# Claude Fable 5：首个 Mythos 模型--强大、昂贵且经过严格过滤

- 来源：The Decoder：AI News（RSS）
- 作者：Maximilian Schreiner
- 发布时间：2026-06-10 21:34
- AIHOT 分数：76
- AIHOT 链接：https://aihot.virxact.com/items/cmq84jl6i04h2slepcmxdpv9i
- 原文链接：https://the-decoder.com/claude-fable-5-the-first-mythos-model-is-powerful-expensive-and-heavily-filtered

## AI 摘要

Anthropic 发布 Claude Fable 5，这是新 Mythos 类别的首个模型。该模型在 SWE-bench Verified 上达到 95%，几乎在所有基准测试中领先，但成本是 Opus 4.8 的两倍，每百万 token 价格为 10 美元或 50 美元。严格的安全过滤器会阻止约 9% 的请求，同时新增 30 天数据保留政策，即使签订零数据保留合同也适用。

## 正文

Claude Fable 5: The first Mythos model is powerful, expensive, and heavily filtered

Anthropic has released Claude Fable 5, the first publicly available model in its so-called Mythos class. Early tests show a major leap in coding performance, but safety filters, pricing, and data retention policies are drawing sharp criticism.

With Claude Fable 5, Anthropic has shipped a model that tops nearly every benchmark. Fable 5 is the first publicly available version of the "Mythos class." According to Anthropic, Fable shares its base model with Claude Mythos 5 but adds strict guardrails that block potentially harmful requests related to cybersecurity, biology, chemistry, and model distillation. Mythos 5 is also available but limited to a small group of users.

What "Mythos" actually means on a technical level is mostly guesswork. Every CEO Dan Shipper, whose team had early access, reports that Anthropic staff told him there's nothing special about the architecture. Within the Haiku, Sonnet, and Opus family, Mythos simply refers to the largest and most capable model. Developer Simon Willison suspects the same, that it's the biggest Anthropic model publicly available to date. Fable just feels "big," Willison writes, "not just in terms of speed and cost, but also in how much it knows." Artificial Analysis backs this up: on its AA-Omniscience knowledge and hallucination benchmark, Fable scores 40 points, seven more than the previous leader, Gemini 3.1 Pro. Among open-weight models, that kind of gap typically tracks with model size.

Benchmark leader, but skeptics aren't convinced

Fable 5 sits atop nearly every leaderboard. On the Artificial Analysis Intelligence Index, it hits 64.9 points, roughly five ahead of GPT-5.5 as the closest competitor. On GDPval-AA, an agentic benchmark for real-world work tasks, it posts an Elo score of 1,932. On Humanity's Last Exam, Fable reaches 53 percent, more than seven points above Opus 4.8. A single run of that test cost about $2,200, including fallback costs.

The evaluation service Vals ranks Fable 5 first on its overall index and across all coding benchmarks, including SWE-bench Verified at 95 percent and Vibe Code Bench at 90.35 percent. That last number stands out: six months ago, no model cracked 20 percent. The coding tool Devin also reports a top score on its internal FrontierCode benchmark.

Still, parts of the community aren't buying it. On Hacker News, some users call the jump incremental rather than revolutionary and point to possible benchmark overfitting. Willison himself admits his impressions are "all vibes, if you want a more scientific comparison you'll have to look elsewhere," since he didn't run a proper side-by-side test.

A "warp drive" for big tasks, when it works

The vibes, however, seem to hit right: Ethan Mollick had Fable build a fully researched isochrone travel time map using Claude Code. The model spun up cheaper sub-agents on its own, pulling data on over 2,200 flight routes, train schedules, and road speeds from academic papers, all while writing code and having other agents test it in parallel. Another project, a research tool for calibrating human and AI judgments, took nine and a half hours.

At Every, single prompts produced a walkable 3D rendering of Borges' "Library of Babel" and a survey analysis that, according to Shipper, pinpointed a conversion problem more precisely than weeks of human work. On the company's senior engineer benchmark, Fable scored 91 out of 100, compared to 63 for Opus 4.8 and 62 for GPT-5.5. Shipper calls the model a "warp drive": ideal for tackling large, well-defined tasks asynchronously, but a poor fit for quick back-and-forth interaction. Willison kept his first impression simple: Fable is "a beast."

