓Candidate✓✓✓✓✓✓✓✕
  4. 4

    Roll out with gates

    Shadow first, then a randomized split that grows only while cost, your standard and your outcomes hold.

    1. Shadow3/3 gates
    2. Split 10%2/3 gates
    3. Full

You pay less than today, for the same traffic

Duet Machines earns on the inference its models serve, billed below what your current model charges for the same tokens.

  • A fixed share of today's priceEach answer our model serves is billed at a share of what your current model charges for its tokens, at the public list price or your verified invoice, agreed before anything serves.
  • Only what we serve, once it holdsTraffic moves only while the gates hold on your own conversations. Shadow answers, failovers to your model and throttled requests are never billed.
  • A statement you can checkEvery month: your model's cost for the same tokens, what we metered and the difference, each closed month fixed with a SHA-256.
Billing in the docs

One month at full rollout

Example

What GPT-5.4 would charge for the tokens our model served

$18,400.00
Billed by Duet Machines
$9,200.00
50% of that price, in this example
You keep
$9,200.00
Against what you pay today

Start with a migration report

Connect one agent and we hand back what moving it would take, built from your own traffic.

Ask for a reportWhat it contains
  • Failure taxonomyThe ways your agent fails today, in your expert's categories.
  • Grader validationHow often each grader agrees with your expert, with its interval and sample.
  • CandidatesCheaper models priced on your real tokens, with their test results beside your current model's.
  • RecommendationThe model to move to and its projected monthly cost, or why none qualifies yet.