Anthropic releases Claude Fable 5.2 and Mythos 5.2

On 9 November Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. They're the same underlying model. Fable has stronger cyber and biology safeguards and is open to everyone. Mythos goes only to vetted organisations, through Anthropic's Cyber and Life Sciences Verification Programs and Claude Security in Claude Enterprise. Pricing is $8 per million input tokens and $40 per million output tokens. A fast mode, up to 2.5 times quicker, costs $16 and $80. Anthropic says output is up to 30% faster than the 5.1 models.

Anthropic's own benchmarks put Fable 5.2 ahead of Opus 5.5 on FrontierSWE v2 (69.9% against 62.3%) and Terminal-Bench-Science (74.6% against 58.7%). A team of about 1,000 Fable 5.2 agents formalised Fermat's Last Theorem in Lean in roughly 30 hours. Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 on Anthropic's internal CVE Rediscovery evaluation. Anthropic still recommends Opus 5.5 as the everyday model for most tasks.

For builders: The API ID is claude-fable-5-2, and it's on AWS, Google Cloud and Azure. Use it for long, open-ended engineering jobs where the higher token price pays off, and keep Opus 5.5 for routine work.

Anthropic starts its IPO roadshow

On 10 November Anthropic launched the roadshow for its initial public offering. It had filed its S-1 with the SEC on 26 October. It's offering 825 million Class A shares at $70 to $78 each. About 801 million are new shares and the rest come from existing holders. Underwriters have a 30-day option on 123.75 million more. The stock is set to list on the Nasdaq Global Select Market under the ticker "ANT". Across the range, the new shares would raise about $56 billion to $62.5 billion before fees.

The same day, Anthropic put Claude Meetings into beta on Team and Enterprise plans. Claude joins Google Meet calls as a visible participant, drafts plans and answers questions in the chat during the call. It leaves automatically if an external guest joins, unless an admin has allowed it.

For builders: Once it's listed, Anthropic will publish regular financial reports, which gives teams that depend on its API better insight into the company's costs and capacity.

Nvidia's board investigates Jensen Huang

On Sunday 8 November Nvidia's board opened an inquiry into reports that CEO Jensen Huang was the target of a foreign intelligence effort. This followed a 29 October New York Times report about his relationship with a Chinese chip buyer. The board said it had "no reason to believe" that national security or Nvidia's secrets had been compromised. In a letter to staff, Huang admitted being unfaithful to his wife, apologised and said he'd cooperate with the review. He stays on as CEO while it runs. Nvidia's third-quarter results are due on 17 November.

For builders: GPU supply and CUDA don't change today. Teams with large Nvidia commitments should watch the 17 November results and any leadership news.

Fireworks AI widens its breach disclosure

Fireworks AI updated its October security bulletin on 9 November. A third-party cloud provider had a security incident on 23 October, and on 24 October an attacker used it to reach Fireworks development machines. The exposed data includes billing addresses, invoice and payment history, and credit usage. For a limited set of accounts it also includes secrets, datasets and incomplete model weights. Fireworks says it has detected no impact on Inference, the Training SDK or Nexus, which keep no customer data by default. It links the attacker to the "LLMJacking" activity that Google Threat Intelligence Group and Mandiant described on 2 November.

For builders: If you store keys or datasets with any hosted inference provider, rotate those secrets and keep only what you need on their platform.

Google adds TPU 8tx and TPU 8i Max

Google added two systems to its eighth-generation TPU line on 9 November. TPU 8i Max puts two TPU 8i chips in one package, doubling compute, SRAM and HBM bandwidth, and it's aimed at very long context windows. TPU 8tx pairs TPU 8t silicon with 5th Generation AMD EPYC hosts for hardware-level confidential computing. It has up to 2x lower latency than TPU 8t and targets regulated enterprises and governments. Google hasn't given availability dates.

For builders: Teams in finance, health or government that keep sensitive data off shared accelerators now have a confidential inference option on Google Cloud to test.

Meta releases Muse Ember

Meta released Muse Ember on 9 November. It's a lightweight model for running agent tasks cheaply, priced at $0.10 per million input tokens. Meta's cost-per-task comparisons with rival small models are its own figures.

For builders: It could handle high-volume steps such as tool calls and routing. Run it on your own evals before you replace an existing small model.

Meituan publishes LongCat-2.5 weights

Meituan put the LongCat-2.5 weights on Hugging Face on 9 November under the MIT licence, about a week after introducing the model. It's a mixture-of-experts model with 1.6 trillion total parameters and about 48 billion active per token. The model card shows its Terminal Bench v2.1 score rising to 88.4, up from 70.8 for LongCat-2.0.

For builders: The MIT licence allows commercial use. With a 1.8TB checkpoint, though, self-hosting only makes sense for teams with large multi-GPU clusters.

Sources

Talk about it with people building with AI

Our Auckland meetups are open to everyone. Members get a ForgeBox and the members' projects area for NZ$599 a year.

Become a member

Get the daily AI news by RSS