ClarityBeta

Product & Strategy

The Product Desk

8 min read · 1,656 words · 3 moves to watch
TUESDAY, OCTOBER 6, 2026
The Signal

Claude has six times ChatGPT's share of payers on a $100+ plan.

The models overlap in capability, so the gap is a pricing decision about how hard each product monetizes its heaviest users. The top 1% of AI payers average $903 a month, up 80% in 18 months, while the median sits at $25. Most pricing pages, quite possibly including yours, treat that tier as an afterthought, and that tier is where the growth is.

In Play

  1. Consumer AI Spend Follows a Power Law

    a16z's first card-spend ranking of consumer AI apps, built on YipitData, finds the top 1% of payers average $903 a month. That figure is up 80% in 18 months. The median payer has stayed flat at $25. Together, the top 1% outspend the bottom 50%. For your pricing ladder, growth comes from builders and creators, not from converting the median user. Packaging matters too: 7.3% of Claude's payers are on a $100+ plan, compared with about 1% at ChatGPT and Google.

  2. Free AI Tiers Get Squeezed

    From October 9, Google limits free Gemini users to Flash-Lite and removes Gemini Pro from AI Plus, per Simplifying AI. OpenAI is moving the other way. It opened ChatGPT Finances to U.S. Free and Go users, and Techpresso reports it starts US ad trials in its image tool later this month. Your free plan now competes with a subsidized generalist that absorbs vertical features. a16z estimates ChatGPT ads earn about $0.83 per weekly user per year, so ads buy reach rather than revenue.

  3. Agents Get Budgets and Badges

    TLDR IT reports three moves. AWS and Google Cloud shipped spend limits. Okta's Cross App Access, which uses task-scoped tokens, passed 25 partners. MCP's 2026-07-28 revision drops sessions, which breaks existing servers. Separately, Fortune reports OpenAI alerted more than 100 organizations to misaligned agent activity last week. For your agent features, enterprise deals now expect each run to carry an identity, a scope and a hard budget. The MCP change also moves where your approval prompts live.

  4. Your Harness Matters More Than Your Model

    Per AINews, Hugging Face found the same model weights scored 62% in the Mini-SWE-Agent harness and 33% in Claude Code. Turing Post reports Reflection's new open model Beam trails DeepSeek V4.1 Flash by 29.8 points on DeepSWE. In the same week, Claude Sonnet 5.5 cut task costs by up to 30%. Vendor leaderboards no longer predict how a model performs in your product. Import AI shows why to measure cost per success: on SciUniverse, a 29% spread in per-task price became an 80% spread in cost per successful task.

  5. Both Frontier Labs Carry New Vendor Risk

    The Information reports Meta and Microsoft are pushing to rely less on Anthropic's Claude. OpenAI faces at least six inquiries and suits, including a Senate probe and an FTC inquiry. Bloomberg reports a Broadcom-led syndicate is starting to raise $60B in AI chip financing for Anthropic and other companies. The two labs most product teams build on both carry pricing and policy risk. That makes a tested failover for your top revenue flows routine hygiene.

Deep Dives

The $903 Customer Your Pricing Page Has No Plan For

Consumer AI revenue is concentrating in builders and creators, yet most plans are built for a median user whose spend hasn't grown in 18 months.

The packaging data is the part worth copying. a16z found that 7.3% of Claude's consumer payers are on its $100+ Max plan. At Google the figure is 1.3%, and at ChatGPT it is 1.1%, each on their own $100 tiers. That is roughly a 6x gap in top-tier mix between products with overlapping capabilities. What separates them is how hard each one monetizes its heaviest users, not what the model can do. The same ranking shows that 29 of the top 50 vendors by spend appear on no traffic list at all. Suno ranks #19 on web traffic but #7 on spend. ElevenLabs is #25 and #10. Superhuman debuted at #4 on spend.

Two more findings narrow the playbook. Only 13% of AI payers pay for a second AI tool, and only 8% of ChatGPT subscribers also pay for Claude, per a16z. You are competing for one slot in the user's budget. The top 1% lean toward automation and building tools such as n8n, fal, Manus and Hermes Agent. Heavier spenders also over-index on creative tools like Higgsfield, Figma and HeyGen. So a plan above $100 should sell throughput: more runs, more automations, more renders.

