Claude Fable 5 is a revolutionary model like Mythos, but safer. This means you can work with it too.

The race for the most powerful artificial intelligence has accelerated over the past year, with companies vying to offer a more capable model. In recent months, Claude has gained increasing recognition, even though Anthropic, the company behind it, plays the role of the more cautious one. It kept its strongest Mythos model under wraps since April, releasing it only to a handful of vetted organizations, fearing what its capabilities could achieve in the wrong hands. Now, the company has reportedly found a way to offer the same power to ordinary users. It’s called Fable 5 and can be described simply: it’s Mythos with a muzzle

Mythos with a muzzle: what Fable 5 is

Mythos class models rank above Anthropic’s current top-tier Opus series. The first of these, Mythos Preview, was launched by the company in April as part of the Project Glasswing program, exclusively for a narrow circle of cybersecurity specialists and critical infrastructure operators. Last week, access was expanded to approximately 150 organizations from over fifteen countries. However, the model has so far avoided the public – it possesses unprecedented capabilities in hacking and vulnerability detection, which could cause significant damage in the hands of attackers.

Fable 5 and Mythos 5 are the same model under the hood. They differ in only one, but crucial, aspect: Fable has its safety mechanisms enabled, Mythos does not. Anthropic explains this with the choice of names, where the Latin fabula and Greek mythos essentially mean the same thing, “story” or “narrative”. According to the manufacturer, Fable 5 is the most powerful model it has ever made available to the general public. It holds the best results across most tests, and Anthropic states that the longer and more complex the task the model receives, the greater its lead over the competition.

Work that would take months, it handles in days

According to available figures, Fable 5 is best suited for programming. Payment company Stripe reports that the model managed to rewrite a vast Ruby codebase of 50 million lines in a single day, whereas manual work would have taken an entire team over two months. In Anthropic’s own benchmarks, Fable 5 achieved 80.3% in the SWE-Bench Pro test and 88% in Terminal-Bench 2.1, leaving Opus 4.8 and competing GPT 5.5 and Gemini 3.1 Pro behind. Furthermore, it is more token-efficient than previous models, thus consuming less for the same work.

Anthropic also reports decent results outside of code. In the office work test GDPval-AA, Fable 5 achieved a score of 1932 compared to 1890 for Opus 4.8. Analytics firm Hex claims it was the first model to break the 90% barrier in its demanding test of long analytical tasks. Image processing is an interesting chapter, where Fable 5, according to the manufacturer, becomes the new top performer. It can extract precise numbers from scientific graphs or reconstruct the source code of a web application solely from screenshots. As proof, Anthropic also published that the model completed the game Pokémon FireRed based only on raw game screenshots, without maps and navigation aids that earlier versions necessarily required.

The model also shows progress in memory. In the board game Slay the Spire, Fable 5 improved its performance three times more significantly than Opus 4.8, thanks to access to notes stored in files. This suggests that the model can maintain attention and context even for tasks spanning millions of tokens.

Safeguards that turn a powerful model into an obedient servant

Here comes the part for which Anthropic held Mythos back for so long. Fable 5 received a set of independent monitoring systems, so-called classifiers, which observe whether someone is trying to misuse the model. They cover three areas: cybersecurity, biology with chemistry, and so-called distillation, meaning attempts to extract the model’s capabilities and train a competing system on them. As soon as a classifier detects such a query, the weaker Opus 4.8 handles the response for Fable 5, and the user is informed about it.

Claude Fable 5 is a revolutionary model like Mythos, but safer. This means you can work with it too. Adam Kurfürst News Adam Kurfürst News

Anthropic admits that it intentionally set the safeguards strictly, so occasionally, they might block even a harmless query. However, according to the company’s own data, the switch to Opus occurs in less than five percent of conversations, meaning that in over 95% of cases, the user is fully operating on Fable 5. The company also tested the robustness of its defense with an external premium competition for hackers, where participants, over more than a thousand hours, did not find a universal way to bypass the safeguards. It should be added, however, that the UK agency UK AISI, according to Anthropic itself, came close to one such breakthrough during brief testing.

From a user’s perspective, there’s one catch. If you encounter a sensitive topic, you won’t get the model you’re paying for. This is also confirmed by a note below the benchmark table: in cybersecurity and biology tests, Fable 5’s results are closer to Opus’s weaker performance due to the switching. For a programmer addressing the security of their own application or a biologist working on legitimate research, this can be an annoying limitation in practice.

Mythos 5: the same model, but without a muzzle

While everyone received Fable 5, Anthropic reserved the full power of Mythos only for the chosen few. Concurrently, it also launched Mythos 5, which has its cybersecurity safeguards removed. It is provided to partners from the Project Glasswing program in cooperation with the US government, and according to the manufacturer, it is the model with the strongest cybersecurity capabilities in the world. A program for selected biologists is also expected to be added soon, for whom Anthropic will lift restrictions in the areas of biology and chemistry.

It is precisely with Mythos 5 that one can see why the company hesitated so much. In drug design, it reportedly accelerated part of the process for internal experts by about tenfold and found promising candidates for nine out of fourteen tested protein targets. In molecular biology, according to Anthropic, scientists prefer its hypothesis proposals in about 80% of cases, and one such idea regarding E. coli bacteria was independently confirmed by another laboratory. In genomics, the model worked almost autonomously for a week and trained its own model, which surpassed a recently published work in Science magazine, despite being a hundred times smaller. However, these same capabilities that advance science could also aid in the design of dangerous viruses in the wrong hands.

Price, deadlines, and the catch with your data

Both Fable 5 and Mythos 5 cost 10 dollars per million input tokens and 50 dollars per million output tokens. Anthropic emphasizes that this is less than half the price of the earlier Mythos Preview. From a regular user’s perspective, however, it’s more significant that it’s double the price of Opus 4.8. For demanding and highly autonomous tasks, this might be worthwhile, as noted by Rakuten, which states that the model, with maximum effort, checks and verifies its own work. For ordinary use, however, this premium might not make sense.

Availability is phased. Fable 5 is immediately available via API and for consumption-based enterprise plans. For subscriptions, it’s more complex. From launch until June 22, Pro, Max, Team, and selected Enterprise plans include it for free. From June 23, Anthropic will remove it from these plans, and further use will require credits, with the promise that it will eventually be reinstated as a standard part of the subscription. The deadlines are worth watching, as the company itself admits it’s difficult to estimate demand.

The change in data handling deserves the most attention. For Mythos class models, Anthropic is introducing mandatory 30-day retention of all traffic, even for companies that previously had contractual agreements for zero data retention. The company promises not to use the data for training or anything other than security, logs access, and deletes it after a month. However, TechCrunch points out that this could set a precedent where access to the most powerful models is permanently linked to mandatory data retention under the guise of security.

Fable 5 is thus the most powerful publicly available model you can run today, and it makes sense for developers, analysts, or researchers with extensive and autonomous tasks. For everyone else, there are several “buts”: you won’t get full power on sensitive topics, you pay double compared to Opus 4.8, subscription access is temporary for now, and 30-day data retention must be accepted as a condition. Those who don’t need extreme performance for long tasks will likely not go wrong with the cheaper Opus 4.8.

Would you try Fable 5, or is the cheaper Opus 4.8 enough for you?

Sources: Anthropic, TechCrunch