At $0.25 per million, a cache read costs one fortieth of a fresh input token on Fable 5.1; at the old $1.00 it was one tenth [1][1][2]. The arithmetic that decides whether you care: take a 200,000-token prefix, a system prompt plus a tools array plus a retrieved corpus, and read it on 1,000 turns. That is 200 million cache-read tokens, $200 last week and $50 now [4]. Send 2,000 tokens of fresh input per turn and those 2 million tokens still cost $20 at the unchanged $10 rate [5]. The New Stack's pricing paragraph stops at input, output and cache reads with no write price [3], so treat the $50 as a floor rather than a total.
The saving is only available while the prefix stays stable, and 5.1 adds a second rule watching the same bytes. Every thinking block is bound to the prefix that produced it: modify the system prompt, the tools array, or an earlier message ahead of a thinking block, and the next request errors, or drops the block if you opted into that [9]. The docs list moving cache_control markers among the edits that keep later blocks valid, along with server-side compaction, context editing, and changing effort between requests [10]. So the two mechanisms were built not to fight each other, which is more than you usually get. Removing a thinking block from anywhere other than the start of the run invalidates every block after it [11].
Enforcement is where migrations get interesting. The check binds for accounts created on or after August 31, 2026; earlier accounts record the mismatch and act on it only when the request sets thinking.block_binding.prefix_mismatch_behavior [12]. The same messages array can pass on one account and come back as a 400 reading "The block is bound to a different conversation" on another [12][13]. If Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK assemble your turns, the prefix is kept intact for you [22]; if your own code builds the array, that is what to check before you swap the model string [9].
Forced tool use is the breaking change with a stated reason. tool_choice set to {"type":"any"} or to a named tool returns a 400, on the token counting endpoint too [6]. Thinking is always on for these models, and a forced call would skip it, so the model writes its working-out into the tool arguments and argument quality drops [7]. The replacement is auto with strict:true, or the schema moved to structured outputs, plus a prompt line stating when the tool applies [8].
On the benchmark table: Terminal-Bench-Science 0.1 moves 27.9 points, a 2.13x [6]. For that to land in your numbers, your workload has to look like agentic scientific research run on Anthropic's harness. Everything else sits within 2 to 4 points of Fable 5 or Opus 5 [17], which is inside the noise of most in-house eval sets. The claim worth testing yourself is the token one: Anthropic says you can step reasoning down one or two notches and match Fable 5 at its higher settings while spending fewer tokens overall [18]. That is checkable against your own traces, and it is why Fable 5.1 ships on high effort in Claude Code but medium on Claude Cowork and claude.ai [19].