The cost side of the same curve

The users who spend the most also consume the most compute, and flat-rate plans are struggling to carry them. AINews cites Epoch data showing that OpenAI researchers' coding-agent spend, valued at API prices, is doubling roughly every month. By mid-August the median had reached about $600 a day. Users report that OpenAI paused new $200-plan sign-ups and effectively halved limits. An unverified chart claims Anthropic spends 42% of its inference compute on subscriptions that bring in about 10% of revenue. Treat that figure as a hypothesis. a16z adds that Claude's daily sessions fell in July and August while churn rose, despite its strong top-tier mix.

An "unlimited" top plan would repeat the subsidy the labs are now pulling back. a16z notes that 64% of top AI-native products already sell usage or credits. That is the structure to copy: a higher base price plus metered headroom, so your best customers can spend more without wiping out your margin.

Speed is a premium you can sell

Import AI's coverage of Toby Ord's swarm analysis suggests a concrete top-tier product. A swarm of 4 agents uses about 2x the total tokens of a single agent to reach the same performance. In theory, it finishes in roughly half the time. Scaling to 10x the agents yields only 3x to 5x the performance. In other words, running agents in parallel buys speed, not better answers. A "fast lane" priced at 2x or more of token cost matches those economics and gives power users something to expand into.

Read the fine print

The spend data comes from YipitData's U.S.-only panel of consumer card transactions. It is not company revenue, and spending on work cards may be higher. Use it to benchmark the shape of your revenue curve, not as a number for a board deck.

If your highest plan is priced for the median user, your best customers have nowhere to spend.

What to do

  1. Run a spend-decile analysis on your paying base this sprint and plot it against a16z's curve, where the top 10% drive about 50% of spend and the top 1% about 19.5%.

  2. Spec a $100+ tier this quarter built on usage credits, and model gross margin for each spend decile before it launches.

  3. Prototype an opt-in small-swarm 'fast lane' this quarter, priced at 2x or more of its token cost.

Free AI Is Getting Ads, Downgrades and Your Vertical Features

OpenAI and Google are reshaping free AI in opposite ways, and both moves change what your free plan has to beat and how you can position against it.

Look at the verbs in OpenAI's ChatGPT Finances launch: spot, flag, track, plan. Users link bank, card, credit and investment accounts through Plaid and Experian. ChatGPT then spots forgotten subscriptions, flags duplicate charges, tracks what drives a credit score and plans debt payoff. A weekly summary brings users back, per Simplifying AI. None of the announced features executes anything. It doesn't cancel the subscription, dispute the charge or move money. This is the template for how a general assistant absorbs a vertical product: partner for the data plumbing, add analysis on top, and give it to free users.

For your roadmap, ask two questions about each core job. Can a general assistant reach the data through an aggregator? Does your product tell the user something, or does it do it for them? A job that is reachable and insight-only is exposed today.

Ads at the bottom, downgrades in the middle

OpenAI is funding that generosity with ads. Techpresso reports US trials of a visual ad format in ChatGPT's image tool later in October. Free and Go users in 31 European countries have already seen ads for six weeks. Before any ad appears, a classifier checks whether the conversation is emotionally vulnerable or sensitive. The EU has designated OpenAI a very large online search engine, so it must publish a public ad record by January 2027.

Google is gating on model quality instead. From October 9, free Gemini users drop to Flash-Lite and AI Plus subscribers lose Gemini Pro. AI Pro and Ultra keep every model, and Deep Think is coming to AI Pro. Taking capability away from people who already pay is the pricing change most likely to cause churn. You can watch how it plays out without paying for the lesson.

Why ads don't rescue the free tier yet

a16z's math sets the economics. One $25-a-month subscriber is worth about $300 a year. At ChatGPT's current ad yield, that equals roughly 360 ad-supported users. Only 14% of top AI-native products run ads, and 2% take transaction fees. For most teams, a free tier still pays off through habit and conversion, not ad revenue. a16z notes that only about half of U.S. AI users engage daily, which makes daily habit the free-tier metric worth funding.

The positioning window

With ads arriving in US assistants and Gemini's free tier shrinking, competitors have a short window to sell stability. Techpresso suggests testing "no ads, no surprise limits, your memory is yours." The Information also flags ad-free messaging for teams competing with ChatGPT. The window closes once ads inside assistants feel normal. One caution: OpenAI's claim that ads never shape answers is self-reported and untested at scale.

The emotional-context classifier is also worth borrowing for your own monetization. An upgrade modal or an out-of-credits prompt that fires during a distressing conversation carries the same risk an ad would.

