Anthropic made Claude Fable 5 worse at AI development, users call it anticompetitive behaviour
Anthropic's Claude Fable 5 has drawn criticism from users who say it becomes less helpful on AI research tasks. The backlash centres on hidden safeguards that may switch responses without making the restriction visible.
Anthropic has once again created a buzz in the AI industry with the launch of Claude Fable 5, its new Mythos-class AI model. The company says Fable 5 is its most capable model yet, showing exceptional performance in software engineering, knowledge work, vision, scientific research and many other areas. This model comes with Anthropic's newest safety measures: protections against AI distillation, which refers to the process of using one model's outputs to help train another model. But just after its release, some users are talking less about its capabilities and more about what they believe the model is deliberately refusing to do.
Is Anthropic protecting safety or protecting its competitive edge?
The debate started after users began digging into what Anthropic's distillation safeguards could mean in practice. Several users argued that the company is not just trying to prevent misuse, but is also making it harder for competitors to use Claude to build rival AI systems.
One user summarised the concern by claiming that Fable 5 may not provide its full capabilities when it detects work related to building or improving frontier AI models. According to the user, this could include tasks such as designing large-model training pipelines, planning distributed training across massive GPU clusters, optimising model-parallel systems, building infrastructure for large-scale pretraining runs, working on AI chips and accelerators, distilling frontier models, or finding ways to make competing AI models stronger, faster or cheaper.
The user argued that Anthropic's approach is unusual because the model may not explicitly refuse such requests. Instead, it could still appear helpful while quietly becoming less capable in that specific category of work.
"As a paying user, that matters," the user wrote. "The model can still sound helpful while being intentionally less capable in a narrow but important category of work."
That possibility has fueled accusations that Anthropic's safeguards are motivated by competition as much as safety.
Some critics have gone as far as describing the move as anti-competitive behavior.
"I feel like Anthropic's whole shtick is using safety-ism in the service of anticompetitive behavior," one user wrote.
Another user argued that the restrictions could affect a broader range of AI-related work including AI development or AI research-related tasks than Anthropic intends.
"Claude Fable will be deliberately bad at frontier LLM training. By extension it will also likely be bad at LLM inference, given the overlap in workloads. Very sad," the user wrote.
"There is a lot of justified anger at Anthropic for sandbagging Fable 5 for AI development tasks," one X user wrote.
The user argued that the issue goes beyond day-to-day usage. According to them, the restrictions make it difficult for third-party researchers to evaluate the model fairly because they cannot tell whether the model genuinely failed a task or whether Anthropic's safeguards stepped in behind the scenes.
"Fable 5 would be a perfect test candidate. But because of Anthropic's guardrails, we can't know if the model failed or if their classifiers blocked the capability," the user added.
Another user criticised the idea that the model could intentionally perform worse on frontier AI research tasks without users being aware of it.
"Fable 5 will be bad on purpose on AI frontier LLM research tasks. This is very, very sad for the research community," the user wrote.
Why Anthropic is holding back
Anthropic says it has identified large-scale attempts to extract, or "distill," Claude's capabilities. To address this, the company has built special safeguards into Fable 5. The company says it has taken this step because distilling Fable 5's capabilities could contribute to the spread of near-frontier AI systems that may be released without similar safeguards.
According to Anthropic, if its systems detect that a request may be part of an effort to distill Claude's abilities, the model will fall back to Claude Opus 4.8.
The controversial part: Users may never know
What makes the safeguard controversial is that Anthropic says the protections will not be visible to users. As a result, some researchers believe they may not know when the model's behaviour is being influenced by safety systems.
Several users have speculated that instead of refusing requests outright or clearly indicating that restrictions have been triggered, the model may simply provide weaker assistance or modify its responses.
This has led to concerns about transparency, particularly among researchers who rely on AI models for benchmarking and evaluation.
For them, a key question remains unanswered: Did the model actually fail the task, or was a hidden safeguard responsible?
Anthropic says the risks are worth addressing
Anthropic argues that the restrictions are necessary because Fable 5 is an unusually capable model. The company says Fable 5 exceeds the capabilities of all of its generally available models and has shown exceptional performance in software engineering, knowledge work, vision, scientific research and many other areas.
While the model does not pose the same cybersecurity risks that Anthropic previously associated with Mythos-class systems, the company says releasing a model with these capabilities still comes with significant risks.
According to Anthropic, capabilities in areas such as cybersecurity could potentially be misused to cause serious harm if appropriate safeguards are not in place.
Distillation isn't the only thing Fable 5 is restricted from doing
The anti-distillation measures are only one part of Fable 5's safety framework. The company has also added safeguards in several other sensitive areas. Anthropic says that when Fable 5 detects requests related to cybersecurity, biology and chemistry-related requests. The company says that when Fable 5 receives requests in these sensitive areas, the response is automatically handled by Claude Opus 4.8 instead.
