The Product Desk

The Pricing Mechanics That Just Broke

A backend lead opened her Cursor tab on a Tuesday morning, ran the same Claude-powered refactor she runs every Tuesday, and did not notice anything different. She will notice on June 15, when Anthropic splits third-party Claude tool usage (Cursor, Cline, Zed, OpenCode, T3 Code) into a separate credit pool equal to the plan's dollar value, with overages at full API rates. The implicit discount developers were riding — estimated at 70-90% below API pricing — is gone. The per-developer cost assumption for any team running Claude through a non-Anthropic harness is wrong by roughly an order of magnitude.

This is not generosity ending. Anthropic hired a CFO and is aiming at an October 2026 IPO. The S-1 needs revenue-per-user numbers that implicit subsidies cannot produce. PMs building on Claude should model at least one more pricing adjustment before October.

The ServiceNow Warning Shot

ServiceNow CDIO Kellie Romack reported her team's full-year Anthropic budget was consumed by May 2026, five months into a twelve-month plan. She cannot say which users or workloads burned it, because Anthropic does not ship per-user telemetry. PagerDuty and National Life Group report the same. National Life's Nimesh Mehta called Anthropic "great for consumer usage but not great for companies."

AI costs are structurally unpredictable, and the model providers have not built the instrumentation customers need to govern them.

OpenAI's Displacement Offer

Sam Altman offered 2 months of free Codex to enterprise customers who switch within 30 days, timed to the frustration. Ramp's data has Anthropic at 34.4% versus OpenAI's 32.3%. OpenAI lost the business adoption lead for the first time and is counter-attacking with displacement pricing.

The Decision Framework

Two axes. First: is Claude usage load-bearing for production workflows, or exploratory? Second: is the harness replaceable with Anthropic-native tooling at similar quality?

• Load-bearing + replaceable: Renegotiate with Anthropic inside the 30-day window while leverage exists
• Load-bearing + not replaceable: Pilot Codex on the free offer this week, not next month
• Exploratory: Stop paying metered rates for exploration. Move to whichever vendor is currently subsidizing

The 50% rate limit increase Anthropic bundled in is a grace period, not a concession. Build cost telemetry before June 15, not after.

Three Vendors, One Week, One Standard

SAP, ServiceNow, and Salesforce all shipped autonomous agent architectures last week, and they converged on the same execution layer: headless workflows callable over MCP. The pitch deck calls this an industry standard emerging. What is actually happening is three vendors making the same platform bet at the same time. SAP backed it with a €100M partner fund and a Knowledge Graph for agent context. ServiceNow decoupled workflow logic from UI via Action Fabric, exposing it for any third-party agent to call. Companies do not stand up hundred-million-euro funds for features. They do it for bets they intend to defend for years.

The Procurement Shift Already Happened

A procurement manager at a Fortune 500 asked three enterprise software vendors the same question last week: "Can our agents call this directly, or do my people have to click through your UI?" Two vendors had no answer. The third moved to the next stage. That question is showing up in RFPs 2-3 quarters before most product teams expect it, which means the roadmap conversation is already late by the time the sales team forwards the email.

If your APIs are not agent-consumable, agents will route around you to a competitor whose APIs are.

What "Agent-Callable" Actually Requires

The work is smaller than the deck will suggest for most teams: 1 week of scoping, 2-4 weeks of build, assuming the underlying API is not already broken. The separate decision — whether the core UI should be restructured around agents as primary users — is a roadmap question, not a sprint question. Teams conflate the two and ship neither. Getting the sequence wrong in either direction is expensive.

Notion's Contrasting Play

Notion did not build its own agent. It launched an External Agents API letting Claude, Codex, Cursor, and Devin operate directly inside Notion workspaces. That is the context-layer bet: make the product the shared workspace every agent needs. The 2x2 resolves on one axis. If the product holds collaborative data but cannot out-build foundation vendors on agent capability, the Notion path is correct. If it can ship a better agent for a specific workflow, building the agent is the right call. Hedging both is what most decks recommend and what most roadmaps cannot afford.

