{"slug": "anthropic-is-subsidizing-our-ai-coding-at-13x-how-long-will-it-last", "title": "Anthropic is subsidizing our AI coding at 13x. How long will it last?", "summary": "Anthropic is subsidizing AI coding costs for bundled-plan customers by an average of 13x, according to an analysis by Upbound, a company on Anthropic's Team plan. Upbound engineers using Claude Code on a $125/month bundled seat averaged $5,500/month in token costs when priced at API list rates, with the heaviest user reaching $6,470 — 52x the seat price. The subsidy multiple, calculated from June usage logs for 20 engineers, ranges from under $10 to $6,470, and Upbound warns the cheap bundled deal may not last as Anthropic moves Enterprise customers to usage-based billing.", "body_md": "We are on Anthropic's Team plan at Upbound, and we are about as AI-pilled as a company gets. All of our engineers write code with Claude Code every day, on a bundled seat. It's a flat monthly price with limits, and you don't really see the token meter.\n\nThen one of our engineers switched to opencode and started running Opus 4.8 straight through the API instead of on his bundled plan. I looked at his usage a few weeks in. He was averaging about $5,500 a month. Same person, same work, roughly what he had been doing on the bundled plan at $125/mo. The only thing that changed was that we could suddenly see the token meter.\n\nThat made me want to know what it would cost us if everyone moved off bundled seats and\nonto per-token pricing. It is not just hypothetical. Anthropic\n[moved its Enterprise customers to usage-based billing](https://www.theregister.com/2026/04/16/anthropic_ejects_bundled_tokens_enterprise/)\nearlier this year, tokens on top of the seat, and\n[shut off bundled-seat usage from other coding agents](https://thenewstack.io/anthropic-claudecode-opencode-split/).\nAs the labs start optimizing for margins, that cheap bundled deal will not last, and I\nwanted to know how exposed we are.\n\n## It's not easy to see the tokens on a bundled plan[#](#its-not-easy-to-see-the-tokens-on-a-bundled-plan)\n\nI was surprised by how hard it is to see the tokens consumed on a bundled plan. The analytics dashboards do not report tokens. The usage API that would report them needs an Enterprise account, as far as I can tell, which we do not have.\n\nLuckily, Claude Code writes a local log of every session, with token counts per message,\nand keeps roughly the last thirty days before it prunes. We wrote a small script that\nreads those logs and totals the tokens by model, priced at Anthropic's published rates.\nIt emits only counts and costs, no prompts and no code, so people can run it and share\nthe output without leaking anything. The script is\n[here](https://gist.github.com/bassam/27fa951ac01e0998b73b36d90297678b), and it runs on\nyour own machine in about a minute.\n\nAfter we built ours, someone pointed me to [ccusage](https://ccusage.com), an open-source\nCLI that reads the same local logs and reports usage and cost across several coding\nagents. I had not seen it before writing our own, and it will get you to the same numbers.\n\nSo we pulled the June numbers for 20 of our engineers.\n\n## The subsidy multiple[#](#the-subsidy-multiple)\n\nPriced at list, the engineers' Claude Code usage ranged from under $10 to $6,470. More than 90% of it is Opus, mostly Opus 4.8. These are floors, because the local logs had already pruned part of the month for most people by the time we looked. A few people were also out on PTO for a few weeks in June, so their months are partial, and we counted everyone anyway, which pulls the average down, not up.\n\nSet that against Anthropic's Premium seat at $125 a month, and that's an average of 13x\nmore expensive if we switched to paying for tokens. The average covers a wide range: the\nmedian engineer was about 7x, some barely touched it, and our heaviest ran $6,470 in\nJune, 52x a single seat. I have started\ncalling this the **subsidy multiple**: the tokens you burn in a bundled plan, priced at list, divided by\nwhat you pay for the seat.