According to Anthropic, these restrictions are designed to prevent misuse while still allowing users to benefit from the model's advanced capabilities.
Anthropic has once again created a buzz in the AI industry with the launch of Claude Fable 5, its new Mythos-class AI model. The company says Fable 5 is its most capable model yet, showing exceptional performance in software engineering, knowledge work, vision, scientific research and many other areas. This model comes with Anthropic's newest safety measures: protections against AI distillation, which refers to the process of using one model's outputs to help train another model. But just after its release, some users are talking less about its capabilities and more about what they believe the model is deliberately refusing to do.
Is Anthropic protecting safety or protecting its competitive edge?
The debate started after users began digging into what Anthropic's distillation safeguards could mean in practice. Several users argued that the company is not just trying to prevent misuse, but is also making it harder for competitors to use Claude to build rival AI systems.
One user summarised the concern by claiming that Fable 5 may not provide its full capabilities when it detects work related to building or improving frontier AI models. According to the user, this could include tasks such as designing large-model training pipelines, planning distributed training across massive GPU clusters, optimising model-parallel systems, building infrastructure for large-scale pretraining runs, working on AI chips and accelerators, distilling frontier models, or finding ways to make competing AI models stronger, faster or cheaper.
The user argued that Anthropic's approach is unusual because the model may not explicitly refuse such requests. Instead, it could still appear helpful while quietly becoming less capable in that specific category of work.
"As a paying user, that matters," the user wrote. "The model can still sound helpful while being intentionally less capable in a narrow but important category of work."
That possibility has fueled accusations that Anthropic's safeguards are motivated by competition as much as safety.
Some critics have gone as far as describing the move as anti-competitive behavior.
"I feel like Anthropic's whole shtick is using safety-ism in the service of anticompetitive behavior," one user wrote.
Another user argued that the restrictions could affect a broader range of AI-related work including AI development or AI research-related tasks than Anthropic intends.
"Claude Fable will be deliberately bad at frontier LLM training. By extension it will also likely be bad at LLM inference, given the overlap in workloads. Very sad," the user wrote.
"There is a lot of justified anger at Anthropic for sandbagging Fable 5 for AI development tasks," one X user wrote.
The user argued that the issue goes beyond day-to-day usage. According to them, the restrictions make it difficult for third-party researchers to evaluate the model fairly because they cannot tell whether the model genuinely failed a task or whether Anthropic's safeguards stepped in behind the scenes.
"Fable 5 would be a perfect test candidate. But because of Anthropic's guardrails, we can't know if the model failed or if their classifiers blocked the capability," the user added.
Another user criticised the idea that the model could intentionally perform worse on frontier AI research tasks without users being aware of it.
"Fable 5 will be bad on purpose on AI frontier LLM research tasks. This is very, very sad for the research community," the user wrote.
Why Anthropic is holding back
Anthropic says it has identified large-scale attempts to extract, or "distill," Claude's capabilities. To address this, the company has built special safeguards into Fable 5. The company says it has taken this step because distilling Fable 5's capabilities could contribute to the spread of near-frontier AI systems that may be released without similar safeguards.
According to Anthropic, if its systems detect that a request may be part of an effort to distill Claude's abilities, the model will fall back to Claude Opus 4.8.
The controversial part: Users may never know
What makes the safeguard controversial is that Anthropic says the protections will not be visible to users. As a result, some researchers believe they may not know when the model's behaviour is being influenced by safety systems.
Several users have speculated that instead of refusing requests outright or clearly indicating that restrictions have been triggered, the model may simply provide weaker assistance or modify its responses.
This has led to concerns about transparency, particularly among researchers who rely on AI models for benchmarking and evaluation.
For them, a key question remains unanswered: Did the model actually fail the task, or was a hidden safeguard responsible?
Anthropic says the risks are worth addressing
Anthropic argues that the restrictions are necessary because Fable 5 is an unusually capable model. The company says Fable 5 exceeds the capabilities of all of its generally available models and has shown exceptional performance in software engineering, knowledge work, vision, scientific research and many other areas.
While the model does not pose the same cybersecurity risks that Anthropic previously associated with Mythos-class systems, the company says releasing a model with these capabilities still comes with significant risks.
According to Anthropic, capabilities in areas such as cybersecurity could potentially be misused to cause serious harm if appropriate safeguards are not in place.
Distillation isn't the only thing Fable 5 is restricted from doing
The anti-distillation measures are only one part of Fable 5's safety framework. The company has also added safeguards in several other sensitive areas. Anthropic says that when Fable 5 detects requests related to cybersecurity, biology and chemistry-related requests. The company says that when Fable 5 receives requests in these sensitive areas, the response is automatically handled by Claude Opus 4.8 instead.
According to Anthropic, these restrictions are designed to prevent misuse while still allowing users to benefit from the model's advanced capabilities.
