ClarityBeta

Product & Strategy

The Product Desk

9 min read · 1,814 words · 3 moves to watch
The Signal

GPT-6 Astra drove Figma and a workflow builder for 105 minutes without a human click.

The headless rewrite sitting in someone's architecture doc was justified by one sentence in a planning meeting: agents can't use our interface. That premise is gone, and what replaces it is arithmetic. Input runs $10 per million tokens against $50 for output, and a session that emits actions and observations continuously is output-heavy by construction. So the cost lever is not the model or the interface. It is which tasks get handed over. Sort the candidate list by how much output a run generates before you sort it by how impressive the demo looks.

In Play

  1. Agents Now Operate Your Product's UI

    Claire Vo's tests of GPT-6 Astra, reported in Lenny's Newsletter, had the model add nodes and rewire logic inside Adio's workflow builder, work inside Figma, and run a 1 hour 45 minute unattended QA session with no human clicks. That undercuts any epic on your plan justified by 'agents can't use our interface.' Astra is priced at $10 per million input tokens and $50 per million output, and computer-use sessions are output-heavy by design.

    Ask Clarity
    Sign in to ask more about this

  2. Five Sources Decide Your Shortlist

    Latent.Space's Frontier AEO Tracker ran six prompt variants across seven search-enabled models over 161 categories. Astra retrieves a median of five sources per query; Anthropic's Fable retrieves fifteen. Only 28 of the 161 categories have one universally dominant pick, so most are still contestable. Broken markdown content negotiation on your docs quietly stops models reading the site at all, and no analytics dashboard reports that failure.

    Ask Clarity
    Sign in to ask more about this

  3. Four Assistants Degraded In One Window

    ChatGPT, Claude, Grok and Gemini all degraded inside the same few-hour window on Thursday, and no common cause was ever identified, per TLDR Fintech; Last Week in AI logged three of them failing that morning from unrelated causes. A second model provider does not cover that failure mode. Banks moved within hours to degraded operating modes, manual fallbacks and locally hosted models — three of those four measures assume the model is simply gone.

    Ask Clarity
    Sign in to ask more about this

  4. A Vendor Harness Added 37 Points

    ARC Prize ran GPT-6 Astra twice, per AI Breakfast: 62.7% on its neutral Standard harness at $26,098, and 99.9% on OpenAI's Provider Adapter harness at $18,817. The higher, cheaper score came from preserving the provider's opaque reasoning state between requests. Every capability number in your model-selection doc needs a field naming the harness that produced it, because staying vendor-agnostic now costs measurable score and dollars.

    Ask Clarity
    Sign in to ask more about this

  5. Configuration Stopped Being A Moat

    OpenAI's Codex now scans a developer's machine, imports their Claude Code setup, and renames CLAUDE.md to AGENTS.md automatically, per Simplifying AI; a companion plugin runs Codex inside Claude Code. Months of accumulated setup transfers in minutes, so 'switching is painful' can no longer carry a retention narrative in your planning doc. The defender's opening is real though: hooks and output styles have no Codex equivalent, and custom subagents need a manual pass.

    Ask Clarity
    Sign in to ask more about this

Deep Dives

  1. The Headless Rewrite Just Lost Its Justification

    One power user's unattended session undercuts an epic many teams still fund, and Stripe's numbers say its adoption came from configuration surfaces, not from a better model.

    Priced against the work it does best An engineer kicks off a 105-minute QA run. The agent emits actions, observations and notes the whole time, because computer-use work is output-heavy by construction. Astra lists at $10 per million input tokens…

    3 action items

    ●

  2. Five Sources Decide Your Category's Shortlist

    The buyer's candidate list is assembled from a handful of retrieved pages, and what most often removes you is a server-configuration failure nobody on the growth team owns.

    Four models, four different channel sizes A developer asks four assistants the same question and gets four differently shaped answers. Anthropic widened retrieval from Opus's 11 sources to Fable's 15. OpenAI went the other way, narrowing from Sol's 9 to…

    3 action items

    ●

  3. Model Portability Now Has A Measurable Price

    Buyers now demand model neutrality, a benchmark rerun shows what neutrality costs in score and dollars, and one Thursday morning showed what it does not buy.

    Supply can be revoked for reasons no contract review would flag Somewhere there is a vendor scorecard for Cursor with rows for uptime and data handling, and no row for who owns the company. OpenAI ended Cursor's model access after…

    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