\n\nA few caveats: all prices are at list, not at a rate a big customer negotiates. A large share of it is cache reads from long agentic sessions, which bill at a tenth of the input price, so someone will argue the cost is softer than it looks. A bundled plan is also not all-you-can-eat: past the included limit you either get throttled or, with extra usage turned on, pay the overage at API rates. We have it on and barely touch it, a few hundred dollars a month, even as hard as we lean on Claude Code. Even after all of that, the shape does not move.\n\n## We're not the only ones seeing this[#](#were-not-the-only-ones-seeing-this)\n\nI sat on these numbers for a bit, wondering if we were an outlier, and then a couple of\nmuch larger companies put the same thing on the record. Uber's CTO\n[told The Information](https://www.theinformation.com/newsletters/applied-ai/uber-cto-shows-claude-code-can-blow-ai-budgets)\nthey burned through their entire 2026 AI budget in four months, as Claude Code spread\nfrom 32% to 84% of a roughly 5,000-engineer org. The average engineer ran $150 to $250 a\nmonth, power users $500 to $2,000, and he mentioned spending $1,200 himself in a single\ntwo-hour session. They have\n[since capped engineers at $1,500 a month](https://techcrunch.com/2026/06/02/uber-caps-employee-ai-spending-after-blowing-through-budget-in-four-months/).\nNothing broke, and no one was gaming it. They used Claude Code for exactly the work it is\ngood at, and the bill still blew up. That is the part that matches our data: nothing went\nwrong to produce these numbers. Claude Code working as intended is what produces them.\n\nTesla did the same thing a few weeks ago,\n[capping employees at $200 a week](https://electrek.co/2026/07/02/tesla-caps-employee-ai-spending-200-week/)\nafter some engineers were running up thousands of dollars in tokens weekly. What stood\nout to me is that the same company had been ranking engineers on internal leaderboards by\nhow many tokens they burned, to push adoption. So it went from gamifying token\nconsumption to metering it, and Meta, Amazon, and Walmart have all moved the same way,\ncapping usage or steering engineers to cheaper models. The pattern is consistent once\nusage-based billing makes the per-prompt cost visible at the individual level: the bill\nshows up, and the open tap gets a valve on it. Our 13x is that same wall. We just measured\nit before anyone made us.\n\n## Owning our intelligence[#](#owning-our-intelligence)\n\nSo we are doing the thing that follows from taking that seriously. Open-weight models are\nclosing on the frontier faster than anyone expected, and developers are voting with their\ntokens: on\n[Vercel's AI Gateway](https://vercel.com/blog/ai-gateway-production-index-july-2026),\nopen-weight models went from 11% of volume in April to 29% in June, on under 4% of spend.\nWe've started looking at running our own models, where the cost is compute we control\nrather than a token price someone else sets and can reset whenever the market lets them.\nIt also keeps our prompts and context ours. Satya Nadella\n[made the point well](https://techcrunch.com/2026/07/13/satya-nadella-has-issued-a-shocking-warning-to-companies-using-ai/):\nwith someone else's model you \"pay for intelligence twice, once with money, and again with\nsomething even more valuable: the proprietary knowledge you must reveal to make that\nintelligence useful.\"\n\nA big correction is coming in AI, and that's why we built\n[Modelplane](https://github.com/modelplaneai/modelplane), a control plane for inference,\nin the open. I'm curious what your subsidy multiple is.", "url": "https://wpnews.pro/news/anthropic-is-subsidizing-our-ai-coding-at-13x-how-long-will-it-last", "canonical_source": "https://modelplane.ai/blog/ai-coding-subsidy-multiple", "published_at": "2026-07-23 00:00:00+00:00", "updated_at": "2026-07-23 13:14:52.799217+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-products", "ai-policy", "developer-tools"], "entities": ["Anthropic", "Upbound", "Claude Code", "Opus 4.8", "Bassam", "ccusage"], "alternates": {"html": "https://wpnews.pro/news/anthropic-is-subsidizing-our-ai-coding-at-13x-how-long-will-it-last", "markdown": "https://wpnews.pro/news/anthropic-is-subsidizing-our-ai-coding-at-13x-how-long-will-it-last.md", "text": "https://wpnews.pro/news/anthropic-is-subsidizing-our-ai-coding-at-13x-how-long-will-it-last.txt", "jsonld": "https://wpnews.pro/news/anthropic-is-subsidizing-our-ai-coding-at-13x-how-long-will-it-last.jsonld"}}