The Product Desk

The Implicit Discount That Built Your Unit Economics Is Being Withdrawn

ServiceNow's CDIO Kellie Romack opened a budget report and saw 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 per-user telemetry that would answer those questions. PagerDuty and National Life Group describe the same pattern. National Life Group's Nimesh Mehta calls Anthropic "great for consumer usage but not great for companies."

Here is what is actually happening. Starting June 15, Anthropic splits third-party tool usage (Conductor, Zed, OpenClaw, T3 Code) into a separate credit pool equal to the plan's dollar value. Once credits exhaust, you are on API rates. The 70-90% implicit discount that developers have been quietly building on vanishes. A $200/month plan that was covering $1,500-2,000 of effective API usage now covers $200 of third-party usage plus the original first-party allocation.

The era of subsidized AI inference through integrations is ending. Model the cost impact now, not after the bill arrives.

The Displacement War Is Live

OpenAI responded within hours of Anthropic's announcement. Sam Altman offered 2 months of free Codex to enterprise customers who switch within 30 days. That is displacement pricing timed to developer frustration. Public criticism from Theo, Jeremy Howard, Matt Pocock, and Omar Sanseviero landed in the same window. The Ramp data showing Anthropic at 34.4% versus OpenAI's 32.3% in business adoption explains the urgency: OpenAI lost the business adoption lead for the first time and is fighting to reclaim it.

Tokenmaxxing Is a Real Problem, Not a Meme

The pitch is "AI adoption." What teams actually do is game the leaderboard. Multiple sources confirm what Duolingo quantified publicly: AI-generated content at scale produces ~20% unusable output requiring human QC, and blanket AI mandates produced performative adoption without productivity gains. Amazon's internal AI usage mandates have staff gaming token leaderboards. The metric that looks like adoption is Goodhart's Law in production.

The 2x2 for This Sprint

One axis: is inference cost fixed per seat or variable per call. Other axis: does the customer pay 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.

Three Mega-Vendors Converged on the Same Architecture This Week

SAP shipped a Knowledge Graph for agent context and committed €100M to an Autonomous Enterprise partner fund. ServiceNow launched Action Fabric, which decouples workflow logic from the UI and exposes it via MCP servers any third-party agent can execute. Salesforce added native WhatsApp voice to Agentforce. Three of the five largest enterprise vendors landing in the same place in the same week is not coincidence. It is a commitment to a specific architecture: workflow logic accessible to agents, not just to humans clicking through screens.

A procurement manager at a Fortune 500 opened three enterprise software 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 did not have an answer. The third moved to the next stage.

59% of Token Volume Is Now Agentic

Vercel's AI Gateway production data across 200,000+ teams confirms it: 59% of all token volume flows through agentic workloads. What teams tell themselves is that they will pick a model. What teams actually do is route across several. Anthropic captures 61% of spend, mostly Opus for reasoning. Google captures 38% of volume, mostly Flash for cheap, fast tasks. The interaction-model shift that defines this product cycle is not upcoming. It already happened in production traffic.

Apple's Agent App Store Adds a Platform Constraint

Apple is building AI agent governance into the App Store, likely revealed at WWDC June 2026. They are solving three problems at once: agent review and approval, preventing agents from spawning unauthorized sub-apps, and ensuring agents cannot route around App Store fees. The thing being pitched is governance. The thing being done is fee protection. Roadmaps that involve dynamic interface creation on iOS should architect for Apple's constraints now, not after they ship.

The Build Decision

For most teams, shipping an MCP server against an existing API is a week of scoping, 2-4 weeks of build, assuming the underlying API is not already a mess. That is the easy decision. The harder one is whether the product's core UI should be restructured around an agent as the primary first-touch user. The 2x2 to apply this sprint: on one axis, is the API agent-ready or not. On the other, is the agent the first-touch user or a secondary path. The agent-ready and agent-first cell is where the three vendors above are heading. Pick a cell deliberately. Hedging both is the expensive answer.

