What’s actually changing
This is a synthesis cluster, not a single dated announcement: it pulls together GPT-5.6’s new caching structure and Anthropic’s own published caching structure to compare the two directly, sourced to each provider’s own current pricing documentation rather than third-party coverage.
OpenAI’s GPT-5.6 requires explicit cache breakpoints and charges a reported 1.25x premium on the first write to a breakpoint, while reads against an established cache entry get roughly a 90% discount off the standard input rate, figures that match OpenAI’s own prompt caching documentation. Anthropic’s Claude caching charges the identical 1.25x premium for a 5-minute cache write, or 2x for a 1-hour cache write, with a 0.1x (90% off) rate on cache reads either way, per Anthropic’s own published pricing multipliers. Both providers charge a real, documented write premium; the actual difference is that Anthropic gives you a choice of two durations at two different premiums, where OpenAI has one fixed 30-minute tier.
Structural Comparison Matrix
| Caching Dimension | OpenAI (GPT-5.6) | Anthropic (Claude) |
|---|---|---|
| Cache write cost | 1.25x premium | 1.25x premium (5-minute cache), 2x premium (1-hour cache) |
| Cache read discount | ~90% off | ~90% off (0.1x base rate), same for either duration |
| Cache duration | Fixed at 30 minutes | Choice of 5 minutes or 1 hour |
Fix it: model the write side, not just the read discount
Picking a provider based on caching-friendliness used to be a simple read-discount comparison. It no longer is for OpenAI, and it never really was for Anthropic either, since Anthropic’s own multipliers show the same tradeoff. The real cost of caching a given prefix depends on how many times it gets reused before expiring: reuse it enough times and the write premium is a rounding error against the accumulated read savings; reuse it once or twice and the premium may cost more than caching saved.
Anthropic’s two-tier structure adds one more variable: the 1-hour tier’s steeper 2x premium can still win over the 5-minute tier’s 1.25x if it means paying the write premium once instead of paying the 1.25x premium repeatedly across an hour of reuse. Do that comparison explicitly rather than assuming the cheaper-looking multiplier is actually cheaper for your access pattern.
Run this exact comparison on your own numbers
This post’s whole argument, that the write premium and cache duration have to be modeled together, not just the read discount, is exactly what the LLM Pricing Calculator computes. Enter your own token counts and reuse pattern and see both providers’ real cache-write and cache-read costs side by side, instead of doing this arithmetic by hand.
Two related pieces of this reset
See GPT-5.6 Ends Free Prompt-Cache Writes and OpenAI Extends Prompt Cache Retention to 24 Hours for the two OpenAI-side changes this comparison draws on.
Confirmed version
Sourced directly to OpenAI’s prompt caching documentation and Anthropic’s pricing documentation, re-checked 2026-08-13. An earlier version of this post cited third-party cost-math coverage and stated Anthropic had no documented write premium; that was wrong, corrected here against each provider’s own primary pricing page. Browse more coverage in the AI Productivity archive, or start from The 2026 LLM Token & Pricing Reset hub.
