The Product Desk

The Arbitrage Is Closing

A developer opened her Cursor session this morning, watched Claude burn through context for an hour, and saw a charge that, last month, would have been absorbed somewhere upstream. Anthropic announced that every Claude subscription now includes API credits equal to the plan's dollar amount — $200 plan, $200 API credits. For users of first-party Claude tools, the framing is generous. For the cohort running Claude through Cursor, Cline, OpenCode, Aider, Conductor, or Zed, it is a price increase of roughly an order of magnitude. Starting June 15, third-party tool usage draws from a separate credit pool, with overage billed at full API rates.

The thing being pitched is "fair credit allocation." The thing being done is unwinding the implicit subsidy that made third-party harnesses viable. Model the cost impact this week.

Why Now: The IPO Math

Anthropic hired a CFO and is likely targeting an October 2026 IPO. Power users on enormous implicit subsidies do not produce the revenue-per-user numbers a public-market book wants to show. Multiple sources confirm ARR grew from ~$9B to $30B+ in four months. At a $900B valuation, Anthropic needs $40-50B in annual revenue to justify even a generous forward multiple. There will be at least one more pricing adjustment before the S-1 drops.

OpenAI's Counter-Move

Sam Altman offered 2 months of free Codex to enterprise customers who switch within 30 days. This is displacement pricing aimed at Anthropic's moment of maximum developer frustration. The Ramp data explains the urgency. Anthropic reached 34.4% of business customers against OpenAI's 32.3% — the first time OpenAI lost the lead. OpenAI is buying back share at the exact moment developers are annoyed at the other vendor.

The Decision Framework

Draw a 2x2 this week. 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 to explore. Move to whichever vendor is currently subsidizing it.

The 50% rate limit increase Anthropic offered for two months is a grace period, not a permanent concession.

ServiceNow Is the Canary

ServiceNow's CDIO Kellie Romack opened the spend dashboard and watched her team's full-year Anthropic budget get consumed before mid-2026. She cannot tell you which users drove it, or which workloads. 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.'

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

Tokenmaxxing Compounds the Problem

Several reports this week document the parallel failure mode. Employees gaming AI usage metrics to appear productive. Amazon mandated 80%+ weekly AI tool usage. Duolingo mandated evaluation on AI usage across all roles. Both produced performative adoption without productivity gains. Duolingo's CEO publicly acknowledged the policy failed. Their data shows ~20% of AI-generated content is 'slop' requiring human rework. They have reversed the mandate.

The Structural Problem

Here is what teams tell themselves: we priced the feature with a healthy margin over inference cost. Here is what users actually do: find the workflow that saves them two hours, 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. 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 on a feature the board deck says is working.

Two Categories Are Being Pulled Into Existence

First: AI cost governance as a product category. ServiceNow built AI Control Tower internally and now sells it to customers. Per-customer, per-feature cost attribution has moved from nice-to-have to procurement blocker. Second: multi-model abstraction stops being an engineering convenience and becomes strategic infrastructure the moment a provider raises prices without SLAs.

The Forcing Function

Two axes for this sprint. Does the product have per-customer, per-feature inference cost telemetry in production today. And is there a services motion that can be billed against deployment. The cell to be in is yes telemetry + yes services attach. Every other cell is a finance conversation waiting to happen.

Three Giants Converged on the Same Architecture

SAP, ServiceNow, and Salesforce all shipped headless agent architectures this week. All three landed on MCP (Model Context Protocol) as the enterprise agent standard. ServiceNow's Action Fabric decouples workflow logic from UI and exposes it for any third-party agent to call. SAP shipped a Knowledge Graph for agent context, plus a €100M partner fund to seed the ecosystem. These are platform commitments. Companies do not make those lightly.

A procurement manager spent forty minutes clicking through a vendor onboarding flow. Next quarter, an agent will. She will review the result. Products that can't be called headlessly get bypassed.

The Evidence Is Quantified

Vercel's production AI Gateway data across 200,000+ teams shows 59% of all token volume is now agentic workloads. This is the primary consumption pattern, not a pilot. Multiple sources report enterprise buyers asking the same question in demos: 'Can our agents call this directly, or do my people have to click through your UI?' Two vendors without an answer lost to the one that had one.

The Build Decision Is Smaller Than It Seems

Here is what teams tell themselves: agent-readiness is a quarter of work. Here is what most teams actually need: shipping an MCP server against an existing API is a week of scoping, 2-4 weeks of build, assuming the underlying API isn't already a mess. The larger question — whether the product's core UI should be restructured for agent-first interaction — is a roadmap question, not a sprint question. Don't conflate them.

Where Glean's Data Challenges the Default

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 intelligence layer above raw MCP — structured context, domain knowledge, workflow semantics — is where differentiation lives. Shipping the protocol is table stakes. Shipping the context is the product.

The Two-Axis Sort for Your Backlog

One axis: does the workflow have a stable, documented API surface an external agent can call? Other axis: does the outcome require a human approval step that only makes sense inside the product UI? Workflows with both survive the shift. Workflows with neither get bypassed inside two quarters. The interesting cases worth arguing about before next planning are the ones with one but not the other. Pick one of those to defend this sprint. Hedging both is worse than picking wrong.

The Worked Example

Elena Verna led growth at Amplitude, Miro, Dropbox, and SurveyMonkey. In December 2025, Lovable flattened its org and moved her into a pure IC role called HI-C (High-Impact Contributor). She now spends 90% of her time building, sits in almost no meetings, and personally shipped Lovable's enterprise pricing page to production. The traditional version of that project takes a PM, a designer, engineers, and a week of calendar time. Verna shipped it alone.

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.

Lovable Has Zero PMs

This is not an oversight. Engineers talk to users, write specs, ship code, and read feedback. The company is growing fast enough that the absence is the model. Here is what teams tell themselves PMs do: own the roadmap. Here is what Lovable's engineers actually do: own the roadmap. The bet is that when a single operator can design and ship without handoffs, the coordinator role becomes overhead. Lovable is hiring Growth PMs parallel to Verna, not reporting to her — a collection of autonomous operators, not a growth team.

The Abridge Counter-Pattern

Abridge at $5.3B runs the opposite model and it also works. Their three-act ladder — save time, then save money, then save lives — required someone to decide when the wedge had earned the right to expand. They employ 'clinician scientists' (MDs who are also full-stack engineers) and they maintain written PRDs. The judgment about sequencing is the product work. The variable across both companies is not headcount. It is what the PM actually does with the calendar.

The Diagnostic for Monday

Two questions, in order. First: of the work shipped last quarter, how much of the PM contribution was judgment about what to build versus coordination of the people building it? Second: if the coordination half went to zero tomorrow, does the judgment half justify the role?

If the answer to the second question is yes, the job gets better. If it is no, the job gets done by someone like Verna. Pick which half to invest in this quarter. Hedging both is the outcome that ends badly.