ClarityBeta

Product & Strategy

The Product Desk

9 min read · 1,716 words · 3 moves to watch
The Signal

Two OpenAI engineers cloned Jev's decision API in a week, with no new model.

The Decisions API is existing GPT-6 Luna weights, constrained structured outputs and batched questions, with no reasoning step. Matching the calibration is a separate and harder job. In Iozzo's test the original scored 0.948 F1 to Sonnet 5's 0.911 at 1/47th the cost, so run your own question set before treating the two as interchangeable.

In Play

  1. A Week-Old Clone Lacks Calibration

    Latent Space reports that OpenAI built its Decisions API in about a week. Two engineers did it on existing Luna weights after Jev launched a similar product. For you, this means any AI feature whose value is its API shape can be copied and sold at a lab's prices. Vincenzo Iozzo's test of Jev, reported by TLDR IT, shows what doesn't copy: calibrated accuracy. Jev scored 0.948 F1 against 0.911 for Claude Sonnet 5, at 1/47th the cost.

    Ask Clarity
    Sign in to ask more about this

  2. Agents Get In Without Your API

    OpenAI says Computer Use is now 7x faster. Its product leads told Latent Space that agents now debug and retry instead of stalling. In the same news cycle, OpenAI, Perplexity and Meta all shipped agents that start on events rather than prompts. For you, having no API no longer keeps agents out. They can come in through your UI, your accessibility tree, or another company's chat. DoorDash keeps its own assistant talking in iMessage while renting reach through ChatGPT.

    Ask Clarity
    Sign in to ask more about this

  3. Credentials Must Sit Outside the Agent

    The Information reports that a VC tested Instinct's agent with a fake stored card. The agent read the card number and his passwords back in plain text by running JavaScript on the page. Instinct raised at a $10B valuation the same week. For you, an agent's promise about what it won't see is worthless unless your architecture enforces it. DoorDash's gateway, Stripe's single-use cards and Nvidia's off-host monitor all keep credentials out of the agent's reach.

    Ask Clarity
    Sign in to ask more about this

  4. Copilot Puts a Meter on Agents

    Per The Register, Microsoft now bills Copilot's agents (Cowork, Code, Autopilot) and frontier models by usage, on top of the per-user seat. Admins get spending caps and alerts. Enterprise buyers will compare your agent pricing to that split and expect spend controls at launch. Morning Brew points the other way on consumers: Carnival posted a record $8.43B quarter because anxious customers paid for fixed upfront costs.

    Ask Clarity
    Sign in to ask more about this

  5. AI Impact Claims Outrun Measurement

    a16z's State of Markets II finds that nearly 30% of S&P 500 companies claim quantifiable impact from AI. Only about 2% report any tracked metric. When your customers' CFOs cut 2027 budgets, AI features backed only by anecdotes will be the easiest to drop. The same report says only a tiny share of users run agents at any scale.

    Ask Clarity
    Sign in to ask more about this

Deep Dives

  1. OpenAI Copied the API in a Week. Jev's Edge Is the Calibration It Couldn't Copy

    A lab can now give away the interface for free, so the question for your roadmap is which of your AI features rest on something a fast copy leaves out.

    OpenAI's user-ops team moved all of its support-ticket classification onto the Decisions API as soon as it existed. That is the clearest sign of how little OpenAI had to build. In Latent Space's post-keynote interview, API product lead Nikunj Handa…

    3 action items

    ●

  2. Agents Will Reach Your Product Through the Screen, a Trigger, or Someone Else's Chat

    Agents can now get into products that never built an API for them, which turns your accessibility tree, event stream and partner surfaces into roadmap decisions.

    Swyx told Latent Space he runs YouTube's A/B testing and community posts through the UI. YouTube exposes neither by API. OpenAI's own devex team does the same thing. Better Computer Use means having no API no longer keeps agents out…

    3 action items

    ●

  3. 'Won't See Your Card' Is a Policy. 'Can't See It' Is an Architecture

    One public test exposed the gap between what an agent says and what its runtime allows. Three companies have already shipped designs that close it.

    Yuechen Zhao, a partner at Informed Ventures, stored a fake card in Instinct's Vault after the agent told him it wouldn't see card numbers. Then he asked it to read them back. It did, in plain text, by running JavaScript.…

    3 action items

    ●

The edition continues

Take the signal into the room.

Sign up or log in to read all 3 deep dives in full, plus the final take.

Read the full edition

Continue with LinkedIn