DevDay brought GPT-6.1 Sol and dots. I still do not have 6.1 Sol in my own account, so I cannot tell you how it feels in my work yet. What I can inspect today is OpenJEV: a public route to TypeSafe's Jev model, built around the small decisions applications make all day.
The appeal is concrete. Give a system some state, ask a bounded question, receive a typed answer, and decide in your own code what happens next. That is a useful piece of intelligence even if nobody calls it an agent.
What OpenJEV actually offers
OpenJEV's API documentation describes one endpoint for three answer types: choice selects among named options, score rates something on an ordered scale, and noul estimates a yes-or-no proposition. Multiple questions can share the same input state in one request. The service is independent of TypeSafe; it says it provides access to TypeSafe's Jev through OpenRouter.
Think of a support message. Jev can choose the queue, score urgency and estimate whether a refund was requested. The application still owns the consequential step: route it, flag it, or send it to a person. OpenJEV's example response is explicitly illustrative, and its docs warn that a confidence value is a signal rather than a guarantee. That distinction is the whole design discipline: test the awkward cases and set review thresholds from your own data.
Diogo Almeida introduces Jev and its structured-decision approach.
The unusual part is the access model
OpenJEV proposes funding public API access with creator fees from $JEV trading. The site says callers do not need to buy or hold the token. Its public treasury shows fee and estimated call-coverage figures; those are not the same as audited inference spending, a service-level guarantee, or proof that free capacity will remain unlimited. The thesis itself acknowledges variable trading fees and the need to manage capacity as usage grows.
I like the attempt to make the funding loop visible. I would judge it by the boring things over time: useful integrations, error rates, actual model costs, remaining funds, and whether a developer can depend on the endpoint when the trade volume is quiet. An interesting mechanism becomes infrastructure only when people can build on it repeatedly.
DevDay made the contrast sharper
OpenAI introduced dots as always-on agents powered by Astra, while its 6.1 Sol documentation positions that model for complex coding, computer use and professional work. Those announcements are exciting. They also live at a different scale from a single typed decision inside an app. My missing 6.1 Sol access is an account-level observation, not a claim about everybody's rollout or the model's quality.
OpenAI introduces dots, its always-on agent product, at DevDay.
What sticks with me is that the big systems still need good little judgments. A dot may do the visible work, but software around it must decide what to route, what to retry, what to trust and when to ask a human. OpenJEV is a bet that those moments deserve their own tool and their own economics. I want to see whether the public endpoint earns a place in real workflows, one carefully measured decision at a time.
What I verified
On September 29, I checked OpenJEV's API docs and thesis, TypeSafe's documentation, and OpenAI's 6.1 Sol model page. I did not run an authenticated OpenJEV call or benchmark Jev against an LLM. I have not tested 6.1 Sol in my account. The image is an editorial illustration, not a product capture.