Fable 5.1 Doubled Its Science Score. Older Models Can't Read Its Mind.
TL;DR
On September 1 Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1, the same model under two names: Fable for everyone, Mythos for vetted US organizations in Project Glasswing. Input and output stay at $10 and $50 per million tokens, but cache reads drop from $1.00 to $0.25, which Anthropic says cuts typical bills about 25% and agentic ones up to 45%. The headline capability number is Terminal-Bench-Science, where 5.1 scores 52.6% against Fable 5's 24.7%. The fine print matters more: three API changes break existing Fable 5 code on purpose, thinking blocks are now readable in one direction only, and the model you can buy scores 5.1 points below the identical model you cannot.
What shipped, and what stayed put
The model page lists the specs: a 1M-token context window at flat per-token pricing, 128K max output, a June 2026 knowledge cutoff, adaptive thinking that cannot be switched off, and a default effort of high. The model ID is claude-fable-5-1 on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry, and Anthropic commits to keeping it live until at least September 1, 2027.
Mythos 5.1 shares every spec and every price. The difference is safeguards: Fable 5.1 runs with production classifiers for cyber and biology, Mythos 5.1 runs with them loosened, and Mythos access goes only through Glasswing and your account team. The docs also contain the least salesy line of any launch this year: "For most workloads, start with Claude Opus 5." Fable 5.1 is pitched for the cases where Opus 5 at higher effort still falls short.
The science score more than doubled
Anthropic leads with Terminal-Bench-Science 0.1, an agentic research benchmark where the model has to run real scientific workflows from a terminal. Fable 5.1 posts 52.6%. Fable 5 scored 24.7%, Opus 5 29.0%, and OpenAI's GPT-5.6 Sol 22.4%, all per Anthropic's own table.
The rest of the table is a steadier climb. On Humanity's Last Exam it reaches 60.9% without tools and 65.0% with them, up from 57.8% and 63.8% for Fable 5. On OSWorld 2.0's strict computer-use scoring it moves from 36.1% to 41.7%. CursorBench 3.2.0 goes from 70.5% to 73.4%, GDPval-AA v2 from 1723 to 1853, and AutomationBench, a business-workflow eval, from 17.1% to 31.4%.
The one independent number so far comes from ARC Prize, which verified five effort variants on launch day. At max effort Fable 5.1 scores 90.0% on ARC-AGI-2 at $4.49 per task, against Fable 5's 89.2% at $5.45. On ARC-AGI-1 it actually slips, 97.5% versus 98.5%, though at $1.40 per task instead of $2.11. Read that as a model tuned to spend fewer tokens per answer rather than a smarter one on grid puzzles. ARC-AGI-3 results are not yet posted.
The 5.1-point safeguard tax
The most interesting row in the announcement is Terminal-Bench 4.0, the agentic coding benchmark, because it lists both variants of the same weights. Mythos 5.1 scores 60.9%. Fable 5.1 scores 55.8%. Mythos 5 sat at 42.0% and GPT-5.6 Sol at 37.3%.
Anthropic's footnote explains the gap: Fable 5.1 was evaluated with production safeguards on, and when a classifier fires on a cyber-flavored task the request is redirected to Opus 4.8, or to Opus 5 for biology. Those fallbacks count against the score. The tax is what you pay in benchmark points for the classifiers, and a 5.1-point levy on a 5.1 release is at least easy to remember.
The safeguards did get looser. Anthropic says Fable 5.1 can now identify software vulnerabilities, though not write exploits, and that Claude Code sessions see about 60% fewer cyber interventions than on Fable 5. Penetration testing, exploit generation, and binary scanning still bounce to the Opus models. One Hacker News commenter who runs a security org reported spending months clearing Anthropic's Cyber Verification Program only to learn it does not unlock Fable at all. The unlocked model is Mythos, and Mythos means Glasswing.
Cache reads at 2.5% of input
The pricing change is small in the table and large on the invoice. Every other Claude model bills a cache read at 10% of base input. Fable 5.1 bills it at 2.5%: $0.25 per million tokens, down from Fable 5's $1.00. Cache writes are unchanged at $12.50 for the five-minute tier and $20 for one hour, the batch API stays at half price, and the 512-token minimum cacheable prompt stays.
