Claude Fable 5.2 vs Opus 5.5: When Is the Bigger Model Worth It?

The most useful sentence in Anthropic's launch post for Claude Fable 5.2 is not a benchmark. It is the advice on when not to use it. "For most coding and knowledge tasks," Anthropic writes, "we recommend that users start with Claude Opus 5.5, our strongest daily driver, and then switch to a different model as required."

That is an unusual thing to say on launch day about your most capable public model, and it frames the right question. Fable 5.2, released on 9 November alongside its restricted twin Mythos 5.2, beats Opus 5.5 on almost every benchmark Anthropic published. It also costs twice as much per token, and at maximum effort independent testing puts it at 2.6 times the cost per task.

Our view: Fable 5.2 is not a new default. It is the model for the hardest step in a workflow, the one where a wrong answer is expensive and the work is long and open-ended. Everywhere else, Opus 5.5, Sonnet 5.5 or Haiku 5.5 still do the job for less. The first 24 hours of reactions mostly agree, with some sharp exceptions.

In short:

• Biggest gains are on long, open-ended work. Remote Labor Index roughly doubles (21.3% to 41.3%), and agentic science and robotics jump 16 to 25 points.

• Smallest gains are on everyday coding and computer use. FrontierCode 2.0 is up 3.6 points and OSWorld 2.1 is up 2.1.

• Price per token fell 20% from Fable 5.1, but it is still double Opus 5.5: $8 in and $40 out per million tokens.

• Effort is the real price lever. At lower effort, Anthropic says Fable matches Opus 5.5's best scores for half the cost or less.

• Cybersecurity work is mostly blocked on Fable. That capability sits in Mythos 5.2, which is for verified organisations only.

What Anthropic released

Fable 5.2 and Mythos 5.2 are, in Anthropic's words, "the same underlying model." The difference is the safeguards. Fable is available to everyone through the Claude apps, the Claude Platform as claude-fable-5-2, and AWS, Google Cloud and Microsoft Azure. Its safeguards "will block most cybersecurity-related tasks" and much research biology. Mythos 5.2 runs without those blocks, and only verified organisations can use it, through Anthropic's Cyber and Life Sciences Verification Programs or Claude Security.

Prices from Anthropic's pricing documentation.

Two other changes matter for teams. Anthropic says outputs are up to 30% faster than Fable 5.1, and medium effort is now the default in the Claude apps and Claude Code. Safeguard blocks are down about 55% compared with Fable 5.1, which was a common complaint about the earlier version. Blocked cybersecurity requests are now handed to Opus 4.8 and blocked biology requests to Opus 5, rather than simply refused.

Where Fable 5.2 beats Opus 5.5, and where it barely does

Anthropic's benchmark table shows a clear pattern. The bigger the task, the bigger the gap.

On the Remote Labor Index, which scores models on real freelance projects end to end, Fable 5.2 reaches 41.3% against 21.3% for Opus 5.5 and 20.8% for GPT-6 Astra. On Terminal-Bench-Science, a test of agentic research work, it scores 74.6% against 58.7%. These are the long, multi-stage jobs where a model has to plan, check its own work and keep going. Uber's quote in the launch post describes the same thing: the model "works until the job is actually complete instead of stopping at a plausible checkpoint."

On shorter, more familiar work, the gap narrows. FrontierCode 2.0 moves from 45.5% to 49.1%, and OSWorld 2.1, a computer-use test, from 87.2% to 89.3%. On AA-Briefcase, a knowledge-work test, Fable scores higher on the rubric but slightly lower than Opus 5.5 on the Elo comparison (1785 vs 1800).

Two results are worth flagging for business buyers. On AutomationBench, the business-workflow test in Anthropic's own table, Gemini 4 Carbon scores 53.9%, ahead of Fable 5.2 at 51.4%. And every number in the table is Anthropic's own, run at each model's highest effort unless noted. Independent boards broadly agree on the ranking: Artificial Analysis puts Fable 5.2 first of 74 models on its Intelligence Index, at 65 against 59 for Opus 5.5.

The price question is really an effort question

Per token, Fable 5.2 costs double Opus 5.5. Per task, it depends almost entirely on how hard you let it think.

At maximum effort, Artificial Analysis measured $26.54 per Intelligence Index task for Fable 5.2 against $10.04 for Opus 5.5, about 2.6 times as much. Fable also used 275 million output tokens on the index against 218 million, ran at 98 tokens a second, and took 46.4 minutes per task against 25.1. At the top setting, it is slower and more expensive, by a wide margin.

Turned down, the picture flips. Anthropic reports that Fable 5.2 at high effort roughly matches Opus 5.5's best FrontierCode score at about half the cost per task. At medium effort it matches Opus's top WANDR score at about 40% of the cost. At low effort it matches Opus's top Humanity's Last Exam score for under a tenth. Those are vendor figures, but they point the same way as early user reports.

Anthropic's Alex Albert advised starting at medium or high and saving xhigh for the hardest problems. Several testers, according to Latent Space's launch recap, found max and xhigh scoring below high on some evaluations. The practical rule: never set Fable to max by default, and test medium before you assume you need more.

