blog: when does prompt caching actually save money

[Makatzrina]

What this adds

One blog post: When Does Prompt Caching Actually Save Money? — 1,245 words, six question-shaped H2s, FAQPage schema for answer engines.

Plus its cover SVG and the one-line slug registration that stops the blog card falling back to the generic placeholder.

The angle

Every guide on prompt caching quotes "up to 90% savings" and almost none do the arithmetic on the write premium, which is what actually decides whether caching saves anything.

The post works the break-even for all three pricing shapes:

Write premium	Ahead from
None	the 1st reuse
1.25x (5-minute)	the 2nd use
2x (1-hour)	the 3rd use

Then a worked example at published Claude Opus 5 rates where input cost falls 82% but the bill only falls 67%, because output is never cached and is priced five times higher. That gap is the honest correction to the headline number.

It also names the trap: a one-shot workload that never re-reads its prefix doubles its input cost on a 2x write premium, and the long cache window looks like the safer option while carrying the higher premium.

Sourcing

The post makes no claims about AntSeed. Every figure comes from OpenAI's and Anthropic's own pricing and caching documentation, fetched and checked 1 September 2026, so nothing here needs internal verification.

Target keyword: when does prompt caching save money. The head term prompt caching is held by established domains and isn't reachable from current authority; the question form is, and it's the shape answer engines quote.

Verification

• All 3 internal links statically checked against real built routes — the site uses onBrokenLinks: 'throw'
• Both external sources fetched and every quoted figure verified against the live vendor pages
• Arithmetic in both tables recomputed independently

Not run: a full docusaurus build. Dependency install did not complete in the working environment, and ci.yml builds tiers 0–3 but never touches apps/website, so the website build is not exercised on PRs either. The only code change is one string added to a Set<string>. Worth a local pnpm --filter @antseed/website build before merge.

Changelog

No user-facing package, CLI, desktop, protocol, or configuration change. Website content only, so CHANGELOG.md is not updated.

🤖 Generated with Claude Code

[Makatzrina]

Build verified

Ran a full production docusaurus build against the repo-pinned pnpm 9.15.4 tree. This closes the caveat in the description.

✔ Server: Compiled successfully
✔ Client: Compiled successfully
[SUCCESS] Generated static files in "build".

Confirmed in the output:

• build/blog/when-does-prompt-caching-save-money/index.html renders
• FAQPage JSON-LD present in the emitted HTML
• Cover SVG copied to build/img/blog/covers/
• onBrokenLinks: 'throw' passed — all three internal links resolve

One warning, pre-existing and unrelated to this PR: a broken anchor on /docs/guides/using-the-api/ pointing at /docs/guides/agents/#define-your-internet-accessible-antseed-endpoint. That heading no longer exists on the agents page. Worth a separate fix.

Note for whoever picks this up: ci.yml builds tiers 0–3 but never touches apps/website, so the website build isn't exercised on PRs. Given the blog is about to receive regular commits, adding a website build step to CI would catch this class of problem automatically.