That is why the same list price produces a 25% to 45% cheaper bill. An agent loop re-reads its whole transcript on every turn, so cache reads dominate long sessions. One HN user doing the math at a 95% hit rate figured their optimal context before auto-compaction roughly doubles, from about 200K tokens to about 400K. Another pointed out that a 40x spread between fresh input and cached input is unusual, and reads more like a real change in how cached prefixes are served than a marketing discount.
Three things that break on purpose
The what's new page is unusually blunt about compatibility. If you call Fable 5 today, three changes will fail your requests:
- Forced tool use returns a 400. A
tool_choiceofanyor a named tool is rejected. Thinking is always on, and a forced call would skip it, so the model would dump its working into the tool arguments. Use strict tool use or structured outputs instead. - Older models cannot read its thinking blocks. Every block records the model that produced it. Fable 5.1 reads Opus 5 and Fable 5 blocks, but nothing older reads Fable 5.1's. A router that falls back mid-conversation silently drops the block, unbilled, unless you send the binding-controls beta header to get the drop reported.
- Editing earlier turns invalidates thinking. Change the system prompt, the tools array, or any earlier message and the next request fails with "The block is bound to a different conversation." It is enforced for accounts created on or after August 31, 2026; older accounts only get a warning unless they opt in. Mythos 5.1 does not run this check.
Anthropic files all three under strengthened anti-distillation. The mechanism works like a cheque that voids itself if anyone alters the payee line: each thinking block carries a signature bound to the exact conversation prefix that produced it, so you cannot harvest reasoning, splice it into a fresh transcript, and replay it into a cheaper model. An HN commenter noted the forced-tool ban also closes a known trick where a bogus "think deeply" tool coaxed the model into writing raw reasoning into its arguments.
Five additions arrive with the breakage: per-message effort changes that keep the cache warm, turn-scoped system messages that expire at the next user turn, a display mode that returns progress updates as text while reasoning stays hidden, the lower cache price, and Anthropic's statistical text watermark on every output, with C2PA credentials on files. The watermark adds no tokens and, Anthropic says, carries nothing about you or your organization.
What Anthropic says got worse
The same docs list regressions you will hit without changing a line of code. Fable 5.1 may issue one tool call per turn where Fable 5 batched several, which costs round trips in bash-and-editor harnesses. It writes fewer progress updates between tool calls, answers from memory more often at low effort, is more likely to rewrite a whole file for a small edit, more likely to paste source passages into a summary without quote marks, and in places writes denser prose with fewer paragraph breaks. Each comes with a fix on the prompting guide.
That last item drew the most heat on HN, where several people said Claude's output already exhausts a human's token limit. An Anthropic engineer replied in the thread that the writing is the thing they like most about 5.1, with fewer stock phrases and more reliable style-following. Both can be true: the model uses fewer tics and packs more into each sentence. Whether that reads as better depends on how fond you were of paragraph breaks.
The compute bill behind it
The night before the launch, Bloomberg and the Wall Street Journal reported a $35 billion, six-year cloud deal between Anthropic and Lambda, covering roughly 350 megawatts at a Hut 8 site in Nueces County, Texas, with Nvidia holding the lease and Lambda installing the chips, as summarized by SiliconANGLE. Grid connection is slated for the first quarter of 2027. A model whose own docs steer most users to the cheaper sibling still needs somewhere to run the users who ignore that advice.
Key Takeaways
- Fable 5.1 and Mythos 5.1 are one model with two safeguard levels. Fable is on the API, Bedrock, Google Cloud, and Microsoft Foundry today; Mythos is Glasswing only.
- List prices are unchanged at $10 in and $50 out, but cache reads fall from $1.00 to $0.25 per million, roughly 25% off typical bills and up to 45% off agentic ones.
- Terminal-Bench-Science jumps from 24.7% to 52.6%. ARC Prize's independent numbers show cheaper tasks and a flat-to-slightly-lower ARC-AGI-1 score.
- Same weights, different scores: Mythos 5.1 posts 60.9% on Terminal-Bench 4.0 to Fable 5.1's 55.8%, the cost of production safeguards.
- Three deliberate breaking changes (no forced tool use, one-way thinking blocks, append-only history) are anti-distillation measures. Audit any code that edits message history before migrating.
- Anthropic documents its own regressions: fewer parallel tool calls, whole-file rewrites, denser prose. Re-run your evals and re-tune effort before switching.
Sources: Anthropic announcement, Claude Fable 5.1 model page, What's new in Claude Fable 5.1, ARC Prize verified results, VentureBeat, MacRumors, SiliconANGLE, Hacker News discussion