Opus or Fable is the practical question behind Anthropic’s newest flagship model. On 9 November 2026 the company released Claude Fable 5.2, together with its sibling model Claude Mythos 5.2. In the vendor’s own tests, Fable ranks ahead of Claude Opus 5.5, Anthropic’s strongest everyday model, in almost every category. Even so, Anthropic advises users to stay with Opus for most work, because Fable drains the usage limit faster. That tension is the subject of a roughly six-minute video published on 12 November by Sebastian Claes on his channel “Sebastian Claes | Claude Code & KI-Agenten”. He does not run his own bake-off; he walks through the claims and examples in the vendor announcement.
How Opus 5.5 and Fable 5.2 differ
Both models belong to Anthropic’s Claude family. According to the documentation, each has a one-million-token context window and up to 128,000 output tokens. The difference is the role. Opus 5.5 is what Anthropic calls its “strongest daily driver”: the starting point for most coding and knowledge work. Fable 5.2 is meant for especially large, open-ended technical jobs – long agent runs, difficult migrations, work where the path is not known in advance.
According to Anthropic, Mythos 5.2 is the same underlying model as Fable, but with weaker cybersecurity and biology safeguards. That is why it is available only to vetted organisations in IT security and biomedical research, through the Life Sciences and Cyber Verification Programs. For ordinary use, Fable 5.2 is the generally available variant. Compared with Fable 5.1, Anthropic promises outputs up to 30 percent faster and an API price that is one-fifth lower: $8 per million input tokens and $40 per million output tokens, down from $10 and $50. Opus 5.5 costs half as much on the API, at $4 and $20. For subscribers, the weekly allowance usually matters more than the token price.
We already reviewed the launch and the vendor benchmarks in a closer look at Claude Fable 5.2. The earlier price cut for Opus 5.5 was a story in its own right.
What the vendor scores actually show
Anthropic writes that Fable 5.2 is “a step above Opus 5.5 in almost all the capabilities we evaluated.” Those figures are vendor claims, not independent measurements. In the comparison table in the announcement, Fable 5.2 leads among other tests on FrontierCode 2.0 (49.1 percent at the highest effort versus 45.5 percent for Opus 5.5), FrontierSWE v2 (69.9 versus 62.3 percent) and the Remote Labor Index (41.3 versus 21.3 percent). On AutomationBench, a business-workflow test, Gemini 4 Carbon is ahead with 53.9 percent, against Fable’s 51.4 percent.
The exception Claes flags in the video is the knowledge-work test AA-Briefcase v1.1, which has two scores. On the rubric, Fable 5.2 leads with 66.1 percent against 61.2 percent for Opus 5.5. On Elo, Opus 5.5 is slightly ahead, at 1,800 versus 1,785. That is the “one of two ratings” in which, as the video notes, Opus even edges out Fable. Independent league tables often tell a different story anyway, as the recent GDPx benchmark from SemiAnalysis showed.
The examples Claes picks up from the announcement
The concrete cases in the video come from Anthropic’s announcement, not from a hands-on test by the channel. One Anthropic engineer is said to have given Fable 5.2 only an empty folder. The model then spent eight days around the clock coordinating more than 1,000 subagents and wrote a prototype operating system that boots on a laptop – with its own kernel, drivers, file storage, windows, a login screen, an installer and built-in Claude functionality. There is no independent way to verify that from the published material.
In another experiment, Anthropic says Fable 5.2 designed three working electronic devices from a one-paragraph description: a home voice-assistant hub, a USB speakerphone and a handheld virtual pet, each with a circuit board, a 3D-printed enclosure and software. The boards were built with the manufacturer Diode; all three devices are said to have worked without electrical modifications. In the video, Claes highlights the hands-free device with a USB port.
Closer to everyday office work are two tester quotes from the announcement. One tester had Fable pull figures from a company’s 10-K filing into a forecasting model. The template did not say whether a metric should be copied or calculated. Fable inferred the intent from the cell formatting, noticed that the regional figures did not add up to the reported total, and got every cell right. Another tester had it review documents for a loan application; the model flagged that the salary deposits did not match the account balance on the statement. These are individual tester reports, not a controlled study.
As an example of a task that may justify Fable, Claes points to migrating a custom-built program to a new technology. Anthropic says an earlier version of Fable 5.2 moved a critical internal production service from Python to Rust; the service then started about 26 times faster. That, too, is a vendor claim.
Whether your plan already includes Fable
Whether Fable runs inside a subscription at no extra cost depends on the plan. According to Anthropic’s help pages, Fable models are included in the weekly limit on Max, Ultra and premium seats on the Team plan, but only up to half of that limit, and they use it faster than other Claude models. On the Pro plan and on standard Team seats, Fable models normally run on separately purchased usage credits.
At launch there is a time-limited exception, which Claes dates to 16 November. The official offers page is more precise: the window ends on 16 November 2026 at 11:59 p.m. Pacific Time, which is 17 November 2026 at 8:59 a.m. CET. Until then, Pro plans and standard Team seats can use all Fable models for up to half of the weekly limit at no extra cost. Eligible accounts also get one Fable limit reset, which expires at the same time. The default model does not change; Fable has to be selected in the model picker. After the offer ends, Fable on Pro goes back to usage credits.
Opus or Fable: the three questions from the video
Claes gives three reasons, consistent with Anthropic’s documentation, why the company still points everyday users to Opus. First, Opus is meant for daily work, Fable for especially large, open-ended tasks. Second, according to the API documentation, Fable 5.2 always uses adaptive thinking; it cannot be turned off. Claes’s conclusion is that even a request for a direct, quick answer still costs waiting time and limit usage that the question may not need. Third, Fable blocks most IT-security tasks; Anthropic says those queries are then answered by the older Opus 4.8 (and biology queries by Opus 5).
From that, Claes derives three checks before switching to Fable:
- Is the task large and open-ended, with many steps whose path you do not fully know in advance?
- Has Opus already failed at this exact task? Anthropic recommends the same order: start with Opus, and move to Fable only if the result is not good enough.
- Does it involve IT security? If so, Fable will usually block the request anyway.
For a client quote, a meeting summary or a quick analysis, Opus therefore remains the better choice. Fable is more plausible if an old custom system needs to be migrated, or if a stack of documents has to be checked for inconsistencies. Claes’s tip is to give Fable, before 16 November, exactly one real task that Opus has already failed at, and to watch consumption under Settings > Usage.
Conclusion
On the vendor’s own numbers, Fable 5.2 is stronger than Opus 5.5 in most of the disciplines Anthropic measured – and it is still not a new everyday model. Anthropic says so itself, and Claes draws the same practical line in his video: start with Opus, and reach for Fable only on the few large jobs where Opus gets stuck. The examples, from the operating system to the loan review, come from the announcement and have not been independently confirmed. Pro users still have a no-extra-cost trial window, up to half of the weekly limit, until 17 November at 8:59 a.m. CET. After that, the rule is the same as before: use Fable deliberately, because the limit shrinks faster.