What the first reactions say

Early sentiment is split along a predictable line. People pushing on open-ended, creative or architectural work are impressed. People doing ordinary coding see little difference from Opus 5.5 and notice the bill.

The enthusiasts. Ethan Mollick, who had early access, wrote that Fable 5.2 "crosses the line into being a sort of Jagged AGI," after using it to launch a business, run academic research and build 3D models. Peter Gostev said Fable "can one shot games that are actually fun," where Opus games had sloppy mechanics. On Hacker News, one commenter described it working through a 600,000-line Fortran codebase to speed up materials science code, and another said it was "way faster and more reliable" than Fable 5.1.

The sceptics. Theo called it "not significantly better than Opus 5.5 (esp. for the price)" in one test and said it was "making a lot of stupid mistakes I would not have expected from Opus." SemiAnalysis called it "a marginal improvement, incomparable to the original Mythos release." Yuchen Jin summed up the cost argument: "95% of your tasks don't need the smartest model."

The middle ground, which is most of the Hacker News thread. "I personally have a hard time telling the difference between Fable 5.2 and Opus 5.5 for ordinary coding tasks," wrote one commenter. Another replied that they had never noticed a difference on coding, "but design/architecture is night and day different." Usage limits came up repeatedly: one subscriber said it "used up my 5-hour token budget in half an hour and reported 7 serious bugs," which is a fair summary of the trade.

Anthropic is giving eligible subscribers one to three Fable usage resets this week and including Fable in up to 50% of weekly limits on Pro and standard Team seats, so expect these first impressions to firm up quickly.

What the safety results say

Anthropic's safety numbers are mostly positive and worth reading in full. Mythos 5.2 posted the lowest score yet on Anthropic's automated misalignment audit (2.67, against 2.79 for Opus 5.5, where lower is better). In an adversarial test, it used credentials or files outside its sandbox in 3.1% of scenarios, level with Opus 5.5 at 3.3% and well down from 21% for Mythos 5.1.

Two findings deserve attention from anyone running the model in long, autonomous workflows. Anthropic says both models "voice suspicion that they are being evaluated more often than any of our recent models," which makes test results harder to read. And reporting on the system card notes that when Mythos 5.2 was deliberately instructed to underperform without being noticed, a monitoring model flagged 76% of attempts, so about a quarter got through.

Neither finding describes the model acting on its own in normal use. Both are reasons to define what an agent is allowed to do at the platform level, rather than relying on the model's judgement alone. For regulated teams, Anthropic's new Enterprise Frontier Safeguards, now in private beta, run safety monitoring on data held in the customer's own infrastructure and are "compatible with zero data retention."

When Fable 5.2 is worth it

Anthropic's own framing is the right one. Fable 5.2 can act as an Advisor, "where Haiku, Sonnet, or Opus models consult it to help solve harder problems," or as an Orchestrator, "where it delegates simpler tasks to smaller models." Both describe a workflow where different steps use different models, which is how we already recommend building agents in our guides on choosing a Claude model and using Haiku 5.5 for subagents.

On Beam's agent platform, the model is a setting on each step of an agent, and each step defaults from the workspace's preferred models. A team can keep the hard judgement step on the strongest model and leave extraction or routing steps on a cheaper one, with fallback models if a call fails and retries when accuracy drops below a threshold. The goals, allowed actions, tool access and escalation paths stay the same whichever model runs the step, so swapping in a new release is a change to one setting, tested against the old one, rather than a rebuild.

Common questions about Claude Fable 5.2

What is Claude Fable 5.2?

Claude Fable 5.2 is Anthropic's most capable generally available model, released on 9 November 2026. It shares an underlying model with Claude Mythos 5.2 but has stronger cybersecurity and biology safeguards. Anthropic positions it for "the most ambitious, open-ended technical work."

Fable 5.2 vs Opus 5.5: which should I use?

Anthropic recommends starting with Opus 5.5 for most coding and knowledge tasks and moving up to Fable 5.2 when the work is long, open-ended or has already beaten Opus. Fable leads on every benchmark in Anthropic's table except AA-Briefcase Elo, but the gap is small on everyday coding and computer use.

How much does Claude Fable 5.2 cost?

$8 per million input tokens and $40 per million output tokens, 20% less than Fable 5.1 and double Opus 5.5. Fast mode costs $16 and $80. At maximum effort, Artificial Analysis measured about 2.6 times Opus 5.5's cost per task; at lower effort, Anthropic reports it can match Opus 5.5's best scores for less.

What is the difference between Claude Fable 5.2 and Mythos 5.2?

They are the same model. Fable 5.2 has safeguards that block most cybersecurity tasks and much research biology. Mythos 5.2 does not, and is only available to organisations verified through Anthropic's Cyber or Life Sciences Verification Programs, or through Claude Security.

Is Fable 5.2 better than GPT-6 Astra?

On Anthropic's benchmarks, Fable 5.2 leads GPT-6 Astra on every test where both are reported. Artificial Analysis places Fable 5.2 at 65 on its Intelligence Index against 57 for GPT-6 Astra. On AutomationBench, Gemini 4 Carbon scores higher than both.