Jonathon's AI Wiki

Claude Fable 5.1 + Mythos 5.1 — Cheaper Cache Reads, ZDR, and Looser Cyber Gates

10 min read

Source: ai-research/anthropic-claude-fable-mythos-5-1-announcement-2026-09-01.md — Anthropic’s launch page (anthropic.com/claude-fable-and-mythos-5-1; the page shows “September 2026”, and the system card is dated 2026-09-01), fetched 2026-09-29. Claude Code facts from v2.1.257; migration facts from the claude-api skill update (anthropics/skills #1704).

Claude Fable 5.1 (claude-fable-5-1) and Claude Mythos 5.1 (claude-mythos-5-1) are the same model with different safeguards: Fable 5.1 is generally available, while Mythos 5.1 is available only through trusted-access programs for cybersecurity and life-sciences work. Anthropic frames the release around the three customer complaints that capped Fable 5’s adoption — price, data retention, and safeguards — and addresses each directly. Three weeks later Opus 5.5 claimed Fable-5.1-level work on most tasks at 40% of the per-token price, so Fable 5.1’s case now rests on the hardest long-horizon work.

Key Takeaways

  • Same list price, much cheaper cache reads. 50 output per MTok (unchanged from Fable 5); cache reads $0.25 per MTok, 75% less. Anthropic’s estimate from four weeks of real August usage: ~25% cheaper for typical workloads, up to ~45% for highly agentic, context-heavy work where cache reads dominate the bill.
  • Zero data retention is back for eligible customers, with a longer-term replacement. Enterprise Frontier Safeguards (EFS) stores data “in cloud infrastructure controlled entirely by the customer,” with any human review done by the customer by default. It rolls out in phases “beginning later this fall” across Claude Code, Claude Enterprise, the Claude Platform, Bedrock, Google’s Agent Platform, and Microsoft Foundry. Until then, eligible customers can use Fable 5.1 (and Fable 5) with zero data retention.
  • Cyber safeguards fire far less. 60% fewer cyber false positives, and Fable 5.1 may now be used to discover software vulnerabilities (not to develop exploits). Claude Code users should see “around 60% fewer interventions per session.” Penetration testing, exploit generation, and binary-based vulnerability scanning still redirect to Opus models. Biology safeguards for Fable 5 and 5.1 now fire 85% less often on benign elementary biology and medical questions.
  • Default effort differs by surface: High in Claude Code, Medium in Claude Cowork and Claude.ai. Anthropic says Fable 5.1 at Low or Medium matches or beats Fable 5 “at a much lower cost.”
  • Anti-distillation: new API accounts can no longer edit history and keep the thinking. For accounts created from launch day onward, manually editing Claude’s prior context in a multi-turn conversation while preserving its prior thinking is blocked. Existing accounts were not affected at launch, “though it will apply to all users with future model releases.” Opus 5.5 extended the same mechanism, preserved thinking, to accounts created on or after 2026-08-31.
  • Watermarked output (EU AI Act). Models released after 2026-08-02 carry an invisible watermark under the EU Code of Practice on Transparency of AI-Generated Content, which Anthropic signed in July 2026. A detection API is in private preview for regulators, media, researchers, and obligated enterprises. See Claude Output Watermarking.
  • Mythos 5.1 is identical but gated. Available through the Cyber Verification Program (Mythos-class access “in the near future”) and the Life Sciences Verification Program, built with the US government. At launch it was limited to a set of US organizations. Claude Security (codebase vulnerability scanning) now runs on Mythos 5.1 — see Claude Security.

Benchmarks (vendor-reported)