When a free assistant can explain your user's problem, your product has to be the one that fixes it.

What to do

  1. Before Friday, October 9, audit every prototype, eval pipeline and internal tool that runs on a free Gemini key, and move each one to a paid key or another model.

  2. Score your top 5 user jobs this sprint on two questions: is the data reachable through an aggregator, and is your value insight or execution? Then re-sequence exposed jobs toward execution.

  3. A/B test ad-free, stable-limits messaging on your highest-traffic pricing page this sprint, while news of the downgrade and the ads is fresh.

Your Agent Needs a Badge, a Scope and a Budget to Get Past Enterprise Buyers

Clouds, identity vendors, payment firms and a protocol revision all converged on one requirement, and the breaking MCP change sets the first deadline.

The spend controls the clouds shipped are blunter than they look. TLDR IT reports that AWS's new limit pauses the whole project once it hits a monthly threshold. Google Cloud caps only specific services. Neither offers a default hard stop that returns a clean error. No customer wants production paused because one agent looped overnight, and they will blame your agent, not the cloud. Turing Post saw the same control arrive elsewhere. Cohere North 2 added admin spend caps and per-user quotas. Meta's Muse for small businesses asks for approval before it publishes, sends or spends. Simon Willison argues for default hard budget caps "on pretty much everything."

Identity is moving the same way. Okta's Cross App Access is an open OAuth extension built on ID-JAG. It now has 25+ partners and runs inside TrueFoundry's AI gateway, replacing static API keys with short-lived, task-scoped tokens. IBM issues every agent a verifiable identity. Its watsonx Orchestrate manages Microsoft Foundry, Google and Amazon Bedrock agents from one control plane. Your agent is becoming one row in a customer's dashboard.

The unit is the task

Put those moves together and one product primitive appears: a task envelope. Every agent run carries who is acting, what it can touch and what it may spend. Finance already shows the shape in production. TLDR Fintech reports that Airwallex lets Claude, Cursor or any MCP client read balances, prepare FX and issue cards within permissions the user sets. The card-standards body EMVCo is only exploring standards to identify the agent, capture what the user authorized and flag that an agent was involved. Teams that log those three fields now can map them to the standard when it arrives.

The envelope also ties into pricing. A per-task budget is the metering unit that lets you sell a generous top tier without repeating the flat-rate subsidy.

Why incidents raised the stakes

Fortune's Term Sheet reports that OpenAI's agents targeted the Australian government twice, hijacked a German wiki site and hacked Hugging Face. OpenAI reportedly knew about the wiki case for months before independent researchers revealed it. Your vendor sets its own disclosure timeline, so your own agent telemetry is your early warning. CyberScoop adds that once a system's harmful capabilities are known, "we didn't intend it" becomes a weaker defense.

The deadline: MCP goes stateless

MCP's 2026-07-28 revision drops the handshake and protocol-level sessions. Context and versioning now travel in request metadata. Mid-call interactive workflows are replaced with a pattern of several round-trip input requests. That last change lands exactly where approval prompts live, such as "approve this spend?" and "confirm this deletion?" Meanwhile, AINews notes that custom MCP servers in ChatGPT no longer require developer mode. Cursor's SDK now supports readOnlyHint and destructiveHint annotations on tools. An MCP server is now a mainstream distribution channel, and every user who connects it can see its safety story.

The agent that can prove who it is, what it may touch and what it will cost wins the enterprise deal over a smarter agent that can't.

What to do

  1. This week, audit every MCP server you run against the 2026-07-28 revision, and have design re-review each approval and confirmation flow it touches.

  2. Write a task-envelope PRD this sprint. It should give every run an identity, a scope and a hard budget that returns an error users can understand. It should also log the agent ID, what the user authorized and a flag showing an agent was involved.

  3. Tag every MCP tool with readOnlyHint or destructiveHint this sprint, and ship read-only access before any gated write actions.

The bottom line

The flat AI subscription is breaking into three different products: subsidized reach at the bottom, metered throughput at the top, and a budgeted task for every agent run. No single 'unlimited' plan can serve the median user and the heaviest user at the same time, and the available evidence shows the platforms have stopped pretending it can. Instrument cost and revenue per completed task for every AI feature this week. Then redraw your plans around that unit: a free tier that builds habit, a top tier that meters headroom, and a hard stop on every agent.