In June 2023, the developer of Apollo — a widely used Reddit client — calculated that Reddit's newly announced API pricing would cost his app roughly $20 million a year. Apollo shut down on June 30. The API still existed. The endpoints still worked. What changed was the price Reddit put on the relationship those endpoints represented.
An API — an application programming interface — is a structured way for one piece of software to request data or actions from another. Your phone's weather app calls a weather service's API. Between roughly 2010 and 2018, the expectation was that this would become the universal way machines talked to each other. A 2015 Harvard Business Review article compared a company without APIs to the internet without the Web. ProgrammableWeb's directory tracked over 19,000 by early 2018.
What providers discovered in the years after was what each interface actually costs to keep open, and they began rationing access according to who justified the expense.
Reddit rationed by price. Spotify rationed by capability: in late 2024 it removed recommendation and audio-analysis endpoints from new developer applications while previously approved apps kept them, and by February 2026 new accounts needed a Premium subscription and were capped at five authorized users. X/Twitter introduced paid tiers in 2023 that priced out most researchers and independent developers; Wired reported that the cheapest full-archive research package covered about 0.3% of monthly platform output.
Slack's 2025 rate-limit changes show the mechanism most cleanly, because nothing about the interface itself changed. The conversation-history endpoints stayed available to everyone. But a new commercial application outside Slack's Marketplace was held to one request per minute returning 15 messages — fifteen messages a minute, total. An internal customer application hitting that same endpoint could make 50 requests per minute and pull up to 1,000 objects per call. Same URL, same documentation, capacity varying by a factor of several thousand depending on your standing with the provider.
In each case the API remained. What varied was whether it was usable for your purpose.
The maintenance burden underneath all this is well documented. A 2024 study interviewing developers and architects at eleven companies found 13 of 17 running multiple API versions at once, and 14 reporting quality degradation over time. The recurring work is unglamorous: version-specific testing, backward-compatible routing, feature backports, ongoing communication with downstream consumers. The researchers called it consumer lock-in — providers unable to evolve their own interfaces because outside parties depend on old behavior, while the compatibility layers accumulate.
A public API is a commitment renewed quarter after quarter: documented, versioned, secured, supported, and eventually deprecated or fenced off when what it costs the provider exceeds what it returns. That arithmetic works for large partners, paying customers, and strategic integrations. It does not work for the long tail of connections that might be useful to somebody, which is precisely the population the API-economy narrative implied would be covered.
So the coverage is real but selective, and the selection follows the money rather than the technology. The distance between "has an API" and "has an API that covers what you need, at the volume you need, at a price you can pay" is where most machine-to-machine relationships still live. Better tooling has not closed it, because the tooling was never what was missing.