Companies are feeling the sting of AI spend. Uber exhausted its full‑year budget in four months, while Microsoft, Salesforce and GitHub have begun throttling employee usage. Frontier models such as GPT 5.5, priced at $5 per million input tokens and $30 per million output tokens, drive those costs. The price tag makes modest tasks, like fixing TypeScript types across dozens of files, quickly add up.

Frontier labs shoulder research, data curation and massive training bills that can reach hundreds of millions. Open‑weight releases let any inference provider add a modest markup, slashing prices dramatically. Anthropic’s Claude 4.8 matches the cost of its 4.7 predecessor, while China’s GLM‑5.2 beats GPT 5.5 on coding benchmarks at one‑tenth the token cost. These shifts erode the premium that early leaders once commanded.

Specialised silicon chips already cut inference spend by 30‑70 % versus Nvidia H100s, and caching or mixture‑of‑experts tricks keep latency low. As hardware costs fall and zero‑switching gateways let apps flip providers instantly, cloud‑only pricing loses its grip. Soon developers will run models locally for routine edits, leaving subscription fees for only the most demanding workloads.