The Fable 5.1 price cut matters more than its benchmarks
The benchmark improvements are real, but the 75% cache-read price cut is what actually changes the game for long-running AI agents.
Anthropic shipped Fable 5.1 today. The headline is the benchmarks. The story is the price.
The launch post leads with a doubling of Terminal-Bench-Science scores, a new high on Humanity’s Last Exam, protein binders validated in real labs. All true, all worth a skim. But read past the charts and you hit the sentence that actually matters: cache reads now cost $0.25 per million tokens, down from $1.00. For typical workloads that is roughly 25% less than Fable 5. For heavily agentic, context-heavy work, Anthropic says the savings reach about 45%.
That is not a footnote. It is the release.
What Fable 5.1 actually does
The model itself is a solid step forward. Terminal-Bench 4.0 goes from 42.0% to 55.8% with safeguards on, and 60.9% in the less restricted Mythos variant. CursorBench 3.2.0 climbs to 73.4%. Red Hat reports it found the root cause of every broken build they tested. Millennium says it finally explained a rare crash their engineers chased for years, by disassembling an external vendor library and matching it against a core dump.
The interesting part is that the gains are not evenly spread. Strip out Terminal-Bench-Science and the deltas over Opus 5 are modest: roughly +3.5% on Terminal-Bench 4.0, +1.5% on GDPval, +2.5% on OSWorld. One HN commenter (GodelNumbering) put it bluntly: for a model one tier up and one version up, those are not big improvements, and it leaves almost no room to place an Opus 5.1 above it. The frontier is not leaping forward in every direction. It is bending hard in two specific ones: long-horizon agentic work, and cost per useful task.
Why cache-read pricing is the sleeper
Long-running agents do not burn tokens on fresh input. They re-send the same transcript, the same context, the same tool history, call after call. At a typical 95% cache hit rate, one commenter (seaurchinzee) calculated their optimal context window before autocompaction roughly doubles, from about 200K to 400K tokens. Simon Willison made the same point in the thread: cache pricing directly hits any long-running agent, because subsequent calls re-read prior transcripts. A vendor at viktor.com reported a ~30% real-world cost reduction in their evals.
This matters because the unit of AI work has changed. Nobody cares what one prompt costs. They care what an eight-hour autonomous session costs, one that wakes up with a full context, plans, runs experiments, fixes its own failures, and comes back with a result at 3am. That class of workload only becomes rational when the re-reads are cheap. Anthropic just made the re-reads a quarter of the price and then pointed the model at exactly that workload with its new effort levels. Fable 5.1 defaults to High effort in Claude Code. The direction of travel is unambiguous.
What the skeptics get right
The HN thread is not buying the narrative wholesale, and not all of it is noise. The prose complaints are real: people are exhausted by Claude’s word density, by “load-bearing” and walls of text, by a model that turns a one-line answer into a page. Anthropic says Fable 5.1 writes better, and early reports agree, but the trust has been damaged. Several commenters describe the same pattern: the model is smart, the output is exhausting, and they have drifted to Codex or open models out of sheer fatigue. That is a product problem, not a benchmark problem, and no price cut fixes it directly.
The pricing has caveats too. The cache discount applies to token-billed usage, not subscription plans, and commenters noticed quota limits were cut 17% around the same time. The safeguards still refuse legitimately weird requests, like a developer’s own security review or a remote-execution coding task, until you remove the word “execution.” And Mythos, the genuinely unrestricted variant, is gated behind US-government-linked trusted access programs, which is its own conversation about who gets frontier capability and who does not.
The takeaway
Watch the cost curve, not the leaderboard. Benchmarks are a snapshot; prices are a strategy. Anthropic is telling you where agents are headed: long-running, context-heavy, unattended work, priced so that the marginal cost of a re-read is near zero. The 45% savings on agentic workloads is the real product. The Venus elevation map, the protein binders, the overnight experiment runs, those are the proof that the strategy works, not the strategy itself.
If you run agents today, your practical move is simple. Re-measure your workloads against Fable 5.1 with cache-heavy settings, especially anything that keeps a long context alive across many calls. The model that runs for eight hours and costs less than the one that runs for three is not just cheaper. It changes what you can justify building at all.