Summary
- Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1 — the same underlying model split into two versions with different safety guardrails
- Lower cache-read pricing cuts costs by about 25% for typical work and up to roughly 45% for heavy agentic tasks
- Enterprise Frontier Safeguards (EFS), rolling out this fall, will let customers store data on infrastructure they control themselves, while cybersecurity false positives have already dropped 60%
About three months after releasing Fable 5 on June 9, Anthropic announced two new models on September 1: Claude Fable 5.1, aimed at coding and knowledge work, and Claude Mythos 5.1, built specifically for cybersecurity and life sciences. The two are actually the same underlying model with different levels of safety tuning — Fable 5.1 is publicly available to anyone, while Mythos 5.1 is restricted to companies and researchers who've passed Anthropic's trusted access program. Anthropic is billing both as "the most advanced models in the world for coding and knowledge work," but the more notable part of the announcement is how directly it addresses complaints customers have raised about pricing, data retention, and safety controls.
Anthropic official website
To put it plainly: until last month, Anthropic's coding model Fable delivered strong quality but at a steep price, pushing developers to split work between Fable for design and cheaper models for repetitive tasks. Access controls and data retention policies had also led some companies and countries to reconsider how they routed their traffic. Fable 5.1 is a direct response to all three complaints.
Fable and Mythos: one model, two safety tiers
Claude Fable 5.1 and Claude Mythos 5.1 share the same base model. The difference lies in how tightly the safety guardrails are set. Fable 5.1 is publicly available for coding, knowledge work, and long-running problem solving, while Mythos 5.1 carries a separate set of safeguards designed to support cybersecurity and life sciences work, and is only available through the trusted access program. Anthropic said the gap between the two models' Terminal-Bench 4.0 scores comes from the older, less refined cyber safeguards interfering with certain tasks, and expects that gap to shrink significantly now that those safeguards have been improved. Fable 5.1 also ships with new distillation defenses. Starting today, newly created API accounts can no longer manually edit context while preserving Claude's prior reasoning trace across multi-turn conversations. That closes off a publicly known distillation technique in which thousands of fake accounts are used to extract a model's reasoning and build unguarded copies of it.
Pricing: cheaper cache reads cut costs by up to 45%
Anthropic estimates Fable 5.1 will run about 25% cheaper than Fable 5 for typical token-based workloads. That comes from a price cut on "cache reads" — the cost of re-reading input that's already been processed and stored — which dropped 75%, from $1 to $0.25 per million tokens. For long-running, high-intensity agentic tasks, the savings can climb to roughly 45%. Only the cache-read price changed, though; the base rates of $10 per million input tokens and $50 per million output tokens are unchanged from Fable 5.
As covered in last week's background piece, low-cost models like GLM 5.2 have emerged at roughly one-ninth of Fable's price, pushing developers to split their workloads across models. This price cut reads as Anthropic's answer to that shift.
One caveat worth noting: benchmarking firm Artificial Analysis points out that the savings vary depending on how the model is used. For agentic tasks dominated by cache reads, the savings come out to about $1.40 per task. But when effort is set to maximum, Fable 5.1 tends to output more tokens, which can push per-task cost about 20% higher than Fable 5. In other words, "up to 45%" is the ceiling figure under default settings with heavy cache reuse — not a universal number.
Usage limits reset for all users
Alongside the launch, Anthropic reset the 5-hour and weekly usage limits for all users. The official developer account ClaudeDevs announced the reset on the same day, meaning that regardless of how much of your existing limit you'd used, it starts fresh as of today — giving users room to try the new models without worrying about hitting a cap.
Data stays with the customer — introducing EFS
Anthropic is introducing a new data retention framework called Enterprise Frontier Safeguards (EFS). Under EFS, data is stored on cloud infrastructure controlled directly by the customer rather than by Anthropic, while still maintaining state-of-the-art abuse detection. Anthropic says this delivers full privacy on par with a zero-data-retention policy, and any human review that becomes necessary is designed to be handled by the customer by default, not by Anthropic. The system was built with more than 100 customers across finance, healthcare, manufacturing, telecom, legal, retail, and public sector, together with AWS, Google Cloud, and Azure, and it's supported across Claude Code, Claude Enterprise, Claude Platform, Amazon Bedrock, Google's agent platform, and Microsoft Foundry. EFS will roll out in phases starting this fall; until then, qualifying customers can use Fable 5.1 under zero-data-retention terms. Companies wanting early access can apply through the contact form.
Fewer false positives on safety, government partnership on biology
On the cybersecurity side, new safeguards have cut mistaken flags on harmless content, reducing interventions per session for Claude Code users by an average of about 60%. That's enabled Fable 5.1 to be used for defensive purposes like finding software vulnerabilities — though building exploits for those vulnerabilities, or dual-use tasks like penetration testing and binary vulnerability scanning, still get routed to the Opus model. Claude Security has now switched over to run on Mythos 5.1. Biology access remains tightly controlled. Fable 5.1's biology safeguards cut false positives on harmless requests — like introductory biology or medical questions — by 85%, but anything resembling R&D-level queries still gets routed to Opus. Advanced biology capabilities are only unlocked on Mythos 5.1 through the Life Sciences Verification Program (LSVP), which Anthropic built together with the U.S. government. The program has completed its first round of participant registration and plans to expand access soon.
