The Product Desk

Why usage data from the boost period cannot be your baseline

An engineer hits the quota wall mid-task and files a bug against the internal agent. The agent is not the problem. Two vendors changed what a dollar buys in the same fortnight, in opposite directions, and both changes corrupt the measurement anyone would use to plan around them. Unwind AI reports that OpenAI is resetting paid Codex and ChatGPT Work usage after fixing token-burn bugs spread across image compaction, memory workers, goals, automations, subagents, computer history, rolling summaries and MCP result handling. Those bugs silently consumed 10-50% of paid quota. Quotas may now stretch that much further. So whatever the dashboard says for the affected period describes a defect, not real team consumption. Baseline after September 14, not before.

That is the practical argument for first-party token accounting. Invoice-level spend is accounting. Per-feature and per-session cost in your own telemetry, reconciled weekly against vendor reports, is measurement. Without the second one, an internal regression and a vendor billing bug produce identical symptoms. One vendor has now demonstrated the billing bug is real, so the reconciliation stops being a hygiene item.

Entitlement is the price lever

Anthropic's list price did not move. The entitlement moved, which raises cost per unit of work for exactly the heavy users extracting the most value. TLDR IT reads this as a vendor managing inference margin without a pricing announcement. Techpresso reads the same cut as a capacity signal, alongside Anthropic's simultaneous legal and infrastructure pressure. The two readings diverge on whether this is temporary rationing or structural repricing. That distinction decides whether the answer is buying seats or re-architecting.

The test is cheap: watch whether the next change also arrives as an entitlement adjustment rather than a rate card. If it does, assume every capacity number in the roadmap is a vendor-controlled variable and stop treating it as a constant.

The arithmetic, and who breaks first

Divide 1.25 by 1.50 and you get roughly 0.83. Any squad consuming more than 83% of the temporary boosted cap loses work mid-month. Finding those squads is a two-hour query against usage data, and the window closes on September 13. Unwind AI's recommendation goes one step further: cap agent workflows at 80% of the new ceiling and add graceful degradation to a cheaper model on quota exhaustion, so exhaustion produces a slower answer rather than a failed one.

Three decisions follow, and they are mutually exclusive enough that hedging is the worst option: buy more seats, tier access to the heaviest users, or absorb roughly 17% less agentic capacity and cut scope. Write the choice into the sprint plan so the slippage has an owner.

The PRD consequence you keep deferring

Every AI feature in design needs four things it probably lacks: metering, soft caps, defined degradation behavior, and a user-visible usage meter. Add a fifth that nobody schedules. It is the communications plan for tightening a limit, drafted before anyone ever loosens one. That message is arriving from a supplier. Assume users will eventually receive one from you.

The best-funded AI vendor on earth cut power users' capacity by 17% and called it a 25% increase, with list price untouched. Build every AI roadmap assuming it happens again.

Outcome pricing is an instrumentation project wearing a pricing decision's clothes

Someone will type "per resolved task" into a pricing page this quarter and assume the hard part was picking the number. Stripe has already published best practices for pursuing outcome-based pricing, and vendors publish best practices when they expect disputes. Treat that as the honest scoping signal for your own version. Before a price can be charged on a completed task, four artifacts have to exist: an unambiguous outcome definition, an attribution model that survives a customer audit, a credit and dispute policy, and the margin tolerance to absorb performance variance previously passed to the buyer. That is one to two quarters of work, and sales will have it on a slide inside three weeks.

Most teams assume the pricing conversation waits for the roadmap. Procurement does not wait. Once a buyer's other AI vendors bill per resolved ticket, the per-seat line becomes the item procurement attacks, and "our AI is different" is not an answer to a unit-of-value objection.

The number that makes the conversation concrete

Terminal-Bench 4.0 changed who reads a leaderboard by publishing resolution rate, tokens and dollars together, per Unwind AI. What teams do with a leaderboard is check whether they moved up a row. What a buyer does with this one is read it as a purchasing table.

GLM-5.3 delivers roughly 81% of the leader's resolution at 45% of the cost, and burns more tokens getting there. Fable 5's overlap with GLM-5.3 inside error bars is the useful discipline here. Three points of headline resolution will not survive a review as a difference.

Where the sources agree, and where they don't

Last Week in AI reports Qwen 3.8, a 27-billion-parameter open-weight model, as rivalling GPT-5.6 and Claude Opus, alongside Thomson Reuters building an in-house model explicitly to cut its Anthropic spend, while Anthropic's annualized revenue surged to $65 billion. Those facts sit together comfortably. The largest payers have found a substitute that fits on hardware they can rent, which turns a renewal into a negotiation instead of a formality.

The divergence is about what cheaper means. TLDR Fintech's "moats in the age of floods" argument holds that model capability is not the bottleneck. Reliable outcomes across messy workflows are, and the argument names pricing against outcomes as a durable moat. Read alongside the token column above, that means a model swap can lower the invoice and leave cost per finished task exactly where it was. The parity claims are reported and vendor-adjacent; treat them as a reason to run your own test, not as a benchmarked result.

The smart move

Pick the single feature where success is objectively verifiable and instrument the outcome event before the pricing page changes. In parallel, put cost per resolved task on the eval scorecard and run the open-weight option against the incumbent on the top three task classes. Reserve the premium model for the one class where ten points of resolution changes what the user actually receives. The forcing function is narrow enough to apply on Monday: a task class that cannot emit an outcome event nobody argues about does not get outcome pricing this year.

Procurement will read that cost column before your AE does, and they will read it out loud in the renewal call.

