Fable 5.1 and Mythos 5.1: One Model, Two Sets of Safeguards
Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1. They are the same underlying model. The difference is entirely in the safeguards: Fable 5.1 is generally available, while Mythos 5.1 is gated behind trusted access programs and carries permissions built for cyberdefenders and life sciences researchers.
The benchmark that stands out is Terminal-Bench-Science 0.1, where Fable 5.1 posts 52.6% against 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol. That is not an increment; that is a doubling of the previous best. Elsewhere: 55.8% on Terminal-Bench 4.0 (Mythos 5.1 hits 60.9% with looser cyber safeguards), 1853 on GDPval-AA v2, 60.9% on Humanity’s Last Exam without tools, 31.4% on AutomationBench against Opus 5’s 26.9%, and 73.4% on CursorBench 3.2.0.
The effort dial matters more than the headline numbers. Fable 5.1 at Low or Medium effort matches or beats Fable 5 at full effort, for meaningfully less money. It defaults to High in Claude Code, Medium in Claude Cowork and on Claude.ai. Anthropic’s own framing is that the model “avoids shortcuts” and fixes root causes — the illustrative case being an investment firm’s one-in-a-million crash that nobody had explained in four to five years, which Fable 5.1 traced by disassembling a vendor library and matching it against the core dump.
Enterprise Frontier Safeguards Answers the Data Retention Complaint
Anthropic introduced Enterprise Frontier Safeguards (EFS), its attempt to square two things enterprises have been told to choose between: zero data retention and state-of-the-art misuse detection. Under EFS, customer data sits in cloud infrastructure the customer controls, not Anthropic’s, and any human review is by default done by the customer.
It was built with more than 100 customers across financial services, healthcare, manufacturing, telecom, law, retail and the public sector, plus AWS, Google Cloud and Microsoft Azure. Coverage spans Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Google’s Agent Platform and Microsoft Foundry. Rollout is phased, starting this fall.
The interim provision is the practical bit: eligible customers can run Fable 5.1 and Fable 5 with zero data retention today, before EFS is ready. If your security team has been the thing blocking a Claude deployment, that objection now has a dated answer instead of a roadmap slide. Note also the phrasing Anthropic used to introduce all of this — price, data retention and safeguards were presented as customer feedback being addressed, which is a different posture than a capability announcement.
Anti-Distillation Lands, and the EU Watermark Gets a Detector
Fable 5.1 ships with strengthened anti-distillation mechanisms, and the specific change is worth reading carefully. API accounts created from today onward can no longer manually edit Claude’s prior context in a multi-turn conversation while preserving the transcript of Claude’s prior thinking. That closes a publicly documented technique for extracting reasoning traces at scale.
Existing accounts are not affected yet, but the change applies to all users on future model releases. Anthropic acknowledges a small number of custom integrations will break. If you have tooling that rewrites assistant turns mid-conversation — evaluation harnesses and multi-agent orchestrators both do this — check it before the next release, not after.
Separately, the EU AI Act Code of Practice obligations arrived in usable form. Models released after August 2, 2026 carry an invisible statistical watermark, and the required detection API is now in private preview for regulators, law enforcement, media, fact-checkers, researchers, educational organizations and EU civil society groups, plus enterprises with their own compliance obligations. Anthropic is explicit that the watermark carries no information about the user, their organization, or their conversations.
Cache Reads Drop 75% and That Is the Whole Price Story
Fable 5.1 costs $10 per million input tokens and $50 per million output tokens — identical to Fable 5. Nothing changed on the headline rate. What changed is cache reads, now $0.25 per million tokens, down 75%.
That single line does the work. Anthropic measured four weeks of actual August usage at default effort and puts the reduction at about 25% for typical workloads (Claude Enterprise, Claude Code and API combined) and up to about 45% for highly agentic workloads — the context-heavy, tool-heavy sessions where cache reads are most of the bill. Model the delta on your cache-read share, not the headline percentage, because the two numbers are 20 points apart for a reason.
The market read arrived fast. Cognition said it is moving Devin’s Opus 5 traffic to Fable 5.1 on launch day, starting with code review, on the grounds that the new cache-read pricing finally makes a Fable-class model economical for work they had kept on Opus. That is the actual mechanism here: this is a price cut aimed at moving Opus workloads down a tier, not at winning a rate-card comparison.
60% Fewer Cyber Interventions, and Vulnerability Discovery Is Now Allowed
The change most likely to alter your day: Fable 5.1 can now be used to identify software vulnerabilities. Not to develop exploits for them — that line is held — but the defensive half of security work is no longer collateral damage. Anthropic projects about 60% fewer safeguard interventions per Claude Code session compared to Fable 5.
Know what still routes to Opus models: penetration testing, exploit generation, and binary-based vulnerability scanning. Those remain dual-use and remain redirected. On the biology side, the newest safeguards fire 85% less often on benign elementary biology and medical questions, though life sciences R&D queries still go to Opus.
Claude Security — the codebase scanner that proposes patches for human review — is now powered by Mythos 5.1. And for teams that genuinely need the unfiltered version, the Cyber Verification Program is being extended to Mythos-class models, while the Life Sciences Verification Program has enrolled its first participants in partnership with the US government. Both are currently US-only, with expansion pending.
