|
OpenAI shipped a 12-hour cache for its API this week, with a "cache forever" option for anyone on ChatGPT Work or Codex plans — a change that looks small until you do the math on repeat prompts.
Caching isn't new to the API, but extending the window this far, and removing it entirely for paid-tier users, turns caching from a latency trick into a real cost lever: any system prompt, codebase context, or agent scaffold reused across calls stops counting as fresh billable tokens after the first hit. For a solo builder running an agent loop that re-sends the same large context on every turn, that's the difference between paying full input price per call and paying it once.
It lands in the same week Anthropic is still working the opposite side of the same lever — Sonnet 5.5's cache-read price already fell from $0.20 to $0.10 per million tokens this month — which is the pattern this topic keeps surfacing: nobody wants to cut the sticker price on a flagship model, so the competition has moved entirely into cache, batch, and tier mechanics that only show up if you read the fine print. OpenAI also quietly folded the caching rollout into a broader release that included Locked Chats (PIN/biometric-protected conversations) and an acquisition of Orchestra Bio to expand its Rosalind life-sciences tools — the kind of week where the actual price-relevant change sits a few bullet points down from the splashiest headline.
For anyone building on the API rather than subscribing to ChatGPT, the takeaway is concrete: if your workflow re-sends large, static context, check whether you're on Work or Codex, because "cache forever" is the quiet unlock that makes those workflows meaningfully cheaper to run at volume.
More this week:
| 