That strength comes with a tradeoff. Mollick describes how little he contributed himself and how few of the model's hundreds of micro-decisions he could actually follow. He went from being the wizard casting a spell to being the client signing a check: "I describe what I want, I pay for it, and I judge the result." The code review service CodeRabbit confirms Fable's strength on underspecified, autonomous coding tasks but warns that it falls behind Opus 4.8 on code review precision and tends to run tasks until the system kills them.

Safety filters make the model useless for many researchers

The most common complaint, by far, is the guardrails. Fable automatically falls back to the weaker Opus 4.8 or refuses to respond when it suspects sensitive topics. According to Artificial Analysis, this happens on roughly eight to nine percent of tasks, mostly scientific ones.

In practice, users report the filters flagging harmless requests constantly. A medical physicist writes: "I genuinely can't use Fable. I'm a medical physicist. I use the word nuclear a lot." Others describe MRI image segmentation being classified as bioterrorism, a question about malaria transmission by mosquitoes getting blocked, and a basic security review being flagged as a cybersecurity risk. "As a scientist, this is perhaps the most useless model I've ever tried," reads one of the sharper reactions.

One detail from the 319-page system card also caught the attention of the user base. Willison flags that Anthropic has built in invisible interventions that deliberately degrade Fable's performance when users try to develop competing frontier models, things like pretraining pipelines or ML accelerator design. Unlike the cyber or bio filters, there's no visible fallback here.

Instead, Anthropic quietly manipulates responses through prompt modification or steering vectors. Anthropic says only about 0.03 percent of traffic is affected. Willison isn't thrilled about a model that secretly distorts its answers on "ML accelerator design" to slow down research that might compete with Anthropic's own goals. In forums, the move is widely seen as openly anticompetitive: "Anthropic's definition of 'unsafe' encompasses 'competing with Anthropic.'"

The economics don't add up for everyone

There's also the question of whether Fable's performance justifies its much higher price once you move past the subsidized flat-rate plans. Companies pay separately for the Harness, and tokens are billed at standard API rates. One developer reports that switching from the $200 Max flat rate to enterprise billing pushed his Opus costs from $200 to $10,000 a month at the same usage level. With Fable, it would be around $20,000.

For many, that creates a straightforward comparison: for $10,000 to $20,000 a month, you can hire one or two experienced developers, depending on location, and give them an AI license on top. Outside the US, the gap is even starker. One developer earning about $2,500 a month puts it bluntly: "nobody I know would pay that, no company will justify spending $20k/month when they can hire 10 more developers instead."

Supporters argue the value can't be measured in tokens alone. Solving hard problems in days instead of weeks can deliver real returns, if Fable reliably matches the output of two or three developers. Skeptics counter that the real difficulty is rarely solving the problem. It's defining scope and acceptance criteria: "the challenge in software development is not to solve a problem, but to define the outcome, the scope, the acceptance criteria etc." AI-generated solutions also often need to be thrown out and rebuilt, eating into any efficiency gains.

The debate shows how rising model costs are reshaping real-world adoption, a trend we covered in detail in our recent AI Radar #3.

A hybrid approach is taking shape in these discussions. Many see cheaper models like DeepSeek v4 as good enough for daily work, while reserving Fable or Opus for complex features, tough bugs, or high-level planning. One user describes a workflow where they draft the initial plan with Opus 4.8, have Fable check it for ways to simplify, and only then move into a cheaper execution loop.

Rate limits drain fast, and data retention is a dealbreaker

Fable 5 has a one-million-token context window and a 128,000-token output cap. Its training data goes up to January 2026. It's available for a limited time on Pro, Max, Team, and Enterprise plans, where it counts as double the Opus usage through June 22. Limits run out fast. One user reports a single session burned through the entire five-hour window on the $200 Max plan without finishing the task. Another says a single prompt blew past the Pro plan limit.

Starting June 23, the model will require credits until Anthropic restores subscription access once capacity allows. In practice, that makes the model significantly more expensive.

There's also a new 30-day data retention policy for Mythos-class models. It applies even when users have a zero-data-retention agreement in place. For companies that relied on that feature, it's an immediate dealbreaker.

This leaves us with an early verdict: Fable 5 seems a genuine leap, especially for complex, delegatable software work in the hands of experienced users. Anyone doing small, quick tasks or working in regulated fields will likely be better off with cheaper or less filtered models. Fable 5 also marks a shift in how Anthropic secures, prices, and quietly steers its most powerful models.

AI News Without the Hype – Curated by Humans