271 Bugs vs. 1 Bug: Same Model Class, Different Harness

A Mozilla security engineer ran the same model class twice this quarter and got two completely different products. Pointed at Firefox through a custom agentic harness built on Anthropic's Mythos Preview — writing reproducible test cases, scaling across ephemeral VMs, wired into the existing security lifecycle — the system surfaced 271 bugs, including sandbox escapes, race conditions, and use-after-free vulnerabilities that fuzzers had missed for years. Pointed at curl's 178K lines of C with less harness investment, the same model class flagged 5 claimed vulnerabilities and exactly 1 became a low-severity CVE. Daniel Stenberg's verdict on the curl submissions was "primarily marketing."

The model is table stakes. The harness is the product. Mozilla is wiring this pipeline into CI to scan patches as they land, which is the benchmark anyone selling into this space will be measured against.

Mythos Achieves Full Network Takeover

Anthropic's Mythos and OpenAI's GPT-5.5-cyber both reached 'full network takeover' in UK AI Security Institute testing, where the previous generation capped at 'advanced persistence.' The old threat model assumed a foothold required human expertise to escalate; the new one assumes AI handles the full kill chain without a human in the loop. Mythos completed both of AISI's hardest tests. GPT-5.5-cyber managed one.

Exploitation Timelines Collapsed

The thing being pitched is "AI for defenders." The thing being done is a compression of attacker timelines. PraisonAI, the open-source multi-agent orchestration project, went from CVE disclosure to actively exploited in 4 hours. An 18-year-old NGINX RCE turned up in the rewrite module, affecting virtually every modern web deployment. EDR products that took weeks to reverse-engineer now take days using LLMs, and the cost of finding exploit chains dropped from a week of human time to 15 minutes of compute.

What This Means for Your Product

The forcing function for product teams this quarter is concrete. If you are evaluating AI security scanning, require vendors to demonstrate custom harness capability against your codebase, not generic scanning against a demo repo. Mozilla's 271-bug result versus curl's 1-CVE outcome is the benchmark frame. If you are shipping software, a 30-day patch SLA assumes attackers take days to weaponize, and that assumption is provably false at 4 hours. Enterprise RFPs will start asking for AI-powered security testing evidence in the build pipeline within two quarters. Decide now what that evidence looks like, before procurement decides for you.

The Existence Proof

Elena Verna — former head of growth at Amplitude, Miro, Dropbox, and SurveyMonkey — pushed Lovable's enterprise pricing page to production by herself last week, with no PM scoping requirements, no designer on mocks, and no engineer on the build. She says she spends ~90% of her time building, almost none of it in meetings. Lovable employs zero product managers. Engineers talk to users, write specs, ship code, and read the feedback. Growth is fast enough that the absence is the operating model.

What the PM Role Actually Decomposes Into

Separate the thing being pitched from the thing being done. The PM value prop splits into three pillars: (1) cross-functional coordination, (2) customer and market judgment, and (3) strategic prioritization. Pillar one is what fills calendars with meetings and Notion with alignment docs, and it is the pillar AI-enabled flat orgs are eliminating. When one operator can design and ship without handoffs, the coordinator becomes overhead rather than enablement.

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

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

Duolingo publicly acknowledged that its blanket 'evaluate all employees on AI usage' policy failed. AI content at scale produced roughly 20% unusable output that required human quality control. Mandating AI use across every role produced performative adoption without productivity gains, and the policy has since been reversed. The quality gate is what separates the Verna model (senior expert plus AI equals production output) from the mandate model (average contributor plus AI equals slop).

Claude Code's /goal Command

Claude Code's new /goal command runs fully unattended multi-turn coding sessions. A separate Haiku model evaluates completion against measurable conditions. That evaluator-judge pattern — a lightweight model reads only transcripts and judges against user-defined criteria — is the reference architecture for any autonomous AI workflow. Vague conditions break it: infinite token-burning loops, or hallucinated success.

The Honest Diagnostic

Look at what shipped last quarter and split the PM contribution into two columns: judgment about what to build, and coordination of the people building it. If coordination went to zero tomorrow, the judgment column has to justify the role on its own. If it does, the job gets better. If it does not, the job gets done by someone like Verna.