The Signal in the Support Queue

The forcing function is concrete. Pull the last twenty support tickets from the top-decile accounts. Count how many assume a human in the seat versus how many assume an agent is doing the work. If that ratio moved even ten points toward agents in the last two quarters, the headless layer is a retention bet on the next renewal cycle, not a platform bet on the next board meeting. Read the tickets before picking the sprint.

The Case Study

Elena Verna shipped Lovable's enterprise pricing page to production alone. She is the former head of growth at Amplitude, Miro, Dropbox, and SurveyMonkey, and on the day she pushed that page live there was no PM scoping requirements, no designer on mocks, no engineer on the build. A workflow that used to take three roles and a week of calendar time collapsed into one person and a few hours. In her HI-C (High-Impact Contributor) role, she reports spending ~90% of time building, with almost no meetings.

Lovable has 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 is the model, not an oversight.

What's Actually Being Unbundled

The PM role is four jobs in a trenchcoat: user research, prioritization, spec-writing, and cross-team coordination. AI tools fold the first three into the builder, but only when the builder is the one talking to users. The fourth job disappears when the org is small enough that coordination happens in one channel. Ravi Mehta's framing is the right one: AI does not make a PM world-class at design or engineering. It makes them average-to-good at everything at once.

The PMs who survive this shift look less like project managers and more like mini-GMs who happen to prototype and iterate directly.

Where the Model Breaks

The pattern works at Lovable's size. It breaks at some larger size, and nobody knows where. Two diagnostic questions separate the cells. One: is prioritization owned by a named person, or does it emerge from whoever shouts loudest? Two: do builders talk to users every week, or does feedback arrive filtered through a deck? Lovable sits in "named owner, direct contact." That cell runs without PMs. Every other cell still needs one.

The Talent Risk for Established Companies

Senior builders who can get autonomy and impact density at a flat org will leave to get it. Not all of them want a VP title. The best ones will go where they can spend 90% of the week creating instead of 90% aligning. Companies that ungate information access pull disproportionate talent. Companies that protect management layers end up with coordinators and no builders.

The Framework for Monday

Two axes. How much of the week goes to building versus coordinating, and whether the org lets a single operator ship a customer-facing surface without a handoff chain. If an enterprise pricing page takes the team a week and an HI-C a few hours, the gap is not speed. It is structure, and it compounds across every experiment and iteration cycle.

The Capability Jump

Anthropic's Mythos became the first model to clear both UK AISI simulated attack ranges, achieving full autonomous network takeover. The previous generation capped at 'advanced persistence.' OpenAI's GPT-5.5-cyber cleared one of the two. Both ran ahead of the trend line of AI cyber tasks doubling every few months. The old threat model assumed attackers could get a foothold and then needed a human to escalate. That assumption is stale.

The Mozilla vs. curl Result Is the Product Lesson

A Mozilla security engineer ran Claude Mythos Preview inside a custom agentic harness her team built. The harness writes reproducible test cases, scales across ephemeral VMs, and wires into the existing security lifecycle. It surfaced 271 bugs, including sandbox escapes, race conditions, and use-after-free vulnerabilities that fuzzers had missed for years.

The same Mythos model, pointed at curl's 178K lines of C with generic scanning, found exactly 1 low-severity CVE out of 5 claimed issues. Daniel Stenberg's verdict: "primarily marketing."

The delta between 271 and 1 is not the model. It is 270 bugs worth of harness engineering.

What This Means for Product Security

Here is what is actually happening to users. The time between vulnerability disclosure and weaponized exploit collapsed from weeks to hours. PraisonAI's auth bypass was actively exploited within 4 hours of disclosure. Patch SLAs of 30-60 days were designed for the old exploit economics. Those economics changed this week.

Palo Alto Networks scanning with AI models found dozens of serious vulnerabilities across 130+ products. That is the base rate for any product with real surface area. Enterprise security teams will notice before most product orgs do. The first symptom is an RFP asking for AI-powered security testing evidence in the build pipeline.

The Build Decision

The framework is one question. Does the team already have a corpus of prior findings and a triage pipeline running? If yes, buy model access and build the harness. If no, the sprint work is building the corpus, not evaluating models. Picking one of those is better than hedging both.