- AI coding agents like Claude Code and OpenAI Codex are now billed at full API token rates for enterprise customers.
- AI coding agents burn vastly more tokens than chatbots, making them far more lucrative for OpenAI and Anthropic.
- Anthropic is rumored to be approaching its first profitable quarter as enterprise AI spending accelerates sharply.
- OpenAI has 703 open roles right now, with nearly a third tied to enterprise sales and go-to-market functions.
- AI coding agents like Claude Code and OpenAI Codex are now billed at full API token rates for enterprise customers.
- AI coding agents burn vastly more tokens than chatbots, making them far more lucrative for OpenAI and Anthropic.
- Anthropic is rumored to be approaching its first profitable quarter as enterprise AI spending accelerates sharply.
- OpenAI has 703 open roles right now, with nearly a third tied to enterprise sales and go-to-market functions.
AI Coding Agents Are Finally Making Real Money
For the past three years, the story of OpenAI and Anthropic has been one of staggering costs chasing stubborn losses. That story may be changing. AI coding agents — specifically Claude Code from Anthropic and Codex from OpenAI — appear to be the product that actually closes the gap between astronomical infrastructure spend and meaningful revenue. And a pair of quiet but significant pricing changes in April 2026 just made that equation a lot more interesting.
Anthropic is strongly rumored to be on the verge of its first profitable quarter. Companies are reportedly getting hit with LLM bills they didn’t anticipate. Neither of these things is a coincidence.
The Pricing Shift Nobody Was Talking About
Here’s what happened. For a long time, Anthropic’s enterprise plan came with a relatively generous usage allowance — the pitch was essentially that Claude seats included enough usage for a typical workday. That changed sometime around November 2025, when Anthropic quietly restructured enterprise pricing to $20 per seat per month, plus full API token pricing on top of usage. Many existing customers are only finding out now, as their annual contracts come up for renewal.
OpenAI followed the same playbook. On April 2, 2026, the company updated Codex pricing to align with API token usage rather than the old per-message model. Initially this applied to new and existing Plus, Pro, and ChatGPT Business plans, as well as new enterprise contracts. By April 23, all existing ChatGPT Enterprise customers — including Edu, Health, Gov, and ChatGPT for Teachers — were brought onto the new structure.
Both companies also released new flagship models in April at higher price points. GPT-5.5, which shipped April 23rd, costs twice as much per API token as GPT-5.4. Anthropic’s Opus 4.7, released April 16th, runs roughly 1.4 times the cost of Opus 4.6 once you account for changes in their tokenizer. So within the same month: enterprise contracts repriced at full API rates, and the API rates themselves went up. That’s a significant double move.
The Consumer Subsidy That Was Hiding in Plain Sight
To understand why this matters, it helps to look at what power users of AI coding agents are actually consuming — and what it costs at API rates versus flat subscription pricing.
Developer and blogger Simon Willison ran the numbers using a tool called ccusage that estimates token costs from local usage logs. Over a 30-day period, his usage of Claude Code would have cost $1,199.79 at API prices, and his Codex usage would have come to $980.37 — a combined $2,180.16 in tokens. He’s paying $200 a month in subscriptions. That’s roughly an 11x subsidy.
His conclusion: he’d assumed enterprise customers running agents at scale were getting similar deals. They weren’t — or at least, they won’t be anymore. The consumer and prosumer tiers are still heavily subsidised. The enterprise tier is now, essentially, cost-plus.
That gap reveals something important. AI coding agents don’t just produce better output than a simple chatbot interaction — they burn tokens at a fundamentally different rate. A back-and-forth conversation uses a modest amount of compute. An agent autonomously writing, testing, debugging, and iterating through a codebase uses orders of magnitude more. That token intensity is exactly what makes these tools expensive to run and, now, expensive to buy at scale.
Why Coding Agents Crack the Revenue Problem ChatGPT Couldn’t
ChatGPT’s user numbers are genuinely impressive. In February 2026, OpenAI reported more than 900 million weekly active users. But only around 50 million of those — about 5.6% — were paying subscribers. At $10 to $20 a month each, that’s a real business, but it’s nowhere near the scale needed to justify the company’s infrastructure commitments, let alone its rumoured $1 trillion long-term capital requirements.
AI coding agents flip that dynamic. The people using these tools most heavily are professional software engineers — some of the best-compensated knowledge workers on the planet. Their employers have both the budget and the incentive to pay for tools that genuinely accelerate their output. And unlike consumer ChatGPT users who might dip in and out, developers are integrating these tools into their daily workflow. The usage is sticky, intensive, and billable at API rates.
Willison argues that November 2025 was the real turning point — the moment when model quality crossed a threshold that made agents genuinely reliable for professional use. Six months of adoption later, companies have started budgeting for this seriously. The surprise bills rolling in now aren’t a sign that AI is overhyped; they’re a sign that it’s actually being used.
There’s also a broader implication worth sitting with. A coding agent isn’t just a tool for writing software. It’s a tool for automating anything you can accomplish by typing commands into a computer. As these systems get better, the addressable market expands well beyond software engineering — into legal research, financial analysis, data work, content operations, and anywhere else that knowledge work is currently done through a screen. The revenue model being built on AI coding agents today is probably the template for enterprise AI more broadly, not a niche carve-out.
The Irony in the Hiring Numbers
There’s a telling footnote to all of this in the job listings. Willison scraped the careers pages of both companies (using Claude Code, naturally) and found that OpenAI currently has 703 open positions, of which roughly 229 — about 32.6% — are enterprise-facing roles: account executives, go-to-market specialists, forward-deployed engineers. Anthropic has 390 open roles, with around 105 (26.9%) in the enterprise bucket.
So the companies building the most talked-about automation technology in the world are currently hiring humans at a serious clip to sell it. Enterprise software contracts, it turns out, still require a substantial number of people in meetings, on phone calls, and managing renewals. The irony is real, but it’s also just how enterprise software works — and the fact that Anthropic and OpenAI are now building out these sales motions at scale is itself a signal. You don’t invest in enterprise go-to-market infrastructure unless you believe the product is ready and the deals are closeable.
What Comes Next
Both companies are eyeing IPOs, and the timing of these pricing moves is unlikely to be accidental. Profitable quarters and growing enterprise revenue tell a very different story to public market investors than loss-making growth. Anthropic’s rumoured profitability milestone, if confirmed, would be a significant moment — not just for the company, but for the broader narrative around whether frontier AI labs can ever build sustainable businesses.
The bet OpenAI and Anthropic are now making is that AI coding agents are sticky enough, and valuable enough, that enterprise customers will absorb the repricing rather than defect. Given that these tools are already embedded in daily developer workflows at major companies, that’s probably a reasonable bet. The question is how quickly competitors — Google’s Gemini, open-weight models, and a growing crop of specialised coding tools — can offer credible alternatives at lower price points. The window for extracting premium pricing from a captive enterprise base never stays open forever.
Source: https://simonwillison.net/2026/May/27/product-market-fit/
