A filter on AI and product news for PMs who’d rather build than scroll.

• Anthropic opens its most capable tier to everyone with Claude Fable 5. Fable 5 tops nearly every benchmark Anthropic tested, scoring 80.3% on SWE-Bench Pro to Opus 4.8’s 69.2%. Anthropic is rationing access behind usage credits after June 22, so the model you prototype on this month may not be the one you can afford to ship on.
  
• Cursor’s Design Mode lets you direct coding agents by pointing, drawing, or talking. You can select an element, draw on the page, or describe a change by voice while the agent edits the code underneath. The lesson for PMs is the interaction model: letting users point, draw, and talk protects their flow, so it’s worth asking where your own product still forces everything through a text box.
  
• Microsoft launches its first in-house models, MAI-Code-1-Flash and MAI-Thinking-1. The coding and reasoning models are Microsoft’s first built in-house, and it claims they hit GPT-5.5-level results with roughly 10 times better cost efficiency. Microsoft spent years renting models from OpenAI, so owning both the model and the enterprise desktop is how default PM tooling gets decided without anyone making a decision.
  
• A German court rules Google is legally liable for its AI Overviews. The Munich court found AI Overviews are Google’s “own words,” so the liability shields that protect ordinary search results don’t cover false claims the summary invents. The court also tossed Google’s “users can verify it themselves” defense, which moves traceable citations in your own AI features from nice-to-have to legal cover.

Addy Osmani, an engineer at Google, borrows Margaret-Anne Storey’s triple-debt model to name a third kind of debt that usually goes untracked. Alongside technical debt (in the code) and cognitive debt (in your head) sits intent debt: the goals, constraints, and rationale behind a decision that never got written down anywhere. Agents can pay down the first two, refactoring a tangled module or re-explaining code you’ve forgotten, but intent is the one input that has to come from a person.

That lands squarely on product managers, because you are usually where intent starts: why a bet is on the roadmap, why you scoped it this narrowly, why you said no to the obvious feature. When that reasoning lives only in your head or a buried Slack thread, the agent drafting the spec or the teammate picking up the work fills the gap with a confident guess, and it is usually wrong. The habit worth building is writing down the “why” the moment you decide, because as Osmani puts it, code is the answer and intent was the question it was meant to solve.

Ethan Mollick had early access to Fable 5 and described a genuine leap: in one test it worked autonomously for nine and a half hours, spinning up its own sub-agents to research, code, and check each other’s work. He puts the change in his own role plainly. He’s no longer the wizard casting the spell but the patron commissioning the work, briefing the model and judging the result without seeing the hundreds of small choices it makes along the way. “Steering is no longer the same as doing,” he writes.

For product leaders, the shift is from doing the work to commissioning and judging it. Mollick can still steer Fable with an ambitious brief, and the more demanding the instruction the better the result, but he sees almost none of the hundreds of choices the model makes along the way. The leverage moves to the two things you still control: how precisely you frame what you want, and how well you can judge what comes back, because the middle of the process has become a black box you don’t get a vote in.

Yoonho Lee pushes back on a belief many ML researchers hold, that real learning only happens in a model’s weights. He argues the “text layer” around a model, its prompts, context, memory, retrieval, and harness code, is a legitimate way to change how a system behaves, and in low-data situations it’s orders of magnitude more sample-efficient than retraining. It also opens what he calls “update-time compute”: a system can learn from a single failure by rereading it, forming a hypothesis, and testing a fix in text, rather than waiting months for the next model to ship.

This matters most for product teams building on models they don’t own. You can’t retrain GPT or Claude, but you control what surrounds them: your proprietary data, your customer signals, your structured memory. Lee’s argument is that this layer is a lasting advantage in its own right, and the teams who treat it that way build something a shared base model can’t hand any competitor.

Shaili Guru argues the free era of MCP servers is ending, and her math shows why: a human touches Salesforce maybe 50 to 200 times a day, while an agent can fire 500 tool calls in ten seconds, landing 10 to 100 times the load on the same infrastructure. She maps three pricing models filling the gap (per-call, where a search costs a cent and a generate costs a dime; re-wrapped human seats like Copilot at $30 a user or Agentforce at $125–150; and outcome-based, like Intercom Fin at $0.99 a resolution), and argues that re-wrapping the seat only postpones the reckoning rather than fixing the unit economics.

Then there’s the dual-identity problem: your product now serves both a human and the agent acting for them, with different traffic, cost, and rate-limit profiles that your pricing and audit logs ignore. For anyone who owns packaging or monetization, this shows up as a near-term line item, so if your pricing assumes humans, map where agents already touch your product before the math surprises you. The PM who can answer “what did our agents cost this month, by workflow” is in a very different conversation than the one who can’t.

FROM THE FEED

Worth a follow this week

• Jeff Gothelf, author of Lean UX, put two AI strategies side by side: the company using AI to cut costs, and IKEA, which reskilled its call-center staff into higher-value roles. His point is that AI should amplify people, not quietly replace them. (Read his take).

• Busra Coskuner, a product coach, walks through how Instaffo adopted AI from the bottom up with no mandate: PMs shipping code, designers building components, while the fundamentals like OKRs, product trios, and user interviews stayed exactly where they were. (Read the case study).

• Sachin Rekhi, longtime product leader and writer, breaks down Zapier’s AI fluency rubric. The part worth stealing is the accountability layer: human judgment defines what success looks like, which makes governance, not adoption, the real constraint as you scale AI across a team. (Read his breakdown).

Make confident roadmap decisions 6× faster with AI built for PMs. Spark synthesizes thousands of customer signals to uncover trends, segment insights, and surface supporting evidence at scale. When the model can build almost anything you can describe, the constraint becomes knowing what’s worth building, and that judgment still runs on customer evidence.

Try Spark for free →