The news
AI usage-based pricing is reaching three budget lines at once. On September 27 and 28, 2026, SaaStr highlighted $213-a-week AI coding bills and new vendor fees for agent access, and The Register reported that Microsoft (MSFT) is making usage-based billing central to advanced Copilot features.
On coding, AI spend-tracking startup Larridin published a benchmark in August 2026 built from provider billing feeds for engineers who both merged code and drew billed AI-coding spend in the four weeks ending August 2, 2026. The median engineer billed $213 a week, or about $920 a month, according to Larridin. The 25th percentile was $86 a week and the 90th percentile $911, a spread of more than 10 times. Larridin did not disclose its full sample size.
Larridin also reported that returns vary with how AI-native an engineer is. Its most AI-native engineers reached 11.8 times the output of that cohort's lightest spenders, as Larridin measures it, at around $1,300 a week and had not hit a ceiling, while for partial adopters the marginal payoff dropped by about half once spend passed roughly $600 a week. Those comparisons rest on small cohorts of 27 to 46 engineers each, and both AI-native cohorts come from a single company, according to Larridin. SaaStr, which featured Larridin on September 27, reported that the company raised a $17 million seed round led by Andreessen Horowitz.
On legacy SaaS, SaaStr founder Jason Lemkin wrote on September 28 that almost every pre-AI vendor SaaStr uses is raising prices for agent access. He cited Atlassian (TEAM) Rovo credits. Atlassian's support documentation says calls through its Rovo Model Context Protocol (MCP) server, the standard that lets AI agents connect to apps, consume credits, and that overage billing starts December 3, 2026 at $0.01 per credit, with 10 credits for a basic billable event. Lemkin wrote that HubSpot (HUBS) is raising prices mainly for its own agents, including custom-agent metering that he said began in July, while its MCP server stays free for outside agents.
Lemkin wrote that Salesforce (CRM) sells extra API capacity at about $83 per million calls, and estimated that a proposed agent meter would land between $5,000 and $100,000 per million. That range is his estimate, not a published Salesforce price. He said SaaStr's own AI agent makes more than 30,000 API calls a day.
On office software, The Register reported on September 28 that Microsoft outlined a two-part Copilot model: a per-user subscription license for everyday features such as Copilot Chat and Copilot in Word, Excel, Outlook and Teams, plus usage-based billing in Copilot Credits for advanced features including Cowork, Code and Autopilot. The Register said the usage charges sit on top of subscriptions and did not report credit prices. It noted that the meter itself is not entirely new, since Microsoft already bills some AI services, including Cowork, by consumption; what changes is how central usage-based billing becomes.
The numbers
- Median billed AI-coding spend per engineer (Larridin, 4 weeks to Aug 2, 2026)
- $213 a week
- 90th percentile AI-coding spend (Larridin)
- $911 a week
- Atlassian Rovo overage rate from Dec 3, 2026 (per Atlassian)
- $0.01 per credit
- Salesforce extra API capacity for integrations (per SaaStr)
- About $83 per million calls
- Daily API calls by SaaStr's own AI agent (per SaaStr)
- 30,000+
Why CEOs should care
For CFOs, the issue is that each meter sits in a different budget: engineering tools, application renewals and the Microsoft 365 agreement. Put one owner over all AI consumption spend and ask for a monthly view by team. Microsoft's documentation says its Copilot cost management dashboard supports organization-level and user-level spending limits, alerts and read-only reporting roles for finance staff. It also says spending policies apply automatically to new supported Copilot services by default, a setting admins can switch off if they want to review new services first.
For CTOs and engineering leaders, Larridin's spread of more than 10 times between the 25th and 90th percentiles means averages hide the real budget. Set per-engineer or per-team token budgets tied to an output measure, and review heavy users rather than capping everyone. Larridin's cohorts are small, so if its finding holds in your data, spend above roughly $600 a week for partial adopters deserves a closer look.
For CIOs and procurement teams, agent access is now a renewal term. Before signing, get in writing how the vendor charges for agent and API calls, whether reads are free, what happens to MCP access, and whether caps exist. Lemkin said SaaStr will not choose another vendor that charges a significant tax on its agents; buyers can use the same position in negotiations.
The bigger picture
Software buyers spent two decades budgeting per seat. According to TechMarketView analyst Simon Baxter, cited by The Register, Microsoft's model will require a mindset shift for organizations used to knowing what their Copilot seats cost; he said customers wanting agentic features will need a consumption mindset and cost governance. Lemkin argued that heavy agent pricing could backfire on incumbents: customers route around the meter, build agents on other systems and shrink the vendor's footprint at renewal. That is his forecast, not an observed trend.
What’s next
Atlassian's Rovo overage billing starts December 3, 2026, according to Atlassian's support documentation. Microsoft says it will add more agents and services to usage-based billing over time, and The Register did not report Copilot Credit rates, so buyers should ask their Microsoft account teams for the rate card and use Microsoft's Copilot Credit Estimator before enabling advanced features widely.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error
Companies in this story