A Venus Map, Protein Binders, and a 2.5x Speedup on Genomics Models
Anthropic published three concrete results the models produced rather than three benchmark scores. Fable 5.1 trained a neural network to build a new elevation map of a third of Venus, working from 30-year-old NASA Magellan radar and an existing map covering a fifth of the planet. Resolution improved from 10–20 km down to 2–3 km, with heights up to 25% more accurate. It is released under Creative Commons on Zenodo, ahead of NASA’s VERITAS and ESA’s EnVision missions.
In protein design, Mythos 5.1 was given open-source design and folding tools and its outputs were sent for external experimental validation. On three targets its binders had 10x the affinity of the best entries to Adaptyv Bio’s competitions, and its hit rate reached nearly 50% across 12 targets — against a field where 10–15% is typical.
The one with the shortest path to your infrastructure: Mythos 5.1 wrote custom GPU kernels that sped up seven open-source deep learning models — ChromBPNet, Flashzoi, Enformer, Profluent-E1, ProGen2 and both Evo 2 sizes — by up to 2.5x with identical outputs, cutting estimated GPU cost on genome-wide analyses by 30–60%. Anthropic says it took days rather than the weeks a performance engineering team would need, and plans to open-source the optimizations. Pair that with last week’s Model Hardware Standard preview, which defines a common driver interface for agents to operate lab equipment, and the shape of the strategy is legible.
Twenty-Two Early Partners, and Claudeforce Opens This Month
Anthropic published 22 early-access partner quotes with the launch, which is an unusually large block and worth reading as a distribution map rather than testimonials. The pattern: Cognition moving Devin off Opus, Datadog using it for production incident root-cause analysis, Crosby reporting contract redlining scores jumping 47.9 to 57.0, Hebbia calling it the best deck-generation model it has tested, Browserbase at 82% of tasks on its hardest browser-agent benchmark against 74% for Opus 5 and 57% for Fable 5, and Glean reporting judges preferred it roughly 2-to-1 over Fable 5.
Notice how many of those are agent companies reporting fewer tokens for the same or better result — Rogo at 20% fewer tokens at matched accuracy, Block calling it “far more efficient per token than Opus 5,” Every describing it as roughly twice as fast as Opus 5 at half the tokens. Combined with the cache-read cut, the cost delta for an orchestration product is compounding, not additive.
Meanwhile the Claudeforce partnership announced with Salesforce on August 26 reaches open beta this month. Salesforce in Claude ships as a plugin with 37 prebuilt sales skills for reasoning over live revenue context and taking governed action from inside Claude; additional skills begin landing in late 2026. It is currently limited to select pilot customers.
The Release Where Anthropic Priced the Model Against Itself
Yesterday this brief noted that nothing in the news was about a model. Twenty-four hours later, everything is. But look at what Anthropic actually chose to lead with, because it is not the benchmark table. The launch post opens capabilities, then immediately pivots to price, data retention, and safeguards — framed explicitly as customer feedback being addressed. A frontier lab announcing its best model has never sounded quite so much like a procurement response.
The pricing structure gives that away. Input and output rates did not move. The entire reduction lives in cache reads, which is the line item that dominates exactly one thing: long-running agentic sessions. Anthropic did not cut the price of using Claude. It cut the price of leaving Claude running. Combined with a model that matches its predecessor’s full-effort output at Low or Medium effort, the strategy is to make the expensive tier structurally unnecessary. Cognition’s day-one migration of Devin traffic off Opus 5 is the intended outcome, stated out loud by the customer.
The safeguard changes carry more weight than the numbers suggest. Permitting vulnerability discovery while still blocking exploit development is a real line to draw, and it required Anthropic to trust its own classifiers enough to accept a class of request it previously refused wholesale. The 60% reduction in Claude Code interventions is the visible half; the invisible half is that the security research market was being handed to competitors every time a legitimate defensive query got blocked. Same logic on biology: 85% fewer false fires on elementary questions. These are not safety rollbacks — the dual-use categories still route to Opus — they are precision improvements that happened to also be revenue recovery.
Then there is the science. A Venus elevation map, protein binders at 50% hit rate against a 10–15% field, GPU kernels cutting genomics compute 30–60%. It is genuinely impressive work and it is also, unmistakably, S-1 material. A company weeks from a public filing needs a story about where revenue goes after coding assistants saturate, and “our model does original science, here is the Zenodo record and the external lab validation” is a considerably better answer than a TAM slide. The Model Hardware Standard preview, letting agents drive microscopes and liquid handlers directly, is the same bet with a longer fuse.
The thing to watch is the anti-distillation change, because it is the only item here with a cost to existing users. Blocking context editing for new API accounts is a defensive move against exactly the industrial-scale extraction Anthropic alleged against Qwen-linked accounts — but it also quietly closes a door that legitimate evaluation harnesses and multi-agent frameworks walk through daily. Existing accounts get a reprieve until the next model release. That is a deadline, not an exemption, and almost nobody reading a launch post about benchmark scores noticed it was there.