The Product Desk

What Actually Changed

A developer on the $200 Claude plan opened Cursor this morning and shipped a feature in about ninety minutes of model time. She did not check her usage. She has not checked it for six months, because the math always worked out: she paid Anthropic $200 and got something that, at API rates, would have cost roughly ten times that. Anthropic just announced that every Claude subscription now includes API credits equal to the plan's dollar amount ($200 plan = $200 API credits). The pitch is generous. What it actually does, for the cohort using Claude through Cursor, Cline, Zed, OpenCode, and T3 Code at effective 70-90% discounts to API pricing, is raise the per-developer cost by roughly an order of magnitude. Starting June 15, third-party tool usage gets a separate credit pool. When those credits burn, the meter runs at full API rates.

The era of subsidized AI inference through integrations is ending. The per-developer cost assumption in most budget decks is now wrong by an order of magnitude.

Why Now: The IPO Clock

Anthropic hired a CFO and is likely targeting an October 2026 IPO. The model where power users got large implicit subsidies does not produce the revenue-per-user numbers public market investors want to see. Ramp data has Anthropic at 34.4% business adoption vs. OpenAI's 32.3%, the first time Anthropic has led. They are consolidating pricing power at the moment they have the leverage to do it.

OpenAI's Displacement Play

Sam Altman offered 2 months of free Codex to enterprise customers who switch within 30 days. The timing tracks Anthropic's developer frustration to the hour — Theo, Jeremy Howard, Matt Pocock, and Omar Sanseviero posted criticism within hours of the announcement. OpenAI lost the business adoption lead for the first time and is buying it back with displacement pricing.

The Decision Framework

The Deeper Signal

Multiple sources confirm this is an IPO-driven margin play. PMs building on Claude should model forward at least one more pricing adjustment before October as the S-1 narrative tightens. The 50% rate limit increase for two months is a grace period, not a concession. Teams that have a switching-cost memo ready will move in 72 hours. Teams that do not will spend the quarter in a Slack thread.

The Budget Blowout Pattern

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

The signal is not that AI is expensive. The signal is that AI costs are structurally unpredictable and the model providers have not built the instrumentation customers need to govern them.

The Performative Adoption Trap

Duolingo's CEO publicly acknowledged that the blanket 'evaluate all employees on AI usage' policy failed. Two findings worth keeping: AI content at scale produces roughly 20% unusable 'slop' that needs human QC, and mandating AI usage across all roles produced performative adoption with no productivity gain. They reversed the policy. Separately, 'tokenmaxxing' — employees gaming AI consumption metrics to look productive — is showing up at enterprises including Amazon. What teams tell themselves users do is automate work. What users actually do, when scored on AI usage, is consume tokens.

Why This Is Structural

The cost model breaks on a mismatch most teams have not named. A user finds the workflow that saves her two hours, and she runs it 11 times a day instead of the 3 the pricing model assumed. Retention improves. Usage depth climbs. Gross margin goes sideways, then down. Inference cost is variable. Revenue per seat is fixed. That gap widens with success.

Two Product Categories Being Pulled Into Existence

1. AI cost governance: ServiceNow built AI Control Tower internally and now sells it to customers. Per-customer, per-feature inference telemetry has moved from nice-to-have to procurement blocker.
2. Multi-model abstraction: Strategic the moment a provider raises prices without SLAs or usage transparency. The low switching costs that worry Anthropic are the buyer's leverage.

The Forward-Deployed Engineer Tax

Google Cloud is hiring hundreds of FDEs for Gemini. OpenAI acquired Tomoro to stand up DeployCo with 150 FDEs. ServiceNow and Salesforce are building the same capability in-house. The 'AI replaces services' story does not survive contact with this data. Self-serve AI launches into enterprise without a services component should plan against roughly 20% activation.

The Stable Cell

One axis: inference cost fixed per seat or variable per call. Other axis: customer pays per seat or per outcome. The only stable cell is variable cost matched to variable pricing. Every other cell is a bet that usage will not grow, which is a strange bet to place on a feature the same deck is telling the board is working. The decision a product team can make Monday is which cell they are actually in, and whether the pricing page reflects it.

