The Product Desk

What Actually Changed

An engineering lead opened her Cursor tab on a Tuesday and shipped three PRs before lunch. She has been doing this for months without thinking about token economics. Starting June 15, Anthropic gives Claude usage through third-party tools (Cursor, Cline, OpenCode, Zed, Conductor) a separate credit pool equal to the plan's dollar value. Once burned, the meter switches to full API rates. The implicit 70-90% discount developers have been operating inside quietly disappears in 30 days.

Teams tell themselves they pay for Claude. What they actually pay for is a subsidized harness experience. The per-developer cost assumption in the budget is now wrong by roughly an order of magnitude.

Why It's Happening Now

Anthropic hired a CFO and is targeting an October 2026 IPO. Power users on enormous implicit subsidies do not produce the revenue-per-user numbers public markets reward. ARR moved from $9B to $30B+ in four months. The company is being offered funding at a $900B valuation, ahead of OpenAI's $852B. Customers are absorbing the increases without churning, which is the only signal that matters.

Anthropic is closing an arbitrage that a lot of teams built their unit economics on without realizing it. Those are different conversations than 'pricing alignment.'

The Counter-Move

OpenAI responded within hours. Sam Altman offered 2 months of free Codex to enterprise customers who switch within 30 days. That is displacement pricing, timed to Anthropic's moment of developer frustration. Ramp data has Anthropic at 34.4% versus OpenAI's 32.3% — OpenAI lost the business adoption lead for the first time, and the offer is priced accordingly.

The Decision Framework

Draw a 2x2 before next sprint. One axis: is the Claude usage load-bearing for a production workflow, or exploratory. Other axis: 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 the leverage is real.
• Load-bearing + not replaceable: Pilot Codex on the free offer this week, not next month.
• Exploratory in either cell: Stop paying metered rates for exploration. Move to whichever vendor is currently subsidizing it.

Plan for 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 permanent concession.

Three Enterprise Giants Converged This Week

SAP shipped a Knowledge Graph for agent context and a €100M partner fund for Autonomous Enterprise. ServiceNow's Action Fabric decoupled workflow logic from UI and exposed it through MCP servers for third-party agent execution. Salesforce added native WhatsApp voice to Agentforce. Three of the five largest enterprise vendors landing on the same architecture in the same week is a commitment, not a coincidence.

A procurement manager at a Fortune 500 opened three demos this week and asked the same question: '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 59% Inflection

Vercel's AI Gateway production data across 200,000+ teams shows 59% of token volume is now agentic workloads. Anthropic captures 61% of spend, mostly Opus for reasoning. Google captures 38% of volume, mostly Flash for cheap and fast. The consumption model flipped from human-interactive to agent-mediated, and the bill of materials reflects it.

What Buyers Are Actually Asking

Enterprise buying committees used to ask 'show me the dashboard.' This quarter they ask 'can our agents orchestrate your workflows.' Notion's developer platform launch — markdown API, external data sync, agent tool building, code execution on their infrastructure — is pitched as a developer platform. What it is doing is positioning Notion as the workspace where agents live. Cline released an open-source agent SDK with native teams and subagents. The orchestration layer is commoditizing.

The Build Decision

For most teams, shipping an MCP server against the existing API is 2-4 weeks of build, assuming the underlying API is not a mess. That is the easy decision. The harder one is whether the product's core UI should be restructured around the assumption that an agent, not a human, is the primary first-touch user for a growing share of sessions.

The window before this shows up in RFPs is 2-3 quarters. SAP paying partners to build on this now is a stronger signal than any keynote.

ServiceNow Is the Canary

ServiceNow's CDIO Kellie Romack watched her team's full-year Anthropic budget get consumed before May. She cannot tell which users drove it, or which workloads drove it, because Anthropic does not ship the telemetry that would answer either question. PagerDuty and National Life Group describe the same problem. The signal is not that AI is expensive. It is that AI costs are structurally unpredictable and model providers have not built the instrumentation customers need to govern them.

The Structural Problem

Here is what teams tell themselves: we priced the feature with healthy margin over inference cost. Here is what users actually do: they find the workflow that saves two hours, and they run it 11 times a day instead of the 3 the pricing model assumed. Retention improves. Usage depth climbs. Gross margin goes sideways, then down. That is the base case for any AI feature good enough that users adopt it.

The only stable pricing cell is variable cost matched to variable pricing. Everything else is a bet that usage will not grow, which is a strange bet to place on a feature the deck tells the board is working.

Two Product Categories Emerging

1. AI cost governance tooling: 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 in one budget cycle.
2. Multi-model abstraction layers: Not an engineering convenience anymore. Strategic infrastructure the moment a provider can raise prices without SLAs or usage transparency.

The Readiness Gap Compounds It

Only 15% of organizations have the data foundation for agentic AI, and they are spending millions anyway. Nearly half cite data quality and lineage as the primary blocker. The failure pattern is buy, fail to activate, churn, blame vendor. For teams shipping AI features, onboarding is the product.

Why Costs Won't Fall As Assumed

OpenAI committed $20B to Cerebras in a multi-year capacity lockup. Nebius reports 4+ customers competing for every GPU brought online. When the largest buyers pre-commit at twenty billion dollars, the negotiating position of mid-market buyers gets weaker, not stronger. Roadmaps that assume inference costs drop X% per year should be stress-tested against the floor being set by buyers willing to pre-commit at that scale.

The Evidence Is No Longer Theoretical

Elena Verna — former head of growth at Amplitude, Miro, Dropbox — shipped Lovable's enterprise pricing page to production alone, without a PM scoping requirements, a designer on mocks, or an engineer on the build. She reports spending 90% of her time building, with almost no meetings. Lovable has zero product managers; engineers talk to users, write specs, ship code, and read feedback directly.

What Actually Got Unbundled

The PM role decomposes into four jobs: user research, prioritization, spec-writing, and cross-functional coordination. AI tools compress the first three into a single operator when that operator talks to users directly. The fourth disappears when the org is small enough that coordination happens in a shared channel.

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 variable is ownership rather than title.

The Quality Floor Rises, Not the Ceiling

Ravi Mehta's framing is the useful one: AI doesn't make a PM world-class at design or engineering, it makes them average-to-good at everything at once. Designers are 0.25% of the US workforce; AI tools serve the 99.75% who could never have hired one. The pattern repeats across disciplines.

But: The 20% Slop Rate Is Real

Duolingo ran the experiment and reported back. Mandating AI across all roles produced performative adoption without productivity gains, and AI content at scale yielded roughly 20% unusable output. They reversed the policy. The operating lesson is to measure cycle time and output quality rather than tool logins. AI persona drift degrades meaningfully within 8 dialogue rounds from attention decay, so the three-to-four-turn demo looks clean while the fifteen-plus-turn power user is using a different product.

The Diagnostic

Take last quarter's shipped work and split the PM contribution into two buckets: judgment about what to build versus coordination of people building it. If the coordination half went to zero tomorrow, the question is whether the judgment half justifies the role on its own. If it does, the job gets better. If it doesn't, the job gets done by someone like Verna.