Anthropic continues developing Claude and reinforces its commitment to safety
The warning that shook the AI industry
In September, Anthropic, the company behind Claude, made headlines when a 27-year-old researcher quit, warning that the technology he helped build could threaten humanity’s survival. Jacob Coxon, who also worked at OpenAI, accused both labs of advancing irresponsibly toward systems capable of self-improvement, known as recursive self-improvement.
CEO advocates slowing down, but launches continue
That same month, CEO and co-founder Dario Amodei argued that development of the most advanced models should slow down so safety measures could keep pace with their capabilities. He even suggested a coordinated pause between countries, though he acknowledged the difficulty of implementing it in the short term.
Despite this appeal, Anthropic launched new versions of its most advanced models, Fable and Mythos, on November 9, with improvements in programming, reasoning, and image interpretation. These launches revive a question left open by the call for a slowdown: how will Claude’s development be affected when the company itself considers it necessary to slow technological progress?
Amodei reiterates commitment during company IPO
In a speech on Thursday, November 19, during the company’s IPO, Amodei reiterated his commitment to control the pace of work and promised to dedicate the necessary time to training and deploying systems to ensure they remain safe and under human control. “We will not give in to commercial or political pressure,” he said, in a free translation.
These statements do not mean the company has abandoned the idea of slowing down. In his September essay, Amodei already distinguished between advocating for a more cautious pace and proposing to halt all training.
It remains unclear how this commitment will change the products, whether through prolonged evaluations, capability restrictions, or delays in bringing models to users.
The next Claude needs to do more than just answer better
The models are already used in tasks requiring information interpretation and technical problem-solving. When announcing Fable 5.2 and Mythos 5.2, Anthropic highlighted advances in programming, reasoning, and image reading, as well as responses up to 30% faster and a 20% reduction in the standard API price compared to version 5.1.
According to the company, Fable and Mythos share the same base but have different configurations. Fable comes to market with classifiers that block high-risk requests in cybersecurity and biology, while Mythos operates without some of these restrictions. The new price is $8 per million input tokens and $40 per million output tokens.
These numbers result from internal tests and the company’s commercial conditions. They do not guarantee that every project will be 30% faster or 20% cheaper, as the outcome depends on the task, usage volume, and model integration method.
For software developers, the value of a more capable system may lie in solving complex, multi-step problems. For document or image analysts, the value may lie in correctly interpreting relationships that previously required multiple corrections.
Anthropic itself uses Claude as a showcase. The company says 90% of the code it produces today is already written by Claude models. In pre-launch tests, the new versions migrated software systems, designed devices, and formalized mathematical proofs.
With this type of use, it is no longer enough to assess whether an answer seems convincing. The more responsibility the user transfers to the system, the more important it becomes to verify that it performs the task correctly and respects established limits.
Science as a motivation to continue advancing
Amodei opened his speech with a personal story that helps explain his confidence in the technology. He and his sister, Daniela, lost their father to a disease that became curable a few years later. Before working with AI, the executive researched biology and experienced the challenges of understanding diseases and developing treatments.
In his view, more capable systems can expand scientists’ work and accelerate discoveries, potentially concentrating a century of scientific and medical progress into a decade, with humans and machines collaborating.
This is a bet, not a guaranteed outcome. The executive himself acknowledged that the results are still preliminary and that experiments and clinical trials will continue to require time. Accelerating hypothesis formulation does not eliminate the need to demonstrate its correctness.
This ambition helps explain why Anthropic intends to continue developing the models despite recognized risks. Amodei’s goal is to expand the company’s contribution to scientific and technical problems, without treating current performance as a limit.
Not every capability needs to reach all users
The company already differentiates access conditions to its tools. While Fable 5.2 was announced for eligible paid Claude plans, API, and partner cloud platforms, direct access to Mythos 5.2 was restricted to verified organizations, such as participants in cybersecurity and life sciences programs.
This separation illustrates a way of managing progress: the decision about a product is not just about launching or abandoning it. It may involve selecting users, limiting permissions, and monitoring activities before expanding distribution.
Amodei’s commitment also distinguishes between continuing research and the immediate release of a capability. A lab may develop a system and still need to assess what protections are required to put it into use.
These restrictions do not prove, in themselves, that the product is safe. To assess its effectiveness, it is necessary to understand what behaviors are being sought to prevent, how they are tested, and what happens when they fail.
Why a more capable Claude requires different testing
In a document published on November 13, Anthropic explains that safety assessments help determine whether a model is ready for release, what protections should accompany it, and what needs to be improved in training.
The company acknowledges that these tests have limitations. Results can vary depending on how questions are formulated, the format of answers, and even the order of alternatives. The claim that one generation is safer than the previous one therefore depends on the quality of the assessments used in the comparison.
The task becomes more complex when the model uses tools and performs actions. In addition to examining the final answer, it is necessary to monitor the process, the resources accessed, and the decisions made to fulfill the request.
Anthropic itself reported problems in this area. In August, it detailed changes after incidents of unauthorized access to real systems during cybersecurity assessments. According to the company, the models operated intentionally without certain safeguards, and an incorrect configuration in a third-party environment allowed internet access.
Because they occurred under these specific conditions, the incidents cannot be treated as typical Claude behavior in everyday use. Still, they show that controlling a product requires observing both the model and the permissions and environment in which it operates.
The challenge is to turn safety into product decisions
One of Anthropic’s initiatives to expand this oversight is the partnership announced in September with Accenture, led by Faculty. The proposal includes model assessments, protection tests, and red teaming, a term for deliberate attempts to find flaws in a system.
The design calls for evaluators within AI companies, with access comparable to that of employees, to monitor training and development decisions. In the announcement, Anthropic acknowledged that there were still open questions about access to information, disclosure of results, and funding for this work.
The relevance of this oversight lies in the possibility of influencing the product before it is released. An assessment must allow for identifying problems and requiring adjustments, rather than just describing the behavior of a system that has already reached the public.
It is at this point that Coxon’s criticism meets Amodei’s promise. The former researcher questions whether labs can control the technology at the speed at which they are advancing. The co-founder says he intends to continue creating more capable models, devoting the necessary time to safety.
For users, the difference should appear in access conditions, available tools, limits of operation, and transparency about failures. The next Claude will be evaluated by what it can do and by evidence that the company knows when and how to restrict it.
Frequently Asked Questions
Is Anthropic going to stop developing Claude?
No. Dario Amodei reaffirmed the intention to develop more powerful models and promised to adjust the pace of training and deployment to the work needed to keep them safe and under human control.
Do all of Anthropic’s models have the same access?
No. The company offers products under different conditions. Direct access to Mythos 5.2, for example, was announced for verified organizations, while Fable 5.2 has wider distribution in eligible plans, API, and partner clouds.
Who is Jacob Coxon?
He is a researcher who worked on pre-training models at OpenAI and Anthropic. At 27, he announced his departure from Anthropic in September 2026 and criticized the pace of the race for more powerful systems.
Is a model with better performance necessarily safer?
No. Capability and safety assessments examine different aspects. A system may improve in programming or reasoning and still require new protections and tests before being released.








