The Product Desk

The Pricing Event

A developer opened Cursor on a Tuesday morning and ran Claude against a refactor she had run forty times the week before. The latency felt the same. The bill will not. Starting June 15, Anthropic separates Claude usage through third-party tools (Cursor, Cline, Zed, OpenCode, Conductor) into a credit pool equal to the subscription's dollar value. Burn the pool, pay full API rates. That ends the 70-90% implicit discount power users have been quietly living inside. The 50% rate limit bump for two months is a grace period dressed up as a concession.

The era of subsidized AI inference through integrations is ending. The question is whether your cost model noticed.

The ServiceNow Warning

ServiceNow's CDIO Kellie Romack watched her team's full-year Anthropic budget get consumed before mid-2026. She cannot tell you which users drove it or which workloads, because Anthropic does not ship per-user telemetry. PagerDuty and National Life Group describe the same shape of problem. National Life's Nimesh Mehta calls Anthropic "great for consumer usage but not great for companies."

Separate what was pitched from what is happening. The pitch is per-seat AI productivity. What customers actually do is ship a feature, watch adoption land, and watch usage scale with success. Costs are unpredictable by architecture, not by accident. The pricing model assumed 3x/day usage. Users found it saved them two hours and ran it 11x/day. Retention improved. Gross margin went sideways.

The Competitive Response

OpenAI picked the same week to offer two months free Codex for enterprise teams switching within 30 days. This is displacement pricing timed to Anthropic's moment of developer frustration. Ramp data has Anthropic at 34.4% versus OpenAI's 32.3% in business adoption. OpenAI lost the lead for the first time, which explains the timing.

Why This Is an IPO Signal

Anthropic hired a CFO and is likely targeting an October 2026 IPO. The old model — enormous implicit subsidies for power users — does not produce investor-grade revenue-per-user metrics. Expect at least one more pricing adjustment before October as the S-1 narrative tightens. Model forward accordingly.

The Two Product Categories This Creates

1. AI cost governance — ServiceNow built AI Control Tower internally and now sells it. Per-customer, per-feature inference cost attribution moved from nice-to-have to procurement blocker the day a CDIO could not answer who burned the budget.
2. Multi-model abstraction layers — stop being an engineering convenience and become strategic infrastructure the moment a provider can raise prices without SLAs or usage transparency.

The Worked Example

Elena Verna led growth at Amplitude, Miro, Dropbox, and SurveyMonkey. In December 2025, Lovable moved her into a pure IC role. She now spends 90% of her time building, has almost no meetings, and personally shipped Lovable's enterprise pricing page to production. In a traditional org that ship list needs a PM, a designer, engineers, and about a week of calendar time. She did it alone.

Lovable runs with zero product managers. Engineers talk to users, write specs, ship code, and read the feedback themselves. The company is growing fast enough that the absence reads as a design choice, not an oversight.

The PM value proposition decomposes into three pillars: cross-functional coordination, customer/market judgment, and strategic prioritization. Pillar one is what AI-enabled flat orgs are eliminating.

What AI Actually Enables Here

Ravi Mehta's framing is the useful one. AI does not turn a PM into a world-class designer or engineer. It makes them "average-to-good at everything at once." For a PM who already thinks across functions, that is leverage. Only if the recovered time goes into shipping rather than into coordinating other people shipping. The PMs who survive look less like project managers and more like mini-GMs who prototype directly.

The Counter-Signal: Duolingo's Reversal

Duolingo's CEO admitted the blanket "evaluate all employees on AI usage" policy failed. AI content at scale produced ~20% unusable output requiring human QC, and the mandate produced performative adoption instead of throughput. They reversed it. Forcing AI use without measuring output quality is the thing teams tell themselves is velocity. What it actually is, is theater.

The Structural Economics

Senior builders who can get autonomy at a Lovable-style flat org will leave to get it. Companies that ungate information access attract talent density that compounds. Companies that protect management layers keep coordinators and lose builders. When Verna says some leaders respond "Absolutely not, I need another VP title," that is the filter doing its job.

The Diagnostic

The Lovable cell works without PMs. Every other cell still needs one. Cutting the role before the org has actually moved into that cell is how good products get slower.

The Offensive Threshold

A red-teamer sat down last month with Anthropic's Claude Mythos and watched it clear both UK AISI simulated attack ranges end to end, taking the network without a human in the loop. OpenAI's GPT-5.5-cyber got one of two. The prior generation stalled at "advanced persistence." The threat model most security teams wrote down assumed attackers could land a foothold but needed a senior human to escalate. That assumption is the thing being defended. It is not the thing actually happening.

What attackers actually do now is find-and-chain in near real-time. Palo Alto Networks pointed these models at 130+ products and surfaced dozens of serious vulnerabilities. The 30-60 day patch SLA was negotiated against exploit development that took weeks. The new exploit development takes hours.

The window between vulnerability disclosure and working exploit compressed from weeks to hours. Your SLA was designed for the old economics.

The Defensive Threshold

Mozilla wrapped Claude Mythos Preview in a custom agentic harness and surfaced 271 bugs in Firefox, including sandbox escapes, race conditions, and use-after-free issues fuzzers had missed for years. The same Mythos model aimed at curl's 178K lines produced exactly 1 low-severity CVE from 5 claimed issues. Daniel Stenberg called it "primarily marketing."

Separate the thing being pitched from the thing being done. The pitch is "the model finds bugs." The product is the harness around it: a prior bug corpus for grounding, a triage pipeline that filters noise before a human sees it, and a team deciding what counts as a real finding. curl ran the model against the code. Mozilla built 270 bugs of infrastructure.

New Attack Surface: Your AI Endpoints

The honeypot study quantifies what deployment teams already suspected. AI model servers get indexed by Shodan within 3 hours of exposure. Over one month, researchers logged 113,000+ requests, 23% of them targeting AI-specific paths (/v1/models, /api/tags, .env files). In the final week the same honeypot caught 175 active LLM-hijacking attempts aimed at compute theft and credential exfiltration. LLM-Scanner tooling updated mid-experiment to detect the honeypots. Attackers are iterating faster than most security roadmaps.

What This Changes