Developers ditch paid AI subs for local models

Some developers are dropping $20-a-month artificial intelligence subscriptions and running open-source models on their own machines instead. (openai.com) (anthropic.com) (google.com) The math is simple at the entry tier: ChatGPT Plus costs $20 a month, Claude Pro costs $20 a month, and Google AI Pro is listed at $19.99 a month in the United States. Cancelling all three would cut about $60 a month, or roughly $720 a year, before any hardware costs. (openai.com) (anthropic.com) (google.com) Running a model “locally” means the software generates answers on a user’s own computer instead of on a company’s servers. Ollama says its runtime is designed to let users run open models on their machines, and OpenClaw says it can connect to those local Ollama models through Ollama’s native application programming interface. (github.com) (docs.openclaw.ai) One model drawing attention in that setup is GLM-5.1. Ollama’s model library describes it as a flagship coding model and says it was updated last week, while OpenClaw documents GLM as a supported model family through Z.AI. (ollama.com) (docs.openclaw.ai) The appeal is not only price. OpenClaw’s documentation pitches local use as a way to avoid application programming interface fees, and local inference also keeps prompts and files on the user’s own device unless they turn on outside tools such as web search. (getopenclaw.ai) (docs.ollama.com) That shift is landing as companies push the opposite model: more paid tiers, more usage caps, and more premium plans for heavier users. OpenAI’s pricing pages list ChatGPT Plus at $20 a month and separate application programming interface billing, while Google’s AI plans page bundles Gemini access with storage and other features. (openai.com 1) (openai.com 2) (google.com) A second argument in the debate is about what counts as “productive” artificial intelligence use. The Information reported that Meta took down an internal leaderboard tracking employee token use after the dashboard showed more than 60 trillion tokens used over a recent 30-day period. (theinformation.com) That leaderboard turned token counts into status markers with labels such as “Token Legend,” according to The Information. Outside the company, critics have mocked the idea that burning through more tokens — the chunks of text a model reads and writes — proves better work. (theinformation.com) (msn.com) The local-model crowd is making the opposite pitch: fewer subscriptions, fewer metered prompts, and more control over where the work runs. The companies selling those subscriptions are still betting that convenience, top-end models, and bundled tools will keep users paying. (github.com) (openai.com) (google.com)