ClarityBeta

Product & Strategy

The Product Desk

12 min read · 2,460 words · 4 moves to watch
The Signal

Anthropic's September 14 'increase' cuts Claude Code capacity 17% for heavy users.

The permanent 25% bump sits below the temporary 50% boost that expires September 13, so the invoice tells you nothing. Any squad already running above 83% of today's cap breaks mid-month, and the loss surfaces as sprint carryover rather than a vendor line item anyone can point at. Pull the per-seat usage distribution before the retro decides the team just got slower.

In Play

  1. Claude Code Capacity Drops Mid-September

    Anthropic will raise Claude Code weekly limits a permanent 25% above baseline on September 14, and the temporary 50% boost expires September 13. TLDR IT and Unwind AI both compute the same result: roughly 17% less effective capacity for heavy users, with list price unchanged. That means the loss lands in velocity rather than in a budget review. Techpresso reads it as a capacity constraint; TLDR IT reads it as inference-margin management.

    Ask Clarity
    Sign in to ask more about this

  2. Outcome Pricing Comes From The Category Leader

    OpenAI has begun letting some major customers pay only when its AI completes a task, with a handled support interaction as the stated example, per The Information's reporting. Salesforce and other providers are moving the same way, and Stripe has already published best-practice guidance for avoiding outcome-pricing disputes. Once a buyer's other AI vendors bill per resolved item, your per-seat line becomes the thing procurement attacks at renewal.

    Ask Clarity
    Sign in to ask more about this

  3. Benchmarks Now Publish Cost Per Resolved Task

    Terminal-Bench 4.0 reports resolution rate, token consumption and dollar cost together, per Unwind AI: Opus 5 leads at 51.8% resolution on 6.5B tokens for about $6.0K, while open-weight GLM-5.3 hits 41.8% on 8.7B tokens for about $2.7K. Fable 5's 44.5% overlaps GLM-5.3 within error bars at higher cost. A leaderboard that prints dollars is a procurement document, and your CFO will find it whether or not you do.

    Ask Clarity
    Sign in to ask more about this

  4. Agents Reach Your Product Without Your API

    Chrome and Edge shipped WebMCP, which lets a site declare typed actions an agent can call; on an existing HTML form that is two attributes and no JavaScript, per Daily Dose of Data Science. From the other direction, Simplifying AI reports OpenAI's Appshots gives ChatGPT Work and Codex the content of whatever app is open and lets it act, with no API, partner tier or data agreement. Either way, agents are already completing jobs inside your UI.

    Ask Clarity
    Sign in to ask more about this

  5. Packaging, Not Capability, Decided The Agent Market

    OpenClaw was one of GitHub's fastest-growing projects from November 2025 to March 2026 and by August 2026 was described by QbitAI as over, per ChinAI's post-mortem. The cause was three high-skill setup steps, not model quality: pick the model, wire your own data and terminal access, assemble skills. Codex and Claude Code shipped the same capability packaged. Chinese platforms charged 499 RMB to deploy a free product during its growth phase.

    Ask Clarity
    Sign in to ask more about this

Deep Dives

  1. The '25% Increase' That Removes 17% Of Your September Capacity

    Entitlement changes never appear in a budget review; they surface as unexplained sprint carryover, and this one has a date on it and no owner assigned yet.

    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…

    3 action items

    ●

  2. Outcome Pricing Arrives While Benchmarks Start Printing Dollars

    The category leader absorbing failure risk and a leaderboard publishing cost per resolved task are the same event: the unit your buyer pays for is shifting from access to finished work.

    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…

    2 action items

    ●

  3. Chrome And Edge Shipped WebMCP; Your Site Can Now Tell Agents What It Does

    The marquee benefit, bring your own agent, is not live yet, which sets the correct scope: a flagged read-only spike plus a written policy on what an agent may buy.

    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…

    3 action items

    ●

  4. OpenClaw Died In Five Months Over Three Setup Steps

    Its autopsy and a public three-minute onboarding benchmark point at the same unmeasured surface, and one of them was for sale at 499 RMB while the product was still growing.

    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…

    3 action items

    ●

The edition continues

Take the signal into the room.

Sign up or log in to read all 4 deep dives in full, plus the final take.

Read the full edition

Continue with LinkedIn