Market-structure
Anthropic's Fable 5.1 didn't cut its sticker price, it cut the cost of an agent staying in context all day, and that is the number that decides whether continuous market-monitoring agents are economical
Fable 5.1 keeps Anthropic's list price at $10/$50 per million input/output tokens but drops cached-token reads from $1 to $0.25 per million, a 75% cut Anthropic says makes highly agentic workloads roughly 45% cheaper. For a trading agent that has to hold order books, news and position state in context for hours, that is the line item that mattered.
This site's building-the-stack and llm-agents-vs-quant pages have both made the same point about LLM-orchestrated trading systems since the first draft: the binding cost constraint isn't the per-call price of a model, it's the price of keeping that model's context populated across a long-running session. A quant strategy that re-evaluates a fixed feature vector every bar pays a predictable, small inference bill. An agent watching a market open-to-close, re-reading the same order book snapshot, position ledger, risk limits and news feed on every reasoning step because none of it has left its context window, pays for that re-reading every single time, and that repeated read is exactly what a prompt cache exists to make cheap. Fable 5.1 leaves Anthropic's headline price unchanged at $10 per million input tokens and $50 per million output tokens; it re-prices the cache read from $1 to $0.25 per million tokens instead. Anthropic's own framing, a roughly 45% reduction for highly agentic workloads against a 25% reduction for typical ones, is a tell about where the company thinks the money is actually going: not the first pass through the model, but every pass after it.
The other change matters for the same reason. Fable 5.1 ships a 1 million token input window at standard rates with no long-context surcharge, up from Fable 5's smaller default ceiling. A trading agent that previously had to summarize, truncate or externally retrieve its own market history to stay inside a working context now has more room to just keep it there, which is a workable design simplification but only an economical one if reading that larger context back on every turn doesn't scale the bill linearly with window size. The cache-read cut is what keeps that arithmetic from getting worse as the context grows, and the two changes shipping together, more room plus cheaper repeated reads of that room, reads as a single decision aimed at exactly the always-on agent pattern this site's multi-agent-architecture coverage has described: a supervisor agent and several specialist sub-agents (execution, risk, research) each maintaining their own long-lived context across a trading day, coordinating rather than starting fresh on every message.
None of this changes the questions this site's regulatory-boundary and risk-governance pages keep asking, who is accountable when the agent acts, what kill switch stops it, what audit trail survives it, and a cheaper cache does nothing to answer any of them. What it does is remove a cost objection that has been quietly limiting how continuously an agentic system can afford to watch a market rather than poll it. A firm that concluded eighteen months ago that keeping an LLM agent live and context-aware through a full trading session was too expensive to justify against a simpler polling or scheduled-inference architecture has a materially different unit-economics answer to that question today, independent of any change in what the model itself is capable of reasoning about. That is a market-structure fact worth tracking on its own terms: the frontier-model vendors are now competing on session economics for agentic workloads specifically, not just benchmark scores, and trading is one of the more expensive sustained-context use cases that pricing change will get tested against first.
Read the original: LLM Stats - Anthropic's Fable 5.1 didn't cut its sticker price, it cut the cost of an agent staying in context all day, and that is the number that decides whether continuous market-monitoring agents are economical. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.