Benchmarks you can run. Every number from a live request.

Two live benchmarks from Labs. Decision models answer typed questions with probabilities, scored against known answers. Light assistants prepare the same operational briefs, including a hotel room that has been repaired but not yet inspected. Nothing is filled in before you run it.

01 · Decision models

Two decision models. One benchmark, run live.

TypeSafe’s Jev and OpenAI’s Decisions API both answer typed questions (yes or no, a choice, a score) with probabilities rather than prose. Here they answer the same everyday decisions, scored against known answers. Image cases go to the Decisions API only: Jev reads text and JSON.

Why probabilities matter

Jev TypeSafe · jev-latest

Text cases
—
Image cases
Text only
Typical time
—
Sure of the right answer
—

Decisions API OpenAI · gpt-6-luna

Text cases
—
Image cases
—
Typical time
—
Sure of the right answer
—

Not run yet.

12 cases, 35 questions, live. Each model is timed from our server, so the times include the network. One run is a small sample: a demonstration, not a leaderboard.

The playground. Ask both models your own question.

Any text, and an image if you like. Jev answers from the text; the Decisions API sees the text and the image.

Try an example:
What kind of answer?
The options

What you enter is sent to TypeSafe and OpenAI to answer, and isn’t stored by us. Leave out anything confidential. Privacy

Jev TypeSafe · jev-latest

The answer appears here.

Decisions API OpenAI · gpt-6-luna

The answer appears here.

02 · Operational assistants

Two assistants. The same operational brief.

OpenAI’s GPT-6 Luna and Anthropic’s Haiku 5.5 prepare summaries, next steps and draft replies from the same facts. It opens on the hotel room: the shower is fixed, but housekeeping hasn’t inspected it, so it isn’t ready to sell. Six synthetic cases, no scores until you run them.

Six operational briefs. Two assistants.

The same task, source facts and output budget for both providers. GPT-6 Luna: reasoning none. Haiku 5.5: effort low.

OpenAI GPT-6 Luna

Cases answered
0/6
Checks passed
—
Median elapsed
—

Anthropic Haiku 5.5

Cases answered
0/6
Checks passed
—
Median elapsed
—

No runs yet. Every result below comes from a live request.

Facilities · Synthetic case

Repair complete, room awaiting inspection

Prepare a handover for housekeeping. Distinguish completed maintenance from permission to return the room to service.

The shower repair and leak test are complete. Housekeeping has not inspected the room. Prepare the handover only; do not update its availability.
reference
WO-58213
room
214
repair status
Complete
room status
Awaiting housekeeping inspection
What the checks measure

Expected route: Prepare for review. Reference: WO-58213. Required missing context: none. Selected facts must match their sources. Actions stay with a person. The models do not receive these expected answers.

OpenAI

GPT-6 Luna

No result

GPT-6 Luna · reasoning none

Run a case to see the response, rule checks and measured time.

Anthropic

Haiku 5.5

No result

Haiku 5.5 · effort low

Run a case to see the response, rule checks and measured time.

Rule checks cover the structured response, not every claim in its prose. Time includes the provider request; tokens are reported by each provider. Synthetic cases do not establish performance in your business. Both providers process these examples when you run them. Privacy

03 · Method

How they’re run. And what they can’t tell you.

01

Known answers

Every decision case has an agreed right answer, and every operational brief has checks for its route, reference, facts, missing context and who may act. The models never see them.

02

The same inputs

Both sides of a comparison get the same case, the same source facts and the same output budget. Image cases go to the Decisions API only, because Jev reads text and JSON.

03

Settings stated

Haiku 5.5 runs at effort low and GPT-6 Luna with reasoning set to none. Jev and the Decisions API run with their defaults. Times are measured from our server, so they include the network.

Synthetic cases show how a model behaves on work like yours; they don’t establish how it will perform in your business. What you run is sent to the providers to answer and isn’t stored by us. Privacy · Back to Labs

Want to know which decisions could settle themselves?

Bring one workflow to a free conversation, thirty to forty-five minutes. We’ll show you what could be automated, prepared for approval or left with your people.