Anthropic puts its deepest Opus 5.5 cut on cache reads, at 20 cents a million

Opus 5 charged $5 per million input tokens and $25 per million output; Opus 5.5 charges $4 and $20. That is a fifth off each line. Yahoo Finance and the summary at the top of Decrypt's report both put the cut at 40%, the second specifying default settings, while Decrypt's own body text and Cryptobriefing both say 20%. The published list prices support 20%.

The 40% version becomes arguable once cache reads enter. Those fell 60% to $0.20 per million, from an implied $0.50, and Decrypt says reused context drives most of the cost on long agentic coding sessions. Take a bill split evenly between cache reads and output tokens: the cache half drops to 40% of what it was, the output half to 80%, and the total lands at 60% of the old bill. That is a 40% saving, and it is the workload that produces it.

Against Fable 5.1 the claim is 60% cheaper. Applied line by line to $4 and $20, that puts Fable 5.1 near $10 and $50. Anthropic says the new model performs at Fable 5.1's level on most tasks, and it also concedes that the gap between the two families is narrower than its own benchmark scores suggest.

Those scores, on Anthropic's tests: Terminal-Bench 4.0 has Opus 5.5 completing 66.4% of tasks against Fable 5.1's 55.8% and GPT-6 Astra's 57.9%, a 10.6-point lead over the flagship. FrontierCode puts it at 54.4% against Fable's 50.3%. On GDPval-AA v2.1, an Elo ranking across 44 occupations, Opus 5.5 scored 1846 to Fable's 1735 and Opus 5's 1708, a 111-point margin. Astra takes two: 41.4% to 40.0% on AutomationBench, and 64.6% to 58.7% on Terminal-Bench-Science 0.1.

I would not budget a fifth off model spend on this, because the volume side is unpriced: none of the three accounts says how many tokens an agentic session consumes. Anthropic measures most of its progress through agentic coding, where a model takes multi-step actions on its own instead of answering a single prompt, and those sessions run long enough that cached context dominates the bill.

Cryptobriefing calls the cut pre-IPO positioning, reporting that Anthropic raised $65 billion in May 2026 at a $965 billion post-money valuation, that a confidential S-1 reportedly followed around June, and that internal discussions have centred on valuations above $1 trillion. The same publication says open-weight models and low-cost providers, particularly Asian ones, have been eroding the pricing power frontier labs once enjoyed.

Opus 5.5 is also the first model Anthropic has shipped since Dario Amodei's essay on slowing down. "We must slow the pace at which we improve the capabilities of AI models," Amodei wrote earlier this month. The model launches with the same cybersecurity and biology safeguards as Fable 5.1, and most cybersecurity tasks are automatically rerouted to the older Opus 4.8, which developers have complained about before.

Opus 5 launched on July 24 and Opus 5.5 on September 22, 60 days apart. Repeat a 20% list cut at that cadence for a year and $4 input tokens reach about $1.05. I'd expect segmentation rather than that: Sonnet 5.5 and Haiku 5.5 are expected soon, with Haiku historically the cheapest tier and Sonnet the middle. The case for falling frontier prices breaks if Fable 5.1 holds where it is, because then Anthropic has added a cheaper rung while the top of the ladder stays near $10 and $50.