What the page already knows how to do

A user hands her agent a job the product already performs. There is a search, a cart, a checkout, a booking flow. None of it is written down anywhere a program can read, because it lives inside a layout built for humans. Agents are not failing because they read pages badly; pages were never written for anything but people. WebMCP, walked through by Avi Chawla on August 31, closes that gap by letting a site name each action, describe it in plain English a model can read, and list the typed inputs it accepts.

The implementation cost is close to zero. A tool is a plain object handed to document.modelContext.registerTool(), and on an existing HTML form you add toolname and tooldescription so the browser derives the JSON Schema itself. The execute function calls the same code already sitting behind the button. There is no parallel agent-facing product to build.

The built-in assistant epic is the weakest seat at the table

Chawla ranks six ways an agent can reach an app. Judge each on whose agent does the work, what the user must configure, and what the agent receives on arrival. Raw APIs and backend MCP servers bypass the interface entirely. Computer use guesses from screenshots, and browser automation reads markup that means nothing to it. The built-in site assistant sitting in most backlogs is pitched as ownership. What it actually buys is the bill: you pick the model, you pay the tokens, and because it is not the user's agent, nothing it learns travels with them. WebMCP inverts that arrangement. The user brings the agent and the inference cost, and the interface stays with the team that built it.

The low-risk part is the substrate. JSON Schema tool-calling is already shared across Claude, GPT and Gemini, so describing actions once is not a bet on one vendor.

The two constraints that set v1 scope

One browser family has shipped it, the spec is not final, and per the source, only the browser's own agent can call these tools, which guts the bring-your-own-agent differentiator. Fund it as a spike behind a flag with one escalation trigger: third-party agent invocation shipping.

Integration is trivial. The authorization surface is where v1 gets expensive. Actions run inside the user's authenticated tab, so there is no API key and no token to pass, and the available tools expand with page state, from read-only search when logged out to order history, cart and checkout when signed in. Restated plainly: an agent that reaches your signed-in page inherits purchase authority, and it selects which action to call by reading a plain-English description. A manipulated description or injected page content becomes a path to an unauthorized transaction, and the incident lands on the brand. Two non-negotiables for v1: read-only actions, and an explicit human confirmation gate before any state-changing tool leaves the flag.

The other direction agents are already arriving from

Simplifying AI reports OpenAI's Appshots giving ChatGPT Work and Codex the actual content of whatever app is open — a Slack thread, a form's fields, a dashboard, an iMovie timeline — and letting it act, with no API, partner tier or data agreement. It removes the context tax where users had to describe their screen first. That sets the forcing function for the roadmap: either the agent acts through actions the team described, or through pixels nobody can instrument. One of those shows up in the funnel data. The other shows up in support tickets.

A site that tells an agent what it can do beats a site that makes the agent guess from pixels, and that upgrade costs two HTML attributes and one policy decision about what agents are allowed to buy.

The warning arrived during the growth phase, priced in RMB

A buyer walked into Shenzhen's Huaqiangbei, paid up to 600 RMB above list for a Mac Mini because the thing needed dedicated hardware, then paid a Chinese platform 499 RMB to deploy OpenClaw, which is free. Tencent set up offline booths to help people do the same job. That is what users actually did. What a team tells itself in that moment is that adoption is strong. When a third party can charge real money to finish onboarding, it is quoting the price of a product gap back to you. Per ChinAI's account, that quote showed up months before retention collapsed into the discourse.

Break the adoption path into components. Three high-skill steps: choose the model, connect the agent to your own data and terminals, integrate skills. Shortly after the peak, Codex and Claude Code delivered the same class of capability as a ready-made package, and a DIY assembly framework had no answer to that. Two secondary constraints finished the job. It was designed for 24/7 operation on local hardware, which made inference cost unbounded. That is a diffusion constraint, not just a margin one. It also carried persistent security issues while granting itself data and terminal access. It survived by becoming one-click cloud, with more than 30 derivative products shipping out of Zhipu AI, Tencent and ByteDance.

The benchmark on the other side of the same problem

Graphite's CTO Greg Foster deletes one randomly selected employee's account every day, so everyone re-onboards roughly monthly. Refactoring reports the result: full onboarding, repo linking and settings included, now takes about three minutes. That is a citable time-to-value number for developer tooling, produced with zero tooling budget. Luca Rossi's generalization is the part worth keeping. Whatever users do frequently, the team should do frequently too.

The diagnosis transfers whether or not anyone deletes accounts. Onboarding is the most-used flow in the product and the least-tested by the people who built it. Checked once at launch, then never again, while p99 latency has alerting. The non-destructive version rotates one person per week through a genuinely fresh signup with a written friction log, and captures most of the signal at none of the risk. Pre-commit to eight weeks, because Foster reports the ritual got both love and hate before it got loved.

The measurement failure worth re-reading old docs over

The China-versus-US OpenClaw gap that spawned a policy narrative was wrong. NBC News, citing SecurityScorecard, reported Chinese usage as almost double the US figure, and Council on Foreign Relations researchers built a diffusion-advantage thesis on top of it. Jeffrey Ding queried the same dataset on August 29, 2026 and found 18.7k US instances against 17.0k in China. Near parity, carrying the opposite valence.

Pair that with TLDR IT's Owner.com data point. Its own research found restaurant owners were AI-averse, and that conclusion was obsolete within three months, before the roadmap built on it shipped. AI-sentiment research now has a half-life shorter than a planning cycle. Prioritization arguments being won on first-half research are being argued from expired evidence. The fix is procedural and cheap: a source link, a query date and a named verifier on every quantitative claim, plus an explicit expiry on every capability assumption. Claims missing those do not get a roadmap slot.

When third parties can charge to finish setting up your free product, that price tag is the onboarding work you skipped — and someone else has already collected on it.