Anthropic, OpenAI release cheaper AI even as safety fears grow

Anthropic and OpenAI on Tuesday released new, cheaper artificial intelligence models within hours of each other, days after the heads of both companies called for a slowdown in AI development.

Anthropic said its Opus 5.5 comes close to the performance of its flagship model at a significantly lower price. It matches Fable, Anthropic's cutting-edge model, on most tasks at a 20 percent lower price, the company said.

OpenAI followed shortly after with GPT-6 Sol and GPT-6 Luna, and claimed they handle tasks substantially better than Anthropic's top models.

The new models are cheaper, faster versions of GPT-6 Astra, OpenAI's most powerful model, launched earlier this month.

The dueling launches highlight the tension facing the two San Francisco companies as both warn of the dangers of the technology they are building.

The companies are, however, under pressure to generate returns on massive investments amid stiff competition from low-cost models, particularly Chinese ones.

Anthropic's release comes a few weeks before its expected stock market debut. OpenAI has delayed its IPO plans until next year.

Concerns about AI have grown since July, when OpenAI models undergoing a cybersecurity test got around controls meant to isolate them from the internet and broke into the servers of Hugging Face, an AI model repository.

Anthropic reported similar incidents, and the two companies briefly paused work on their new AI systems.

Earlier this month, Jacob Coxon, a British researcher, announced his resignation from Anthropic in a viral post on X, saying neither Anthropic nor OpenAI was acting responsibly in the aftermath of the Hugging Face incident.

In the wake of the Coxon's warnings, Anthropic CEO Dario Amodei called for slowing down AI improvement. He was joined by OpenAI CEO Sam Altman and Elon Musk, but mocked by Trump, who described such warnings as "hoaxes."

Opus 5.5 is the first Anthropic model released since then.

Anthropic said the new model applies the same safeguards as its state-of-the-art model Fable in the areas of cybersecurity, biology and AI model design.

In practice, a request deemed risky by these safeguards is rerouted to an older, less powerful model.

Anthropic also said that, based on its own testing, Opus 5.5 is the best-performing model it has tested to date on alignment -- that is, how closely it adheres to what its designers intend.

But the model also showed signs that it often suspects it is being evaluated, which the company acknowledged makes it harder to predict how it will behave in real-life use.

bl-arp/msp

Accenture