anthropic.com via Reddit

Anthropic Fable 5 Runs Stripe Migration in One Day

4 sources tracking this story
anthropic generative ai coding tools safety agents model-release ai-safety agent-infrastructure pricing

Key insights

  • Fable 5 scores 80.3% on SWE-Bench Pro versus GPT-5.5's 58.6% and Opus 4.8's 69.2%, with FrontierCode at 29.3% against Opus 4.8's 13.4%.
  • Safety-routed tokens are billed at Opus 4.8 rates rather than Fable 5 rates, making per-token cost variable for any workload that triggers the fallback.
  • The 30-day traffic retention requirement sends data outside AWS's security boundary and supersedes existing zero-retention enterprise contracts.

Why this matters

Benchmark data from The Decoder confirms Fable 5 outpaces GPT-5.5 by 21.7 points on SWE-Bench Pro and more than doubles Opus 4.8 on FrontierCode, with Mythos 5 scoring 78% on ExploitBench. AWS's deployment post reveals that safety-routed tokens are billed at Opus 4.8 rates rather than flat Fable 5 pricing, adding a variable cost layer that enterprises building on the published $10/M figure must model separately. Mythos 5 ships initially through Project Glasswing with US government collaboration, meaning the publicly priced $50/M-token tier is not unconditionally accessible at launch. Hacker News practitioners are not focused on pricing or benchmarks but on the silent routing mechanism: there is no signal to users or production logs when the safety fallback fires and an Opus 4.8 response replaces what Fable 5 would have returned.

Summary

Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9, both built on the same Mythos-class architecture but differing only in safeguard configuration. Fable 5 uses AI classifiers that redirect fewer than 5% of sessions involving cybersecurity, biology, and chemistry to Claude Opus 4.8 rather than refusing outright. Mythos 5 drops those restrictions entirely, currently limited to Project Glasswing cybersecurity partners, with biology researcher access planned as the next expansion. Essentially: (Anthropic, Stripe) the performance headline is Stripe completing a 50-million-line Ruby codebase migration in one day versus a two-month manual estimate. - Pricing lands at $10 per million input tokens and $50 per million output tokens, less than half the previous Mythos Preview rate. - Protein design experts using Mythos 5 accelerated aspects of drug design by around ten times, with nine of 14 protein targets yielding strong candidates. - Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans through June 22, after which usage credits apply. Access controls, not raw capability, will define who benefits as Mythos-class expands into biology and scientific research.

Potential risks and opportunities

Risks

  • UK AISI made initial progress toward a universal jailbreak for Fable 5 during red-teaming; if completed and published, it could invalidate the classifier-based safety model before Anthropic can patch.
  • Enterprise teams on subscription plans face a billing change after June 22, when Fable 5 shifts from included to usage-credit billing, potentially disrupting budget plans for organizations already deploying at scale.
  • Mythos 5's unrestricted biology access, gated to Project Glasswing today, could face regulatory scrutiny if a partner misuses the model before the broader trusted access program's vetting criteria are formalized.

Opportunities

  • Cursor and GitHub, both cited as early Fable 5 testers, are positioned to lock in enterprise contracts before competitors can benchmark equivalent autonomous engineering capability.
  • Protein design firms and biotechs outside Project Glasswing can begin queuing for Mythos 5 biology researcher access, which Anthropic identifies as the next planned expansion of the trusted access program.
  • AI safety tooling vendors could propose complementary classifier layers or audit infrastructure aligned to Anthropic's 30-day data retention and oversight requirements for Mythos-class deployments.

What we don't know yet

  • No numeric benchmark scores comparing Fable 5 to Opus 4.8 are published; the article cites performance leads qualitatively but omits figures like SWE-Bench or equivalents.
  • Whether Project Glasswing partners are contractually bound to the 30-day data retention policy, and what audit mechanism enforces compliance.
  • When biology researchers will gain Mythos 5 access through the planned trusted access program, and what vetting criteria will be applied.

What others are reporting

Coverage cluster as of 24h after publish

  1. Amazon Web Services Read →

    Bedrock billing mechanics, region availability (US East N. Virginia, Europe Stockholm), API endpoint details (bedrock-mantle vs bedrock-runtime), and data-retention compliance specifics absent from the Anthropic announcement.

    Claude Fable 5 makes Mythos-level capabilities available to customers, with strong safeguards designed to make it safe for broader use.
  2. The Decoder Read →

    Full benchmark table: SWE-Bench Pro (80.3%), FrontierCode (29.3%), ExploitBench (Mythos 5: 78%), plus third-party validations from IMC trading and an E. coli protein study.

    Fable 5 beats every generally available model the company has ever shipped and claims state-of-the-art results in nearly all benchmarks tested.
  3. Hacker News Read →

    Practitioner criticism of the silent safety routing: users cannot distinguish Fable 5 from Opus 4.8 fallback responses, raising vendor-trust and observability concerns for production workloads.

    There's already an obvious stench to 'you should scale down your engineering team to a skeleton crew whose core competency is using our product.'

Shared on Bluesky by 6 AI experts (top 5 by trust)

Originally reported by anthropic.com

Read the original article →

Original headline: Anthropic Launches Claude Fable 5 and Claude Mythos 5 — First Public Mythos-Class Model Compresses Stripe's Two-Month Codebase Migration Into One Day, Priced at $10/$50 Per Million Tokens