BenchmarkFable 5.1Fable 5Opus 5GPT-5.6 Sol
Terminal-Bench-Science 0.152.6%24.7%29.0%22.4%
Terminal-Bench 4.055.8% (Mythos 5.1: 60.9%)42.0%52.3%37.3%
GDPval-AA v2 (Elo)1853172318241711
OSWorld 2.0 partial / strict77.9% / 41.7%72.9% / 36.1%75.4% / 39.6%—
Humanity’s Last Exam no tools / tools60.9% / 65.0%57.8% / 63.8%56.6% / 63.6%—
AutomationBench31.4%17.1%26.9%19.6%
CursorBench 3.2.073.4%70.5%70.0%67.2%
  • The Fable/Mythos gap on Terminal-Bench 4.0 (55.8 vs 60.9) is the safeguard cost. Anthropic: the two are the same model, and “the gap between them reflects the tasks on which our earlier, less precise cyber safeguards intervened.” It expects the gap to shrink with the new safeguards.
  • Evaluated with production safeguards on: where they intervened, OSWorld and AutomationBench scored those tasks as zero, and other flagged tasks were completed by Opus 4.8 (cyber) or Opus 5 (biology).
  • These numbers use different benchmark versions from the Opus 5.5 table (GDPval-AA v2 here vs v2.1 there; CursorBench 3.2.0 vs 4.0), so the two tables are not directly comparable.

What testers reported (vendor-selected)

  • Root-cause debugging. Millennium: Fable 5.1 found the cause of a rare crash (about one in a million runs) that its engineers had not explained in four to five years. It disassembled a vendor library, matched it against the core dump, and traced the crash to a bug in that library.
  • Long unattended runs. One tester: a 38-hour unattended ML run that diagnosed a prior result as a label artifact, corrected it, and ran six parallel experiments overnight. Another: “it keeps its own records, reprioritizes as things change, and picks up where it left off.”
  • Economics changed routing. Devin’s team: “We’re moving our Opus 5 traffic in Devin to Claude Fable 5.1 on launch day… with the new cache read pricing a Fable-class model is finally economical for the workloads we’d kept on Opus, starting with code review.” Another tester: “Fable-level intelligence, Opus-level price, Sonnet-speed… about twice as fast as Opus 5 and used half as many tokens.”
  • Documents and decks. One tester reported the best PowerPoint decks of any model it had tested. A contract-redlining benchmark rose from 47.9 to 57.0, and FrontierFinance from 49.2% to 55.9%, credited to going “straight to the call transcript” instead of secondary coverage.

Science results (Mythos 5.1)

  • Protein binders: with open-source design and folding tools, Mythos 5.1’s designs had binding affinities 10× higher than the best Adaptyv Bio competition entries on three targets, and a hit rate of nearly 50% across 12 targets (10–15% is typical), confirmed by two external labs.
  • Venus elevation map: Fable 5.1 trained a network on 30-year-old Magellan radar data to map a third of Venus at 2–3 km detail (was 10–20 km), released under Creative Commons.
  • GPU kernels: custom kernels and caching sped up seven open-source protein and genomics models by up to 2.5× with identical outputs, cutting estimated GPU costs for genome-wide analyses by 30–60%.

Field reports, first month (creators and press, not controlled tests)

  • Subagents inherit Fable by default, and that is where the bill goes. Nathaniel Whittemore (The AI Daily Brief): Fable 5.1’s default is to use “the most advanced model as the coordinator, but also … for all the sub agents,” so a routine research task that spins up five or ten subagents can use “a major percentage” of a usage cap (raw/7_Ways_How_We_Use_AI_Is_Changing.md). Leo Laporte’s fix: tell it “if you create a sub agent, don’t use Fable 5.1. Use Opus or Sonnet,” since subagents are usually “collecting stuff or using tools” (Intelligent Machines, raw/Infinite_Slop_-_The_End_of_Apps_Meet_the_Agentic_OS_Built_for_Your_Phone.md). See Subagents for pinning a subagent’s model.
  • The 25% saving did not show up on one independent harness. Matt Wolfe, reading Artificial Analysis’s cost-per-task chart in launch week: Fable 5.1 was the most expensive model per task at 3.14 for Fable 5 (raw/AI_News_-_The_Most_Insane_Week_So_Far_This_Year.md). Anthropic’s estimate assumes cache-read-heavy workloads; a single-shot benchmark with little cache reuse would not see it.
  • Low effort for drafts, extra for the finish. Nate B Jones ran the same valuation task at low and extra effort. Low produced a finished workbook and deck but “no sources sheet… no check sheet”; extra added linked sources and questions “that could change the investment decision.” His suggested pipeline: build the first version on Fable low, have GPT-5.6 Sol check the structure, then finish on Fable extra (raw/Everyone_s_Testing_Claude_Fable_5.1_On_Code._It_Made_Me_A_37-Second_Film..md). He also found “Claude’s limits still feel much tighter than OpenAI’s,” though 5.1 “went way farther than 5 inside those limits.”
  • Computer use is the gap testers route around. In a head-to-head with GPT-6 Astra, Jones uses Fable to flesh out work but goes to Astra in Codex for computer use: “the speed at which that model can use the computer, it’s night and day” (raw/The_Race_to_Done_-_Fable_5.1_vs_GPT-6_Astra._Who_Wins.md).

