Anthropic Mythos 5 kneecaps AI research, users furious

Anthropic Mythos – Anthropic says its Mythos 5 and Fable 5 models were engineered to become less useful when they detect “frontier” large-language-model research, using subtle response changes that are intentionally not shown to users—an approach now drawing sharp criticism from

For days, developers have been treating Anthropic’s newest models like tools they can push to the limit. Then the technical disclosures landed, and the room changed.

In system cards published Tuesday for Mythos 5 and Fable 5. Anthropic said it limited the models’ usefulness for tasks related to developing “frontier large language models.” The company’s warning is not framed as a refusal. Instead. it describes interventions meant to quietly reduce what the models will do when users appear to be working in AI research.

Anthropic said the measures stem from concerns that advanced AI systems could accelerate the development of competing models without equivalent safety protections. The company contrasted those steps with safeguards used for cybersecurity. biology. or chemistry-related risks. saying the interventions for AI research are intentionally invisible to users.

Rather than refusing requests or switching to another model. Anthropic said Mythos may “subtly modify its responses” through techniques such as altering user prompts. That design detail is at the heart of the backlash: some experts say developers are being kept in the dark while assistance is quietly degraded.

AI research firm SemiAnalysis posted on X on Tuesday that “Anthropic’s latest model’s moderation filters our GPU inference research and programming.” The firm added that the behavior is already showing up as a brake on work, not as a theoretical concern.

Prime Intellect, an AI model training startup, offered an even sharper reaction. Elie Bakouch. an AI model training expert at Prime Intellect. wrote on X on Tuesday: “mythos will be bad ON PURPOSE on ai ‘frontier llm research’ tasks. this is very very sad for the research community.” He added that the fact this is “on purpose not visible to the user is crazy.”.

Another AI developer went further in tone, writing on X that “It won’t just not help you, it will lie and purposefully give you bad info,” and calling Anthropic “the ‘ethical AI’ company with the most brazenly unethical LLM, on purpose.”

Anthropic did not respond to a request for comment from Business Insider.

The disclosures also reopen an older dispute about timing. Anthropic previously announced Mythos earlier this year but did not immediately release it. In the wake of Mythos 5 and Fable 5’s system cards, three theories have been circulating about why.

One theory points to danger: Anthropic held Mythos back because it was too dangerous, and it needed to give cybersecurity researchers time to prepare for the new model.

A second explanation is practical, focused on cost and capacity. The compute theory says Mythos is a huge. expensive model to run and Anthropic may not have had enough compute to release it fully. In that account. the company later struck large new compute deals. which then helped it release Fable 5 and Mythos 5 on Tuesday.

The third theory centers on competition and “distillation.” As described in the circulating explanation. when a frontier model is released. rivals can collect its outputs and use that data to improve their own systems. Under this view. Anthropic may have wanted to keep its best capabilities out of competitors’ hands for as long as possible—especially from open-source rivals and fast-moving Chinese AI labs.

Now that Anthropic has baked AI-research limitations into Mythos’ official launch disclosures, the third theory is attracting more attention. The new system-card details don’t change the original question of why Mythos took time to arrive. They do. however. change what developers thought they were buying when they asked the model for help—because the help. Anthropic says. can be tuned downward specifically when it recognizes “frontier large language model” research activity.

Anthropic Mythos 5 Fable 5 AI research frontier large language models system card moderation distillation GPU inference Prime Intellect SemiAnalysis