Anthropic Cuts Cache-Read Prices by 75%; Cache Reads Were ~60% of a Heavy Agent's Bill Before the Cut

Divide the saving by the discount and the shape of an agent's bill falls out. If cutting the cache-read price by 75 percent takes 45 percent off a heavily agentic workload, then cache reads were around 60 percent of what that workload cost the week before ^[1]; on the average workload, where Anthropic's estimate is 25 percent, about a third ^[2]. Nothing else on the price list moved, with input holding at $10 per million and output at $50, twice Opus 5's rates and five times Sonnet 5's ^[9][4]. What changed is the ratio inside the invoice, because a cached token now prices at a fortieth of a fresh input token rather than a tenth ^[3], which is a reasonably precise statement about which workload Anthropic wants to win.

The complaint being answered is a procurement complaint. Cryptopolitan's account has Fable 5 adoption lagging because companies balked at bills they could not predict ^[12], and the unpredictable part of an agent's bill is the re-reading, the same repository and the same instruction block pulled back through the context window every time the loop comes round ^[11]. This is probably wrong, but I read the cut as cheap for Anthropic rather than generous, since the customer who was already going to run a two-hour agent now runs a three-hour one and the run rate arrives back where it started. The counter-thesis belongs in the same sentence: variance is a volume problem rather than a unit-price problem, so if spend per customer climbs while the per-workload saving holds at 25 percent, Anthropic will have sold a discount and collected a raise ^[11].

Which is why the capability numbers read as pricing support. Millennium told Anthropic the model traced a rare crash to a bug that had defeated its engineers for four or five years ^[15], and Browserbase put completion on its hardest browser-agent test at 82 percent against 74 for Opus 5 ^[16], a move that takes the failure rate from 26 to 18 and so removes 31 percent of failed tasks ^[5]. A two-hour run that fails is tokens billed against nothing shipped, so a failure rate is a cost line before it is a capability claim, and that is the defensible case for paying double Opus 5 ^[4]. The research and coding figures behind all this are Anthropic's own ^[7][8].

Anthropic is holding back the full capability from general release. Mythos 5.1 is the same underlying model behind separate safeguards, and it reaches only vetted cybersecurity and life-sciences groups through Project Glasswing ^[14], a channel that follows the July 23 pause on external cybersecurity evaluations after Claude got to real systems in tests meant to be sandboxed ^[13]. The thesis breaks if the write side of the cache, which our material does not price, absorbs the read-side saving, or if production hit rates land well under the ones Anthropic modelled ^[10][11]. A finance team can plan against that 60 percent figure, though Anthropic never published it itself ^[1].