Jev and ChatGPT Sites: Costs, Uses, and Limitations

Jev is presented as a fast and cost-effective decision engine focused on routing and classification, while OpenAI leverages ChatGPT Sites to create customized internal tools. Claire details the DevDay announcements, the addition of vision to the Decisions API, and the pricing of models like GPT-6.1 Sol and Astra.

Model Pricing and Vision Decisions Change the Equation

GPT-6.1 Sol is offered at $2 per million input tokens and $10 per million output tokens, compared to $10 and $50 for Astra. Claire continues to use Astra for her personal preferences but believes Sol deserves consideration for everyday work. The addition of vision to the Decisions API enables new types of workflows: Claire has used this feature to analyze 100 video images and identify actionable elements.

Jev Preferred for Quick Decisions in a Bounded Action Space

Jev is presented as a decision engine rather than a chatbot, transforming unstructured inputs into structured outputs such as scores, classifications, probabilities, or function calls. Its speed and low cost are highlighted, with a mental model akin to an if/else instruction, where a traditional program would use a condition or a switch. Developers determine the various path options, and then Jev selects the appropriate solution from a natural language prompt. This operation is suitable when the set of possible actions is clearly defined; for tasks such as creative reasoning, open exploration, brainstorming, or image interpretation, LLMs continue to be favored.

Performance and Cost: Trial Numbers and Failure Benchmarks

John Lindquist reports spending 73 cents for 23 development executions with Jev. Claire states she was able to process five gigabytes of JSON data for a cost of 40 cents. In a performance test on a chess game, Jev reviewed the entire game in less than a second, with a speed ten times faster and a cost four times lower than that of a low-reasoning LLM, according to John. He believes a decision model can outperform a generative model when the number of possible actions is limited, and brute-force approaches become feasible as long as each choice is made quickly and at low cost.

Developer Use Cases: Routing, Omnibar, and Layered Classifications

Eight demonstrations of Jev were shown, illustrating scenarios such as a real-time voice assistant, duplicate removal in datasets, application routing, chess game analysis, coordination among multiple agents, and a live presentation coach. Routing is cited as an immediately relevant use case: a simple natural language prompt is enough to identify the right tool, deduce the action to be taken, and launch the corresponding function. The examples of task lists and omnibar illustrate this flexible command layer. For accuracy, John Lindquist favors layered classification, chaining several low-cost steps and combining the results, rather than seeking a single perfect decision.

Sites at OpenAI: Customization through Connectors and Production Infrastructure

Kath Korevec explains that she has been using ChatGPT Sites internally at OpenAI for over a year to transform data from Slack, Notion, and calendars into customized tools, including an incident command center that adapts to each visitor. The Insights plugin allows for automatic display customization, for example by inferring a user's team through their access to channels, while connector authentication manages personalization without dedicated logic. Codex can infer the appropriate plugins from mentions in the prompt, and Kath notes that it often selects correctly without explicit instruction. The Sites integrate D1 storage, R2 buckets, MCP plugin hosting, and co-editing, and are used for production tasks, such as creating slides for the DevDay keynote, although many still consider them a prototyping tool.

Ephemeral Software, Automated Music, and Usage Safeguards

Kath Korevec believes that temporary software is now worth creating, citing the development of a site for a business trip, used for a week and then abandoned. She has automated a music discovery flow that retrieves popular playlists from Reddit every Monday to create and play a Spotify playlist, discovering a song she enjoyed in the process. Skills allow for extending an application without access to the source code: Kath has published a dungeon skill defining dimensions and controls, installable by anyone to generate a room via Astra and add it to a shared crawler. She believes that faster models encourage continuous experimentation; Claire mentions an ultra-fast session with Astra where queries like "jellyfish belly" and "add a chicken" did not interrupt the creative flow. Kath specifies that AI can assist with research and writing, but sending messages remains under her control.

DevDay: Dots and Spaces Tested, Google's Position Questioned

Claire previewed several DevDay announcements, evaluating Dots, Spaces, and Sites, and using the Decisions API to choose podcast thumbnails, then Astra to build a collaborative sketchboard and an interactive 3D world. She finds Dots more capable than they appear, citing examples such as purchases, code generation, and detecting a conflict in her children's schedules, but notes that the relationship between Dot threads, ChatGPT, and Codex remains confusing, leading her to defer her judgment. She considers ChatGPT Spaces a potentially underrated announcement, offering a shared space with access controls for documents and slides. Having developed a similar tool called Claire’s Notebook, she sees OpenAI's version as a serious competitor to Google. For businesses, she presents Sites with grouped plugins as a solution to the proliferation of internal tools, allowing for encapsulating a connector like Snowflake while preserving permissions, so that only authorized employees can access relevant information.