⏱️ Lectura: 10 min
Anthropic cut the price of Claude Fable 5.1 by up to 45% compared to Fable 5 for intensive agentic work, and launched it the same day as Claude Mythos 5.1, the version with reinforced safeguards for cybersecurity and life sciences. Both models share the same core: what changes is how much of a safety filter runs on top of each response.
📑 En este artículo
- TL;DR
- What Happened
- Context and History
- Technical Details and Performance of Claude Fable 5.1
- How to Start Testing It
- Impact and Analysis
- What’s Next
- Frequently Asked Questions
- What’s the difference between Claude Fable 5.1 and Claude Mythos 5.1?
- How much does it cost to use Claude Fable 5.1 compared to Fable 5?
- What is Enterprise Frontier Safeguards (EFS)?
- Can I access Claude Mythos 5.1 as an independent developer?
- Can Fable 5.1 develop exploits for the vulnerabilities it finds?
- What effort level does Fable 5.1 use by default?
- References
The company also debuted a data retention scheme designed for enterprise customers, Enterprise Frontier Safeguards, and adjusted its cybersecurity safeguards to block fewer false positives. Fable 5.1 is already available to any developer; Mythos 5.1 is only accessible through trusted access programs.
TL;DR
- Anthropic launched Claude Fable 5.1 (available to everyone) and Claude Mythos 5.1 (restricted access): same model, different safeguards.
- Fable 5.1 costs 25% less than Fable 5 for typical workloads and up to 45% less for intensive agentic work.
- The savings come from lowering the price of cache reads, according to Anthropic.
- New Enterprise Frontier Safeguards (EFS) scheme: zero data retention using the customer’s own infrastructure, starting this fall.
- Cybersecurity safeguards block 60% fewer false positives than before.
- Fable 5.1 can now search for software vulnerabilities, but Anthropic says it cannot develop exploits.
- On Terminal-Bench 4.0, Fable 5.1 scored 55.8% and Mythos 5.1 60.9%, compared to 37.3% for Fable 5.
- Fable 5.1 uses high effort by default in Claude Code and medium in Claude Cowork and claude.ai.
What Happened
On September 1, Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1 as the same model with two different safeguard levels. Fable 5.1 has general availability: any developer can use it through the API, Claude Code, Claude Cowork, or claude.ai. Mythos 5.1, on the other hand, is only delivered through trusted access programs, designed specifically for work in cybersecurity and life sciences.
According to Anthropic, the performance gap between the two models on cybersecurity tasks reflects cases where the previous, less precise safeguards interrupted the model prematurely. With the improvements announced today, the company expects that gap to narrow.
In biology, Anthropic confirmed an access program developed together with the US government to enable Mythos 5.1’s advanced capabilities in that field. Enrollment for scientists will open soon, though the company did not give an exact date.
Context and History
Fable and Mythos are not new names. The previous generation, Fable 5 and Mythos 5, already followed the same scheme: a base model and a variant with stricter controls for sensitive use cases. What changes with the 5.1 revision is a direct response to three recurring customer complaints: price, data retention, and excessive safety blocks.
The effort-level system (low, medium, high, xhigh, max) isn’t new either: it was introduced with the Fable 5 family so developers can decide how much reasoning, and how much cost, they want to pay for each task. Fable 5.1 inherits that scheme and, according to Anthropic, at low or medium levels it matches or beats Fable 5’s results at a much lower cost.
The Fable/Mythos distinction matters because it separates two audiences with opposite needs: most product teams want a model with standard safeguards and immediate availability (Fable), while security and biology researchers need a model that doesn’t automatically block legitimate but sensitive queries (Mythos), in exchange for going through a verification process.
Technical Details and Performance of Claude Fable 5.1
The numbers Anthropic published show a consistent improvement over Fable 5 across nearly all internal benchmarks. On Terminal-Bench-Science 0.1, a test of agentic scientific research, Fable 5.1 scored 52.6% versus 24.7% for Fable 5. The public leaderboard for that same benchmark, using a Claude Code harness and three runs per task, places Opus 5 at 30.0% and Fable 5 at 21.4%; Anthropic says it reproduced those results at 29.0% and 24.7% respectively, within the margin of error.
On general knowledge work (GDPval-AA v2), Fable 5.1 reached 1853 points versus 1723 for Fable 5 and 1824 for Opus 5. On computer use (OSWorld 2.0), Fable 5.1 hit 77.9% on the partial metric and 41.7% on the strict metric, compared to 72.9% and 36.1% for Fable 5.
The table below summarizes the key differences between the three versions available today:
| Model | Availability | Terminal-Bench 4.0 | CursorBench 3.2.0 |
|---|---|---|---|
| Claude Fable 5.1 | General, for any developer | 55.8% | 73.4% |
| Claude Mythos 5.1 | Trusted access programs (cybersecurity and biology) | 60.9% | Not published |
| Claude Fable 5 | Previous version, still available | 37.3% | 70.5% |
On multidisciplinary reasoning (Humanity’s Last Exam), Fable 5.1 with tools reached 65.0% versus 63.8% for Fable 5 and 63.6% for Opus 5. On enterprise workflows (AutomationBench), the gap was more pronounced: 31.4% for Fable 5.1 against 17.1% for Fable 5.
💭 Key: Fable 5.1 and Mythos 5.1 are, according to Anthropic, the same model. The score difference on Terminal-Bench 4.0 (55.8% versus 60.9%) isn’t extra capability: it’s how much the safeguard layer lets it work before stepping in.
The example Anthropic uses to illustrate this capability is investment firm Millennium: Fable 5.1 found the cause of an intermittent bug in its internal systems that no engineer, and no other model, had managed to explain after several years of attempts.