Three Vendors, One Architecture, One Week

SAP, ServiceNow, and Salesforce all shipped autonomous agent architectures this week, and they converged on the same execution layer: headless workflows callable over MCP. ServiceNow's Action Fabric decoupled workflow logic from the UI and exposed it for any third-party agent to call. SAP shipped a Knowledge Graph for agent context plus a €100M partner fund. Hundred-million-euro funds are not how companies announce features. They are how companies announce platform bets they plan to defend for years.

The enterprise buying committee has moved from 'show me the dashboard' to 'can our agents orchestrate your workflows.' The window before this shows up in RFPs is 2-3 quarters.

The Production Data Confirms It

Vercel's AI Gateway data across 200,000+ production teams shows 59% of all token volume now flows through agentic workloads. Anthropic captures 61% of AI spend, driven by Opus on heavy reasoning. Google captures 38% of volume, driven by Flash on cheap, fast tasks. Most large teams route across multiple providers rather than committing to one. That is not a request-response interaction model anymore. It is autonomous agents picking the model for the task.

What 'Agent-Consumable' Actually Means

A procurement manager at a Fortune 500 opened three enterprise demos this week and asked the same question in each: '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.

The Build Scope Is Smaller Than Expected

For most teams, shipping an MCP server against the existing API is a week of scoping, 2-4 weeks of build, assuming the underlying API is not already a mess. The harder question is whether the product's core UI should be restructured for agent-first interaction. That is a roadmap question, not a sprint question.

Glean's Cautionary Data

Glean benchmarked off-the-shelf MCP against an enterprise knowledge graph. Raw MCP used 30% more tokens and was preferred 2.5x less on agentic tasks. The protocol gets you on the field. The intelligence layer above it (contextual enrichment, domain knowledge) is where the product actually wins. MCP is necessary, not sufficient.

The Forcing Function

Pull the last twenty support tickets from top-decile accounts. Count how many assume a human in the seat versus an agent doing the work. If that ratio shifted even ten points toward agents in the last two quarters, headless is a retention bet on the next renewal cycle, not a platform bet for the next board meeting. That is the decision this sprint, not next year.

The Lovable Evidence

Lovable has no product managers. The company is growing fast enough that the absence is deliberate. It is the operating model. Engineers talk to users, write specs, ship code, and read feedback. Elena Verna, who ran growth at Amplitude, Miro, and Dropbox, now spends 90% of her time building with almost no meetings. She personally shipped Lovable's enterprise pricing page to production. That work previously needed a PM, a designer, an engineer, and roughly a week of calendar time.

The PM role decomposes into three pillars: coordination, judgment, and prioritization. AI-enabled flat orgs are eliminating the first. The question is whether the remaining two justify the title.

What's Actually Being Eliminated

The PM value proposition has four components:

1. User research: understanding what to build
2. Prioritization: deciding what to build next
3. Spec-writing: defining what to build
4. Cross-team coordination: aligning who builds it

Lovable's bet is that components 1-3 collapse into the engineer when the engineer talks to users directly, and component 4 disappears when the org is small enough that coordination happens in a shared channel. Abridge runs the counter-model: clinician-scientists (MDs who are also full-stack engineers) embedded in product teams, with PRDs and written specs still load-bearing. Their release cadence accelerated 4-6x despite keeping the specs.

The HI-C Pattern

Ravi Mehta's framing is the useful 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 look less like project managers and more like mini-GMs who prototype and iterate directly. Companies that ungate information access attract disproportionate talent density. The ones that protect management layers end up with coordinators and no builders.

The Talent Retention Risk

Senior builders who can get autonomy at a Lovable-style flat org will leave to get it. Some will say they need a VP title. The best builders will not. They will go where they can spend 90% of their time creating instead of 90% aligning.

The Diagnostic

Two questions for Monday. First, of the work shipped last quarter, how much of the PM contribution was judgment about what to build vs. coordination of people building it? Second, if the coordination half went to zero tomorrow, does the judgment half justify the role? If yes, the job gets better. If no, the job gets done by someone else.