Anthropic PBC today debuted Claude Fable 5.1 and Claude Mythos 5.1, its most capable large language models to date.
The launch comes a day after the company inked a $35 billion infrastructure deal with cloud startup Lambda Inc. A week earlier, Anthropic signed an even larger hardware contract with Nscale Global Holdings Ltd.
Record-setting benchmark results
Fable 5.1 and Mythos 5.1 are effectively the same model. The main difference is that the Fable 5.1 can’t be used for sensitive cybersecurity and biology tasks. Mythos 5.1, which doesn’t block such use cases, will be accessible to only a limited number of trusted organizations.
Anthropic says that Fable 5.1 has set records across several AI benchmarks. It completed Terminal-Bench-Science 0.1, a test that measures LLMs’ ability to complete research tasks, with a score of 52.6%. That’s more than double what the previous-generation Fable 5 model achieved. Fable 5.1 also scored 13.% higher on the Terminal-Bench 4.0 coding benchmark and aced several non-technical task collections.
Another feature that sets the model apart from its predecessor is that it has more relaxed guardrails. As a result, researchers can use it to discover software vulnerabilities. However, support for more advanced tasks such as finding exploit paths is only available in Mythos 5.1. Anthropic says the latter LLM possesses the “strongest cyber capabilities of any model we’ve released.”
Mythos 5.1 is accessible through two specialized access programs for cybersecurity researchers and biologists. Fable 5.1, meanwhile, is generally available. Anthropic says that an improved prompt caching mechanism enables Fable 5.1 to run “typical workloads” 25% more cost-efficiently than its predecessor. Applications that lean heavily on AI agents are expected to see savings of up to 45%.
The cost cuts indicate that Anthropic has found a way to reduce Fable 5.1’s hardware usage. Nevertheless, it appears that the company expects its infrastructure requirements to grow significantly over time.
New cloud deal
On late Monday, sources told the Wall Street Journal that Anthropic had leased $35 billion worth of infrastructure from Lambda. They didn’t specify the amount of capacity the company is buying, but it’s possible to extrapolate. Nvidia Corp. Chief Executive Officer Jensen Huang recently stated that building 1 gigawatt of AI infrastructure will soon cost $80 billion to $100 billion. That suggests Anthropic is buying somewhere between 350 and 437 megawatts of capacity.
Lambda, the startup with which Anthropic inked the deal, operates an AI-optimized public cloud. The platform enables customers to spin up Nvidia clusters with up to 165,000 graphics cards. Users can customize what firewall a cluster uses, how encryption keys are managed and other configuration details.
Lambda powers its infrastructure using customized versions of popular open-source cluster management platforms. The tooling lineup includes not only Kubernetes and Slurm but also dstack and SkyPilot, two lesser-known alternatives. The latter technologies offer fewer customization options but are simpler to use in certain respects.
The infrastructure that Anthropic is buying from Lambda will reportedly be housed in a data center currently under construction in Nueces County, Texas. The facility is being built by Hut 8 Corp., a Nasdaq-listed AI infrastructure company. Another wrinkle is that Nvidia will hold the lease on the campus. That arrangement may be designed to reduce financial risks for Lambda and Anthropic.
Hut 8’s Nueces County campus will have 1 gigawatt of capacity. In late July, the company disclosed that an unnamed “high-investment-grade” customer had agreed to lease 704 megawatts. A few days later, the Financial Times reported that the customer is Nvidia.
Hut 8 says that the facility is based on the chipmaker’s DSX data center reference architecture. DSX includes not only blueprints but also a simulator that companies can use to test their cluster layouts. Additionally, Nvidia-developed power management tools help data center operators optimize their energy use.
Hut 8 expects to connect its Nueces County campus to the grid in the first quarter of 2027. The first data hall is scheduled to come online about six months later.
The deal comes less than a week after Anthropic inked an even bigger infrastructure contract with Nscale Global Holdings Ltd., a U.K.-based Lambda rival. The AI provider will reportedly pay $45 billion to rent about 460 megawatts of computing capacity in a West Virginia data center. The facility will be equipped with Nvidia Corp.’s latest Rubin graphics cards and Vera central processing units.