How to Start Testing It
Using Fable 5.1 doesn’t require anything special: it’s just another model within the Anthropic API, selectable by its identifier. The first step is installing the SDK or, if you prefer working from the terminal, Claude Code:
npm install -g @anthropic-ai/claude-codeThat same command works the same way on Windows (PowerShell or cmd, with Node.js installed), macOS, and Linux, because npm doesn’t depend on the operating system.
With an API key configured, a direct curl call looks like this:
curl -X POST https://api.anthropic.com/v1/messages -H "x-api-key: $ANTHROPIC_API_KEY" -H "anthropic-version: 2023-06-01" -H "content-type: application/json" -d '{"model": "claude-fable-5-1", "max_tokens": 1024, "thinking": {"type": "enabled", "effort": "high"}, "messages": [{"role": "user", "content": "Find the root cause of this stack trace"}]}'The thinking.effort parameter controls how much the model reasons before responding. Claude Code uses high by default; for simple tasks, lowering it to medium or low reduces cost without losing much quality, according to Anthropic’s own benchmarks.
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-fable-5-1",
max_tokens=2048,
thinking={"type": "enabled", "effort": "medium"},
messages=[
{"role": "user", "content": "Review this authentication module and tell me if there's a race condition"}
],
)
print(response.content[0].text)That script calls Fable 5.1 with medium effort to review a code snippet and look for race conditions. The response.content[0].text field returns the response as plain text.
To confirm you’re paying the new price, check the usage field in the response: cache_read_input_tokens is the part Anthropic says it made cheaper. If your monthly bill didn’t drop compared to Fable 5 for workloads with a lot of repeated context, it’s worth checking whether your client is reusing the cache correctly.
Impact and Analysis
For teams already using Claude in production, the most immediate change is cost: lowering the price of cache reads hits any agentic workflow directly, where the same context (a repository, a long conversation) gets resent at every step. Anthropic estimates the biggest savings there, up to 45%.
The second change, Enterprise Frontier Safeguards, targets a different problem: companies that want total privacy (zero retention) without losing the anti-abuse controls Anthropic applies to traffic. The solution the company proposes is moving storage to infrastructure the customer itself controls. It’s an uncommon design: zero retention normally means less visibility for the provider, not a change in where the data lives.
⚠️ Heads up: EFS still isn’t available to everyone. Anthropic says it will roll out in phases starting this fall; until then, eligible customers can use Fable 5.1 with zero data retention, but without the rest of the EFS scheme.
For security researchers, the most relevant adjustment is the reduction in false positives: 60% fewer than before in the cybersecurity safeguards. Anthropic says this is possible partly because Fable 5.1 can now be used to discover software vulnerabilities, though company policy still prohibits the model from developing exploits for those same vulnerabilities. It’s a distinction that’s hard to audit from the outside: it depends on the model consistently respecting that line, not on an impenetrable technical barrier.
The clearest limitation for developers in Latin America is access to Mythos 5.1: it’s not a model you can request with an API key like Fable 5.1. You have to qualify for a trusted access program, a process Anthropic hasn’t detailed in terms of timelines or public criteria. If your use case is legitimate offensive security research (pentesting, bug bounty) without that access, you’ll keep running into Fable 5.1’s standard limits.
What’s Next
Anthropic confirmed two relative, not exact, dates: EFS starts reaching enterprise customers later this fall, and enrollment for Mythos 5.1’s biology program will open soon for scientists. Neither has a public timeline yet.
What is measurable right now is the performance gap between Fable 5.1 and Mythos 5.1 on cybersecurity tasks (55.8% versus 60.9% on Terminal-Bench 4.0). Anthropic says that gap should narrow as safeguards improve, so it’s a number worth revisiting at the next update.
📖 Summary on Telegram: View summary
Try it yourself: run the curl example above with your own API key and compare the cost of cache_read_input_tokens against your last Fable 5 invoice.
Frequently Asked Questions
What’s the difference between Claude Fable 5.1 and Claude Mythos 5.1?
According to Anthropic, they’re the same model with different safeguard levels. Fable 5.1 has general availability; Mythos 5.1 is only delivered through trusted access programs for cybersecurity and life sciences.
How much does it cost to use Claude Fable 5.1 compared to Fable 5?
Anthropic estimates 25% less for typical workloads and up to 45% less for intensive agentic work, thanks to a lower price on cache reads.
What is Enterprise Frontier Safeguards (EFS)?
It’s a data retention scheme that provides privacy equivalent to zero retention, but stores the information on infrastructure controlled by the customer instead of by Anthropic. It rolls out in phases starting this fall.
Can I access Claude Mythos 5.1 as an independent developer?
Not directly. Mythos 5.1 requires qualifying for a trusted access program focused on cybersecurity or life sciences; Anthropic hasn’t published detailed eligibility criteria.
Can Fable 5.1 develop exploits for the vulnerabilities it finds?
Anthropic says no: the model can discover software vulnerabilities, but its safeguards are designed to block the development of exploits for those same flaws.
What effort level does Fable 5.1 use by default?
High in Claude Code, and medium in Claude Cowork and claude.ai, according to Anthropic’s published documentation.
References
- Anthropic: Introducing Claude Fable 5.1 and Claude Mythos 5.1: official announcement with pricing, benchmarks, and safeguard policy.
- Anthropic API Documentation: reference for parameters like thinking.effort and cache handling.
- anthropic-sdk-python on GitHub: source code for the official SDK used in this article’s examples.
📱 Like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.
Imagen destacada: Foto de Igor Omilaev en Unsplash
0 Comments