Claude Fable 5.2 vs Claude Fable 5.1: A 20% Price Cut That Can Still Raise Your Bill

Claude Fable 5.2 shipped on November 9, 2026 at $8 per million input tokens and $40 per million output, down from the $10 and $50 that Claude Fable 5.1 has carried since September 1 — a straight 20% cut on both sides of the meter, cache reads down from $0.25 to $0.20, batch down from $5/$25 to $4/$20. That is the whole story for anyone who bills by the token, and it is the wrong story for anyone who bills by the task. Because the same release moved the cheaper model into defaults that make it think more: effort still defaults to high, adaptive thinking is still always on, and an instruction that reliably quieted Fable 5.1 — “answer directly without deliberating” — no longer stops its successor from thinking at medium effort and above. On a real workload a drop-in model-ID swap can raise the invoice while lowering the rate card. What follows is when the cut is genuinely yours, what the published scores say with their settings attached, the four API changes that will break code on the way over, and the order to do the migration in.

The two names are both current and both served, which is why the decision keeps landing in front of people mid-quarter. Claude Fable 5.1 is now labelled legacy on A​nthropic's own model page, with a standing note recommending migration and a retirement commitment of “not sooner than September 1, 2027.” Claude Fable 5.2 carries latest, a matching commitment of November 9, 2027, and the same one-million-token context window and 128K output ceiling. Nothing about the envelope changed. Everything about the economics and the behavior did, and they did not all move in the same direction.

The rate card, line by line

A​nthropic publishes both cards side by side, so this part needs no interpretation — only the arithmetic worth doing before a swap.

• Input — Claude Fable 5.2 $8.00 per million tokens vs Claude Fable 5.1 $10.00, a 20% cut.

• Output — $40.00 vs $50.00 per million, the same 20%.

• Five-minute cache write — $10.00 vs $12.50; one-hour cache write — $16.00 vs $20.00.

• Cache read — $0.20 vs $0.25 per million, and both models bill cache hits at 2.5% of base input, the steepest discount in the Claude line.

• Batch API — $4.00 / $20.00 vs $5.00 / $25.00 per million, a flat 50% off each card.

• Fast mode — $16.00 / $80.00 vs the same premium tier on 5.1, so the 2.5×-speed lane also drops 20%.

• Context and output — 1,000,000 tokens and 128,000 tokens on both; knowledge cutoff June 2026 on both.

• Retirement floor — November 9, 2027 vs September 1, 2027.

Two of those lines matter more than their size suggests. Cache reads at 2.5% of input mean the long agent sessions these models are sold for are already mostly cache-priced, so a 20% cut on the input rate is a smaller share of the bill than a 20% cut on output — and long-horizon work is output-heavy. The batch line is the one to check if you run offline evaluation or document processing: $4/$20 against $5/$25 is the same 20%, applied to the cheapest way to use either model. A​nthropic also advertises “30% faster output than 5.1” for the new model; that is a vendor claim, and the independent leaderboard that scores both models currently publishes no output-speed figure for Fable 5.2 at all, so treat the speed gain as unverified until you measure it on your own prompts.

What the scores say, and at which effort

A​nthropic's launch post reports a wide gap between the two Fables on the agentic benchmarks it tracks. Every number below is vendor-reported and unreproduced by us; where the post names a setting, it is included, because on these models the setting moves the score as much as the version does.

• FrontierCode 2.0 — 49.1% at xhigh for Claude Fable 5.2 vs 32.4% for Claude Fable 5.1.

• FrontierSWE v2 — 69.9% vs 56.3%.

• AutomationBench — 51.4% vs 31.4%.

• Terminal-Bench-Science 0.1 — 74.6% vs 52.6%.

• RoboDojo Sim — 64.9% vs 27.6%.

• OSWorld 2.1 — 89.3% vs 83.5%.

• WANDR (relaxed) — 82.1% vs 68.7%.