Stepping into science: protein binders, a Venus map, and GPU kernels
Anthropic put the models' scientific research capabilities front and center in this announcement, calling the results "an early sign that AI is about to make a meaningful contribution to scientific discovery." Three examples were shared. The first is molecular design. Mythos 5.1 designed "binders" — molecules engineered to attach tightly to biological targets in the body — using only open-source protein design tools, and had them experimentally validated by two outside institutions. Across 12 targets, the hit rate came in at around 50%. That's the highest figure Anthropic has measured to date, well above the typical 10-15% hit rate seen in the field today, and on three targets the binding affinity was ten times higher than the top entries in the Adaptyv Bio design competition. The second is computational modeling. Fable 5.1 trained a neural network on radar imagery from NASA's Magellan probe, collected more than 30 years ago, to build a new high-resolution elevation map covering a third of Venus's surface. Resolution improved from the previous 10-20km down to 2-3km, and height accuracy improved by up to 25%. Anthropic released the map under a Creative Commons license to support upcoming NASA and ESA missions to Venus. The third is computational biology. Mythos 5.1 rewrote GPU kernels for seven open-source deep learning models, speeding them up by as much as 2.5x while leaving the results unchanged. For genome-wide analyses, that translates to a 30-60% cut in GPU costs — work that would normally take a team of performance engineers weeks, done in days.
What the benchmarks show
In Anthropic's own benchmark results, Fable 5.1 posted clear gains over its predecessor across several categories. On scientific research (Terminal-Bench-Science 0.1), it scored 52.6% — more than double Fable 5's score — and on terminal coding (Terminal-Bench 4.0), it climbed from 42.0% to 55.8%.
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Scientific research (Terminal-Bench-Science 0.1) | 52.6% | 24.7% | 29.0% | 22.4% |
| Agentic coding (Terminal-Bench 4.0) | 55.8% | 42.0% | 52.3% | 37.3% |
| Knowledge work (GDPval-AA v2) | 1853 | 1723 | 1824 | 1711 |
| Computer use (OSWorld 2.0, partial) | 77.9% | 72.9% | 75.4% | — |
| Multi-domain reasoning (HLE, with tools) | 65.0% | 63.8% | 63.6% | — |
| Work automation (AutomationBench) | 31.4% | 17.1% | 26.9% | 19.6% |
| Coding (CursorBench 3.2.0, max effort) | 73.4% | 70.5% | 70.0% | 67.2% |
| On Terminal-Bench 4.0, Mythos 5.1 scored slightly higher at 60.9%, which Anthropic attributes to the older, less refined cyber safeguards interfering with certain tasks in Fable 5.1's case. Early access customers reported similar impressions. Browserbase, which ran BrowserBench, said Fable 5.1 completed 82% of tasks in about 10 minutes using fewer tokens (compared to 74% for Opus 5 and 57% for Fable 5). On RedlineBench, a contract redlining benchmark, scores rose from 47.9 to 57.0, and on FrontierFinance, an investment research benchmark, from 49.2% to 55.9%. A portfolio manager at investment firm Millennium said a crash occurring once in a million runs had gone unexplained for four to five years — until Fable 5.1 disassembled a third-party vendor library, cross-referenced it against the core dump, and pinpointed the bug inside that library. |
How to try it
Fable 5.1 is available now through claude.ai, Claude Code, and Claude Cowork, and rolls out across all platforms including AWS, Google Cloud, and Azure starting today. To use it via API, head to platform.claude.com and start with the model claude-fable-5-1. The model offers three effort levels — Low, Medium, and High. Claude Code defaults to High, while Claude Cowork and claude.ai default to Medium. Anthropic says even the Low and Medium settings can match or beat Fable 5's High-effort results, at a much lower cost.
Mythos 5.1 isn't something the general public can access directly. It's currently limited to select organizations based in the U.S. — cybersecurity access requires applying through the CVP program, and biology access requires registering for the LSVP mentioned above.
For an example of Fable 5.1 already deployed in real coding work, see the article below.
Cognition's Devin adopts Fable 5.1, cuts coding task costs by 54%
Live demos: Fable 5.1 at work
Anthropic also released three short demo videos showing how Fable 5.1 operates in real workflows.
Editor's take
Read in one line, this update looks like Anthropic getting ahead of a problem before it could shoot itself in the foot. As we covered last month, Fable's high cost had already pushed developers toward cheaper models like GLM 5.2, while OpenAI was slashing GPT-5.6 pricing to go head-to-head with low-cost Chinese models. Cutting prices by up to 45% here reads less like a defensive move and more like getting ahead of the curve. The generational comparison makes the shift clearer. Where Fable 5 only delivered usable results at High effort, Anthropic says Fable 5.1 now clears that bar even at Low and Medium settings. Anyone who's put a model this size to real work tends to land on the same conclusion: what actually differentiates cost isn't peak performance, it's how well a model holds up at lower settings. For enterprises, cost is set by the default configuration, not the top-of-the-line one. That leaves three practical takeaways. First, teams worried about data retention should watch the phased EFS rollout and check now whether they qualify to switch to zero-data-retention terms. Second, security teams should spend a few weeks tracking how much that 60% drop in interventions actually reduces alert fatigue in practice. Third, when evaluating pricing, don't anchor on the "up to 45%" ceiling — calculate savings based on how much cache reuse your own workload actually generates. What stands out most is that Anthropic chose to lead this announcement with scientific research. The protein binders, the Venus map, the GPU kernel optimizations — none of it shows up in coding benchmarks, but all of it signals where the company thinks the next round of competition will play out. The tug-of-war between pricing and safeguards isn't going away anytime soon, and alongside it, a new front is opening up around what it means to be a "model that does science."
Comments