Migrating from Fable 5

From the claude-api skill’s “Migrating to Claude Fable 5.1 from Claude Fable 5” section (anthropics/skills #1704, 2026-09-01):

  • Three breaking changes: (1) forced tool_choice of any or tool returns a 400; (2) thinking blocks are preserved only for the model that produced them or a newer one; (3) they are preserved only in the conversation that produced them, “so edited history replayed with thinking blocks is rejected.”
  • New API surface documented alongside: per-message effort (mid-conversation output-config beta), turn-scoped mid-conversation system messages with clear_at, progress updates between tool calls via thinking.display: "updates", thinking block_binding controls, and a 0.025× cache-read rate on Fable 5.1 with a max_tokens: 0 keep-alive that “usually beats the 1-hour TTL.”
  • Launch-day hedges the skill kept open: Task Budgets support on Fable 5.1, whether Mythos 5.1 shares the 0.025× cache-read rate, and the fallback-credit wording.
  • Claude Code v2.1.257 (2026-09-01): claude-fable-5-1 added and made the default Fable model (1M context, 50, $0.25 cache reads). In Claude apps gateway sessions, fable and best kept resolving to Fable 5 for now, because gateways not yet configured for 5.1 reject it; pick Fable 5.1 in /model explicitly. See Week 36.

Implementation

Tool/Service: Claude Fable 5.1 (claude-fable-5-1); Claude Mythos 5.1 (claude-mythos-5-1, trusted access only)

Setup: Generally available on the Claude API, Amazon Web Services, Google Cloud, and Microsoft Azure. Claude Code v2.1.257+ uses it as the default Fable model except behind apps gateways (see above).

Cost: 50 output / $0.25 cache read per MTok.

Integration notes: Check any integration that forces tool_choice or replays edited history with thinking blocks before switching model IDs. Eligible enterprises should ask about zero data retention now rather than waiting for EFS.

Try It

  1. Re-price your Fable workloads. If your Fable bill is mostly cache reads (long agentic sessions), the ~45% estimate is the one that applies to you. Measure it on a week of real usage.
  2. Retry security work you stopped doing on Fable 5. Vulnerability discovery is now allowed, and interventions dropped ~60% per session.
  3. Pin subagents to a cheaper model. Put model: sonnet or model: opus in your subagent definitions, or say so in the prompt, so a Fable 5.1 coordinator doesn’t fan out Fable subagents.
  4. Compare against Opus 5.5 before committing. Opus 5.5 (2026-09-22) claims Fable-5.1-level results on most work at 20. Keep Fable 5.1 for the tasks where it measurably wins.

Open Questions

  • EFS timing and eligibility. “In phases, beginning later this fall”; which customers are “eligible” for interim ZDR is not defined on the launch page.
  • Does ZDR end the adoption plateau? Anthropic’s Position, August 2026 recorded two competing explanations for Fable 5’s ~11% share of Anthropic tool spend: price or data retention. Anthropic naming both as customer feedback and fixing both at once means spend data after this release cannot separate them. ^[inferred]
    • First spend data point (2026-09): Ramp data, relayed by The AI Daily Brief, had Fable 5.1 at 22.5% of enterprise spend “and rising very quickly” (raw/The_AI_Challenges_Businesses_Are_Actually_Focused_On_Right_Now.md). That fits either explanation, since the release also cut running cost 25–45%.
  • Knowledge cutoff, max output, and tokenizer are not stated on the launch page; the system card (2026-09-01) was not read for this article.
  • SDK model constant. No SDK release in this batch names claude-fable-5-1; whether a constant landed is unconfirmed.

Graph View

Backlinks