Claude Fable 5.1 vs Claude Fable 5: What Anthropic's Own Flagship Refresh Actually Buys

Claude Fable 5.1 is Anth​ropic's current flagship, launched September 1, 2026, and the matchup that matters most for most teams is the one Anth​ropic set up itself: Claude Fable 5.1 against Claude Fable 5, the June 9 flagship it replaces. The price is identical — $10.00 per million input tokens and $50.00 per million output tokens for both — and so is the 1M-token context window. What changed is the character of long agentic work: Anth​ropic's own numbers show Fable 5.1 more than doubling Fable 5 on its hardest science-agent benchmark, and this week's reporting on how enterprises are treating the Fable family over data retention gives the refresh a second, sharper edge for anyone choosing between the two today.

A note on dates and sourcing, up front: Claude Fable 5.1 shipped on September 1, so the model itself is two weeks old, not breaking news. What is current is the reporting from September 14, 2026 — attributed to The Information — that Nvidia, Palantir and Booz Allen are restricting their use of Anth​ropic's Fable family over concerns about retaining sensitive corporate data, and that Anth​ropic introduced rolling 30-day usage logs for safety monitoring in June with the customer option to self-store those logs "still able to be revoked." That reporting is source-attributed and not yet independently confirmed. Everything labeled vendor-reported below is Anth​ropic's own benchmark sheet and has not been reproduced by a neutral party.

The spec sheet, dimension by dimension

• Price — Claude Fable 5.1 $10.00 in / $50.00 out per 1M tokens, the same as Claude Fable 5. Neither price moved at the refresh.

• Cached input — Claude Fable 5.1 $0.25 per 1M tokens, down 75% from Claude Fable 5's $1.00. Cache writes $12.50 (5-minute) / $20 (1-hour); batch mode $5 / $25.

• Context and output — both carry a 1M-token input window and up to 128K output tokens; both take text and image input; both sit behind the same adaptive-reasoning surface with an effort dial from low to max.

• Independent score — on the current version of Artificial Analysis's Intelligence Index, Claude Fable 5.1 measures 53, first of the 201 models tracked, at 67.2 output tokens per second and $7.63 per index task. Claude Fable 5's launch-scale index figure (62) predates the September restatement and is not comparable, so we do not quote a current-scale number for it.

• Vendor benchmarks — Anth​ropic reports Claude Fable 5.1 at 52.6% on Terminal-Bench-Science 0.1 versus 24.7% for Claude Fable 5; 55.8% on Terminal-Bench 4.0 versus 42.0%; 31.4% on AutomationBench versus 17.1%; and 73.4% on CursorBench 3.2.0 versus 70.5%.

• Released — Claude Fable 5.1 on September 1, 2026; Claude Fable 5 on June 9, 2026.

Where the refresh is actually a refresh

The headline trick of the Fable 5.1 launch is that nothing on the rate card moved and almost everything on the benchmark sheet did. The gains concentrate in long-horizon, tool-using, agentic work — Terminal-Bench-Science 0.1 more than doubling from 24.7% to 52.6%, Terminal-Bench 4.0 climbing from 42.0% to 55.8%, AutomationBench nearly doubling from 17.1% to 31.4%. On Humanity's Last Exam, Anth​ropic reports 60.9% without tools and 65.0% with them, against 57.8% and 63.8% for Claude Fable 5. Read these as directions of travel: they are the vendor's own table, and the biggest single-number claims describe the gated sibling, Claude Mythos 5.1, which shares the same weights behind stricter guardrails.

The second lever is cost, and it matters more than any single benchmark. Cached input dropped from $1.00 to $0.25 per million tokens, and long agentic runs are where re-reading dominates: tool output, file contents and prior turns get re-sent over and over. Anth​ropic estimates roughly 25% lower effective cost on typical token-billed workloads and up to 45% on highly agentic ones — a vendor estimate, but the underlying price change is verifiable on the price sheet. A long-horizon agent that re-reads a 100K-token context fifty times over a task pays $5.00 in cache reads on Claude Fable 5's pricing and $1.25 on Claude Fable 5.1's. Across a fleet of agents that line stops being a rounding error.

The third change is the one users of the previous generation feel first: the safety classifier. Developers nicknamed the old behavior "Fable's depression" — flagship money for output that was quietly rerouted to a weaker model when the classifier flagged benign requests. Anth​ropic says Fable 5.1's cyber safeguards now produce roughly 60% fewer interventions per Claude Code session, and that biology safeguards fire 85% less often on benign requests than at Fable 5's launch. If your complaints about the June model were about mid-task handoffs, this is the upgrade that addresses them directly.

This week's retention story, and what it does to the decision

The September 14 reporting is the part that changes the calculation for enterprises, because the Fable family is where Anth​ropic's most capable model sits — and because data-retention posture is exactly the kind of thing a regulated buyer cannot hand-wave. As reported, Anth​ropic introduced rolling 30-day usage logs for safety monitoring in June; customers were told they could self-store those logs instead; and the reporting says that option "can still be revoked." Palantir is described as withholding Fable until it gets irrevocable zero-data-retention guarantees, Nvidia as limiting Fable to less sensitive work, and Booz Allen as excluding it from proprietary cybersecurity work. All of it is attributed reporting, none of it is confirmed by the named companies, and its practical effect is uneven: some customers will care a great deal, and some will not care at all.

What it does to the 5.1-versus-5 decision is worth being precise about. The retention story applies to the Fable family as a whole — it is not a new difference between the two models, since Fable 5 and Fable 5.1 share the same retention posture. What it does change is the risk-weighted math: if your workload is sensitive enough that a revocable self-store option is disqualifying, then neither model clears the bar today, and the conversation belongs with Anth​ropic's Enterprise Frontier Safeguards (EFS) roadmap, which is described as phasing zero-data-retention in from fall 2026 for eligible customers. If your workload is not that sensitive, the retention reporting should not talk you out of a real performance upgrade — but it should be on the decision record rather than discovered later.

Who should pick which

Pick Claude Fable 5.1 when the cost of a wrong or abandoned answer dominates the cost of tokens: long-horizon agentic research, heavy tool use, anything that lives inside Claude Code, and workloads where a 75% cheaper cache read changes your unit economics. Stay on Claude Fable 5 when your traffic is short, single-shot and throughput-tolerant — the price is identical and the old model's scores are still frontier-class — or when you are mid-contract and the migration overhead is real. The honest middle ground is to not pick at all: both models sit behind the same Anth​ropic API surface, and if you call them through a single endpoint you can escalate to Claude Fable 5.1 when a task needs the reasoning ceiling and let the cheaper-identical-priced pair share failover without a second contract.

On OrcaRouter, Claude Fable 5 and Claude Fable 5.1 are both available at their providers' list prices with zero markup, so the $10 / $50 above is the number on the invoice, and Anth​ropic's own price changes pass through the same day rather than being absorbed into a reseller margin. Testing the refresh against your own traffic is a routing-rule edit with automatic failover — the two models share one key, so the switching cost is near zero, and model fusion lets a panel answer together when a task is important enough to want more than one opinion.

The short version: Claude Fable 5.1 is the same price as Claude Fable 5, roughly twice as strong on the agentic-science benchmarks Anth​ropic highlights, 75% cheaper on the cache reads that dominate long sessions, and materially less likely to hand your request off to a weaker model mid-task. The September 14 retention reporting is real and worth weighing if you are an enterprise with strict data requirements — but it applies to the family, not to this specific upgrade, and it should be one line in a decision that the benchmarks and the cache math already decided.