👋 Hey, Egemen here.
Anthropic says Claude Opus 5.5 costs 40% less to run than Opus 5. Ramp’s lead economist measured a 41% drop in token prices since March, so I checked which startups keep the savings. One AI data company grew more than 18x after it changed what it sells.
Here’s a snapshot of what’s on the menu today:
💡 Spotlight: For Product teams moving at AI Speed
🧠 Deep-Dive: The token savings go to finished work
🗺️ Method: Hire anyone, anywhere
⚾️ Catch: Opus 5.5 one shots everything
☝️ Scaled This Past Week: Numeral
💡 Spotlight
Built for Product Teams moving at AI Speed.
Your teams are moving fast, burning tokens, and shipping more than ever.
But more output doesn’t mean more impact.
Jira Product Discovery brings your ideas, customer insights, and business context together so product teams can weigh the evidence, make the tradeoffs, and decide what’s actually worth building. Then connect those decisions directly to delivery in Jira, so everyone knows what you’re building and why.
Jira Product Discovery. For better product decisions in the AI era.
Presented by
🧠 Deep-Dive: The token savings go to finished work
Anthropic released Claude Opus 5.5 on 22 September. The company says it costs 40% less to run than Opus 5 on typical workloads. Ramp sees the same trend in business spend. The blended price per million tokens fell 41%, from $1.15 on 1 March to $0.68 by early September.
I used to count each model price cut as good news for AI startups. Now I split it by how the startup charges.
If you sell credits priced from your model bill, your customers expect each cut to reach them. Your revenue per task drops, so volume has to grow faster than prices fall. Ramp’s lead economist sees that risk for the model labs. He also reports that some businesses now set company-wide defaults that cut the use of frontier models.
If you charge a fixed price for a finished job, the same cut goes to your gross margin. You choose the model, so you keep the savings.
41%
Drop in the blended price per million tokens on Ramp, from the March peak to early September.
Anthropic says Opus 5.5 costs 40% less to run than Opus 5.
(Source: Ramp AI Index, “Cracks in the AI Thesis Part 2”, 9 September 2026)
(Source: Anthropic, “Introducing Claude Opus 5.5”, 22 September 2026)
Seven-day trailing, token-volume-weighted price across input, output and cached tokens. Weekly points come from the data behind Ramp’s chart. The Snorkel figures come from its CEO’s post below.
Snorkel AI shows what happens when you sell the finished job. For years it sold software that helps teams label training data. In September 2025 it started to sell finished datasets and training environments to frontier labs and other model builders.
CEO Alex Ratner wrote on 22 September that Snorkel has grown more than 18x since that launch. The company crossed a $375M annualized run rate the same week and raised $350M at a $3.5B valuation.
My read is that Snorkel now competes for the buyer’s budget for expert work. As a tool vendor, it competed with other software. Snorkel describes its newer data as expert tasks that can take days or weeks. The buyer pays for the finished result, and Snorkel picks how much of each task runs on its own models.
(Source: Snorkel AI, “Data 2.0 and the research era of AI data”, 22 September 2026)
(Source: TechCrunch, “Snorkel AI triples valuation to $3.5B as demand for AI training data booms”, 22 September 2026)
👉 Check who sees your token bill. If your customers buy credits priced from your model costs, pick one workflow and price it per finished job this quarter. Set that price from what the customer pays a person to do the job today. If you already charge per job, rerun your cost per job each time a model ships. Hold the price and spend the savings on the step where an error costs your customer the most.
🗺️ Method
Hire anyone, anywhere — compliant in under 3 days
Found the right hire, but no entity in their country? Remote becomes the legal employer — with contracts, benefits, and tax setup handled, plus direct access to the same in-house team that runs payroll locally.
⚾️ Catch: Opus 5.5 one shots everything
☝️ Scaled This Past Week: Numeral
San Francisco, CA / AI sales tax compliance / announced 23 September 2026
Numeral has raised a $100M Series C. Insight Partners led the round, and Salesforce Ventures, Geodesic, Benchmark, Mayfield, FCVC, Y Combinator and Uncork joined. CEO Sam Ross and CTO Matt DuVall founded the company in 2023. Numeral has raised $157M in total, including a $35M Series B in September 2025.
Numeral runs sales tax for companies that sell across states and countries. It tracks where you owe tax, registers you, calculates the tax, files the returns and pays the states. It also covers VAT and GST in more than 90 countries. Total transaction volume grew 327% year over year.
If you sell SaaS to customers in California, put 1 January 2027 in your plan. California enacted SB 122 in June. From that date, the state applies sales and use tax to prewritten software and SaaS.
Numeral charges $75 per filed return and $150 per state registration, with no contract. If Numeral files late, it pays the penalties and interest.
What to copy
Numeral prices the finished job and backs it with money. You pay for each return it files, and Numeral pays the cost of a late filing. If you finish a task end to end for your customer, charge per finished task and pay for your own errors. With a guarantee, you show buyers that you trust your error rate. Your sales team also gets one simple promise to sell.