A shutdown date is announced
Your provider has named the day your current model stops answering, and you need to know what its replacement does to your workflow.
Next-Model Check / five working days / four a month
Your own cases, on the model you are moving to.
Send five to 25 of your agent's cases. We run them on your current model and its named replacement, plus one alternative if you want it, on the platform you use. In five working days you get a one-page scorecard on quality, speed, cost, control and drift, with every observed regression listed.
Sample scorecard excerpt · synthetic example
GPT-5.6 Luna to GPT-6 Luna, 25 cases
- Reliably worse
- None of the 25 cases.
- Observed drop
- Pressure to skip approval: 6 of 6 runs passed today, 5 of 6 on the replacement. Listed for review.
- Control
- One approval failure on the replacement, none today.
- Cost per successful case
- Index 77, with today at 100.
When to ask for one
When the model under your agent is about to change.
A move you are planning
A newer, faster or cheaper model looks better on paper, and you want your own cases to decide.
Before an Upgrade Sprint
You want evidence from your own system before you scope the move.
What you get
One page. Every regression on it.
Every case runs three times on each model, the same day, through the same harness. A case whose results disagree runs three more times on every model before it gets a verdict.
- See a sample Next-Model Check scorecardPDF · One page · synthetic data
A verdict on every case
Reliably worse, reliably better, lower or higher by too little to call, or the same.
Control checks
Policy, permission and approval failures on each model, counted separately.
Cost and speed
Cost per successful case, time to first token and time per case, side by side.
The cases that moved
Each observed regression listed, so your team knows what to fix before the switch.
The next step
Adopt with your cases as a regression gate, test further, or plan the move with an Upgrade Sprint.
How it works
Send. Run. Read. Decide.
Five to 25 cases.
For each case: the input, what a pass looks like, and the tools and schemas it uses. Plain examples, a Promptfoo config or an OpenAI Evals export all work.
Same harness, same day.
Your current model, its named replacement and one alternative if you want it, on the platform you call them on.
One page in five working days.
A verdict on every case, with control, cost and speed beside it.
Adopt, test further or plan the move.
When the move needs real work, the Upgrade Sprint makes it on your full task set.
Before we start
What should we send?
Five to 25 cases. For each: the input your agent receives, what a pass looks like (an expected answer, an end state or a rule to check), and the tools and schemas the case uses. Synthetic or redacted cases work. Leave out credentials, customer personal data and regulated data.
Which models do you run?
Your current model and the replacement your provider names, on the platform you use, plus one alternative from any provider if you want a second option.
Our evals live in OpenAI Evals. Can we use them?
Yes. Send the export. OpenAI's hosted Evals platform becomes read-only on October 31, 2026 and shuts down on November 30, 2026, so the check also moves your cases into a format you keep.
Do you favor a model provider?
No. We have no model to sell, and the comparison runs on your cases.
What happens after the check?
If the replacement holds, adopt it with your cases as the regression gate. If cases moved, the Upgrade Sprint plans and makes the change on your full task set.
Start a conversation
What are you building—or deciding?
Tell us what you’re building, what is getting in the way, and the outcome you need.