• Humanity's Last Exam with tools — 74.2% vs 65.6%; Chartography without tools — 73.4% vs 44.8%.

• AA-Briefcase v1.1 — 66.1% rubric and 1,785 Elo vs 61.5% and 1,669 Elo.

• RLI — 41.3% at high vs 17.9%.

• BenchCAD 2.1, scored in text — 83.5 vs 44.4.

The pattern is not “a bit better everywhere.” It is a large jump on the benchmarks built around long tool-using runs — FrontierSWE, AutomationBench, RoboDojo, RLI — and a modest one on tasks that were already near the ceiling, like OSWorld. That is consistent with the positioning: A​nthropic sells this tier for agent sessions that run for hours, multistep research, and analysis carried through to a finished document, and its own “what's new” page says most of the capability gain is already present at the default high effort, with uneven returns above it.

The independent check is narrower and points the same way. Artificial Analysis, which runs its own evaluation suite rather than repeating vendor charts, lists Claude Fable 5.2 at an Intelligence Index of 65 and first of 74 models, against 49 and 24th for Claude Fable 5.1. The cost side of that same suite is less flattering: $26.54 per completed index task for 5.2 against $12.99 for 5.1. Read that pair carefully, because it is the entire argument of this article in two numbers — roughly 33% more capability, roughly double the cost per finished task, on a rate card that is 20% cheaper. The benchmark task got harder to complete because the model spends more tokens completing it. That is a real trade, not a pricing trick, and it is invisible if you only compare per-token rates.

Four changes that break code

The migration guide is unusually blunt: A​nthropic lists four changes from Claude Fable 5.1 that can break an integration. None of them is about the model's quality, and all of them fire at request time.

• The prefix check applies to every account. On Fable 5.1, a request that sends back a thinking block after the system prompt, the tools, or an earlier message changed returned a 400 error only on accounts created on or after August 31, 2026. On Fable 5.2 it returns that error for everyone, whatever the account's age. Code that edits history — trimming context, rewriting a system prompt mid-session, swapping tools — can run clean on 5.1 today and fail on 5.2 tomorrow. There is a stopgap: the thinking-binding-controls-2026-08-01 beta header with prefix_mismatch_behavior: "drop_block", which drops the affected blocks and lists them in input_transformations. A​nthropic's own guidance is to treat that as temporary, because dropping blocks makes the model think again and raises output tokens.

• Thinking blocks only travel between 5.2 and 5.2. Only Claude Fable 5.2 and Claude Mythos 5.2 read Fable 5.2's thinking blocks. Moving a conversation to Claude Opus 5.5, Claude Fable 5.1, or the model's own fallback targets (Claude Opus 4.8 and Claude Opus 5) silently loses the reasoning for the turns that run there. The request still succeeds and the dropped blocks are not billed. If a router or a fallback chain moves conversations between models, this is the change that quietly costs you context.

• The old computer-use tool now errors. The computer_20251124 tool returns a 400 on Fable 5.2 on every platform except Amazon Bedrock. Declare the computer_toolset_20260801 toolset instead.

• The advisor has to match the executor. With the beta advisor tool, a Claude Fable 5.2 executor accepts only Claude Fable 5.2 or Claude Mythos 5.2 as its advisor; anything else, including Fable 5.1, returns a 400. Changing the advisor's model is an edit to tools, so it has to happen in a new conversation or the prefix check fires.

One more constraint is not a code change but a contract change, and it applies to both models equally: Claude Fable 5.2, like Fable 5.1, is a Covered Model that requires 30-day data retention and is not available under zero data retention unless A​nthropic expressly authorizes it. Organizations on ZDR stay on models that permit it — that is Claude Opus 5.5's lane, and the comparison article covers it.

