The Product Desk

The Subsidy Era Ends in 30 Days

A platform engineer at a mid-size SaaS company opened her Cursor tab this morning and shipped four PRs before lunch, the way she has every day for six months. She does not know that Anthropic just announced every Claude subscription now includes API credits equal to the plan's dollar amount, and that third-party tool usage gets a separate, limited credit pool. When those credits burn, overages bill at full API rates. For teams running Claude through Cursor, Cline, OpenCode, or any non-Anthropic harness, the effective per-developer cost was subsidized by 70-90% and nobody in finance noticed. That subsidy dies June 15.

The thing being pitched is a credit structure simplification. The thing being done is repricing power users to look like enterprise revenue. Anthropic hired a CFO and is likely targeting an October 2026 IPO. The previous model does not produce the revenue-per-user metrics public investors want to see in an S-1. PMs building on Claude should model at least one more pricing adjustment before October.

ServiceNow Is the Preview of Your Q4

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 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 Displacement Window Is Open

OpenAI responded within hours. Sam Altman offered 2 months of free Codex to enterprise customers who switch within 30 days. The Ramp data showing Anthropic at 34.4% versus OpenAI's 32.3% explains why this offer exists. It is the first time Anthropic has led, and OpenAI is buying back the business adoption lead with displacement pricing.

The Decision Framework

Two axes. First: is your Claude usage load-bearing for a specific production workflow, or exploratory? Second: is the harness replaceable with Anthropic-native tooling at similar quality, or not?

• Load-bearing + replaceable: Renegotiate with Anthropic inside the 30-day window while you still have leverage.
• Load-bearing + not replaceable: Pilot Codex on the free offer this week. Not next month.
• Exploratory (either cell): Move that work to whichever vendor is currently subsidizing it.

The category emerging from this gap is AI cost governance tooling. ServiceNow built AI Control Tower internally and now sells it. For anyone building enterprise software, usage monitoring and cost attribution moved from nice-to-have to procurement blocker the moment a provider raised prices without SLAs or transparency. The multi-model abstraction layer is no longer an engineering convenience. It is the forcing function that determines whether your cost line is a number you can defend in a board meeting next quarter.

Three Giants Picked the Same Week to Ship the Same Thing

SAP shipped a Knowledge Graph for agent context and a €100M partner fund for Autonomous Enterprise. ServiceNow launched Action Fabric, which decouples workflow logic from UI and exposes it via MCP servers for third-party agent execution. Salesforce added native WhatsApp voice to Agentforce Contact Center. Three different pitches. One execution layer underneath: headless workflows callable over MCP.

Companies do not stand up hundred-million-euro funds for features. They stand them up for platform bets they intend to defend for years.

What enterprise buyers are actually doing has already moved. A Fortune 500 procurement lead opened three enterprise software demos this week and asked the same question in each one: "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 is the buying behavior. The deck slide that says "agent-ready" is not.

59% of Token Volume Is Already Agentic

Vercel's AI Gateway production data across 200,000+ teams says it plainly: 59% of token volume now flows through agentic workloads. Anthropic captures 61% of AI spend, mostly Opus on heavy reasoning. Google captures 38% of token volume, mostly Flash on cheap fast tasks. Large teams multi-model route in production — not as a stated strategy, but as the default thing engineers do when one model is too slow and another is too expensive.

The Window Before RFPs Include This

The window before agent-callable APIs show up as line items in enterprise RFPs is two to three quarters. For most teams the work is smaller than the deck will suggest: a week of scoping, two to four weeks of build, assuming the underlying API is not already a mess. The harder, separate question is whether the product's core UI should assume an agent as the primary first-touch user. That is a roadmap question, not a sprint question, and confusing the two is how teams ship MCP endpoints that nobody adopts.

The Diagnostic

Pull the last twenty support tickets and feature requests from top-decile accounts. Count how many assume a human in the seat versus how many assume an agent or integration is doing the work. If that ratio has moved even ten points toward agents in the last two quarters, the headless layer is not a platform bet. It is a retention bet, and the clock runs to the next renewal cycle.

One Person Shipped the Enterprise Pricing Page

Elena Verna joined Lovable in December 2025 as a pure IC, after running growth at Amplitude, Miro, and Dropbox. She spends 90% of her time building, takes almost no meetings, and pushed Lovable's enterprise pricing page to production herself. The traditional version of that project takes a PM, a designer, a couple of engineers, and about a week of calendar time. Verna did it alone in hours.

Lovable has no product managers. Engineers talk to users, write the spec, ship the code, and read the feedback. Growth PMs are being hired in parallel to Verna and do not report to her. That is not a startup quirk. It is a claim that a meaningful share of the PM role was always coordination tax.

The Counter-Signal: Duolingo's 20% Slop Rate

The Duolingo reversal is the cleanest version of a story this beat has been tracking quietly for months. The CEO acknowledged that the blanket 'evaluate all employees on AI usage' policy failed. AI content at scale produced roughly 20% 'slop' that needed human cleanup, and mandating usage across roles generated performative adoption without measurable output gains. They walked it back.

Use 20% as your human-in-the-loop capacity planning assumption until you have your own data. If you're setting team AI adoption goals, measure cycle time and output quality — not tool logins.

The Unbundling Framework

The PM role decomposes into three pillars: cross-functional coordination, customer and market judgment, and strategic prioritization. Pillar one is what fills calendars with meetings and Notion with alignment docs. It is also exactly what AI-enabled flat orgs are deleting. When one operator can design and ship a customer-facing surface without handoffs, the coordinator role becomes overhead.

Ravi Mehta's 'average intelligence' framing is the part teams keep missing in their own planning. AI does not make a PM world-class at design or engineering. It makes them average-to-good at everything at once. For a PM already wired to think across functions, that is an opening. The opening only matters if the time saved goes into shipping rather than into more coordination.

The 2x2 for Monday

• Axis 1: How much of the week goes to building versus coordinating?
• Axis 2: Does the org let a single operator ship a customer-facing surface without a handoff chain?

If an enterprise pricing page takes the team a week and takes an HI-C a few hours, the gap is not speed. It is structure, and it compounds across every experiment and iteration cycle.

The AI Persona Drift Problem Nobody's Testing

The Li et al. paper at COLM 2024 has been sitting in the literature for over a year, and the finding teams keep ignoring is the operational one: significant persona degradation within 8 rounds of dialogue, driven by attention decay. The demo runs three or four turns and looks great. Power users running fifteen-plus-turn sessions are talking to a different product than the one the demo promised. The cheap version of monitoring is to embed a distinctive behavioral marker in the system prompt and watch for its disappearance as a drift-detection canary.