Week of October 25, 2026

Pricing consultant Ulrik Lehrskov-Schmidt finds that 1% of users often drive 50 to 75% of AI costs, while about half barely use AI. One client's worst customer burned 380x its revenue in tokens. He argues for fair usage brakes that touch a few percent of users, which recovered up to 22 margin points in weeks. Figma and Clay used usage credits and AI tiers on existing SaaS to capture value that seat pricing missed.

OpenAI is bringing ChatGPT ads to 26 more countries, for 89 in total, with free and Go tier users seeing labeled ads kept apart from answers. Annualized ad revenue climbed from about $100 million in March to $1 billion by August, and over 100,000 advertisers have bought campaigns. It shows ad-supported AI growing alongside the consumer subscription tiers, while rival Anthropic promises to keep Claude ad free.

Notion now gives each Business seat two separate monthly AI wallets: 1,500 Standard credits for work where Notion picks the model, and 500 Select credits for user-chosen models and Custom Agents, plus a shared workspace pool to top up either. The author traces Notion's path from an $8 AI add-on to bundled plans to usage credits, and argues that splitting balances by who picks the model confuses buyers.

Gainsight is dropping seat-based pricing for Gainsight CS from next year and will instead charge based on customer annual recurring revenue and the number of accounts managed, arguing seats no longer fit when agents and people share the work. Staircase AI is bundled in, MCP access is free, and its Atlas unit sells AI-native services under results-based contracts, now covering over $75 million in customer ARR. The shift reworks how a SaaS vendor packages AI.

Since October 12, 2026, Cursor sells xAI's X Subscription: Plus at $8 a month covers Grok models only, while Premium at $30, Super at $100 and Ultra at $200 multiply one shared weekly pool by 4x, 15x and 50x. Claude, GPT and Gemini sit outside that credit pool on every tier and are billed at API list price. Bugbot code reviews also dropped a $40 seat fee for per-review pricing of about $1 to $1.50.

ChatGPT's advertiser count roughly tripled to about 60,000 between late August and early October, yet OpenAI looks set to miss its projected $2.4 billion in 2026 ad revenue. Clicks cost around $5, far below the $30 to $50 Google charges for comparable software ads, after OpenAI dropped its $200,000 minimum and added CPC bidding and self-serve buying. The piece shows that scaling ad-supported AI inside a consumer AI subscription product needs sustained budgets, not just cheap tests.

Gorgias's pricing lead argues that AI agent pricing remains unsettled. Per-resolution billing, used in outcome-based pricing, ties margin to resolution rate: if resolutions slip while LLM spend holds, each resolution costs more. Per-conversation, cost-plus and seat-based SaaS pricing each bet on a different idea of value. Tracking cost by feature cohort uncovered one money-losing feature, and moving it to a smaller model saved roughly $800,000 a year.

GitHub Copilot now meters chat, agents and CLI through AI Credits worth $0.01 each, with Pro at $10 bundling 1,500 credits, Pro+ at $39 giving 7,000 and Max at $100 giving 20,000. Higher tiers return more usage per dollar, while inline completions stay free. It is a clear case of subscription for usage credits layered onto an AI add-on to existing SaaS, with Business and Enterprise seats feeding a shared pool billed for overage.

Flat seat pricing breaks for AI agents because one heavy user can cost 50 times more in compute than a light user. The article's example is a $40 seat that costs $2 to serve for one developer and over $100 for another. It recommends a hybrid: a seat fee with a usage allotment sized near the median user, then metered overages, like an AI tier on existing software. It also points to Intercom's Fin, which charges about $0.99 per resolution as outcome-based pricing.

Anthropic's draft IPO filing, as reported by the WSJ, shows $16.3 billion in first-half 2026 revenue, up from $1.1 billion a year earlier. Compute spending hit nearly $9 billion in six months, with about $450 billion in cloud commitments locked in over ten years. Two customers supplied roughly 20% of revenue, a sign of how concentrated selling model access through APIs can be at frontier scale.