Where the bill goes up anyway

The behavioral section of A​nthropic's documentation reads like a list of ways the cheaper model spends more. The model ends turns before the work is finished more often at low and medium effort, closing with a list of what remains or an offer to continue. It delegates to subagents readily in harnesses whose prompts invite it, and subagents can account for a large part of what a session costs. It follows style rules and cautious instructions so closely that accumulated “never do X” notes suppress routine work — in one of A​nthropic's simulated setups, only Fable 5.2 held back where Fable 5.1 and Opus 5.5 retried. And at high and above it thinks before almost every reply, including replies to greetings.

A​nthropic quantifies a couple of these. Adding its suggested “finish the whole task” system-prompt text roughly halved the number of recorded sessions that ended with doable work left, at a cost of about 30% more output tokens. Adding an ownership sentence — “this codebase is yours” — made the model merge duplicated code in about half of tries instead of under a tenth, and used about 1.5× the output tokens. Those are vendor measurements on vendor tasks, but the direction is unambiguous and it is the mechanism behind the independent cost-per-task number: you are buying fewer abandoned runs and more autonomous work with tokens.

The practical control is effort, not prompting. A​nthropic's own guidance is to start at high, step up to xhigh only where you have measured a gain, reserve max for ceiling measurements, and step down to medium or low wherever your evals hold. Critically, it says to rerun the effort sweep rather than carrying over the level you tuned on Fable 5.1 — a level does not produce the same amount of thinking across versions. Teams that skip the sweep are the ones who experience a “cheaper” model as a more expensive one.

Running the two side by side

The honest adoption pattern for a version bump this large is a split, not a switch: keep the stable path on Claude Fable 5.1, canary the new model on the workloads where the benchmark gap is widest — long tool-using agent runs, automation pipelines, research that ends in a document — and compare cost per completed task rather than cost per token. OrcaRouter is built for exactly that shape of test. Both models sit behind one API key with the provider's list price passed through at zero markup, so the $8/$40 card is what you pay here and a vendor cut lands the same day rather than at the next billing reconciliation; a routing rule can send a fixed percentage of traffic to the new model, and a fallback chain can return to 5.1 — or to Claude Opus 4.8, one of Fable 5.2's permitted fallback targets — when a call fails or a classifier refuses.

One caveat in the interest of accuracy: checked on November 10, 2026, OrcaRouter's public catalogue lists anthropic/claude-fable-5.1 and anthropic/claude-opus-5.5 but does not yet list Claude Fable 5.2. Fable 5.1 remains available at its full $10/$50 card while the newer model is onboarded, and the same key already reaches the Opus tier if you want a cheaper control arm for the comparison.

The order to migrate in

A​nthropic's migration guide is short enough to follow literally, and the order matters because two of the steps change the prompt prefix and one changes the tool list.

• Set effort explicitly to high before the first request, so you are never measuring an implicit default; then run your evals at medium and xhigh as well.

• Audit your harness for history edits. If your code builds the messages array, run the prefix test before switching, even if Fable 5.1 has never thrown the error for your account.

• Check every model switch in your stack, including fallbacks and any routing layer, for conversations that carry thinking blocks across models.

• Move computer use to computer_toolset_20260801, and repoint any advisor tool at claude-fable-5-2 or claude-mythos-5-2.

• Strip prompt lines written for the old model — “answer directly without deliberating,” token-budget boilerplate, standing delegation instructions — and re-test. Some of them now cost tokens without changing behavior.

• Re-baseline cost and latency on your own tasks, with cost per completed task as the headline number.

Who should move now: teams running long, unattended agent work where a 30-point FrontierSWE delta or a doubling of AutomationBench is the difference between an unattended run and a babysat one, and teams whose eval suite already shows the current model falling short at higher effort. Who should wait a cycle: anyone on zero data retention, anyone whose integration edits conversation history and cannot yet make it append-only, and anyone whose workload is short, well-specified, latency-sensitive chat — at that end of the range the new model's tendency to think before almost every reply is pure cost, and the 20% rate cut does not cover it. The version bump is real and the price cut is real; they simply do not point at the same workloads, and the teams that get value from this release are the ones that measure both before they ship it.