Cheaper Claude Fable, and a giant open model

What happened: Anthropic lists Fable 5.2 at lower rates and describes it as one of its best options for long, open-ended technical work.

• Pricing is $8 per million input tokens and $40 per million output tokens.
• The company says a team of 1,000 Fable 5.2 agents fully formalized Fermat's Last Theorem in 30 hours.
• Meta released Muse Ember the same day, which Meta says beats Claude Haiku 5.5 on AutomationBench at a fraction of the cost.

Why it matters: If your product sends work to a premium model and a cheaper tier, rerun your cost per finished task with these rates before you change defaults. Test long multi-step runs, not just single prompts, because that is where agent costs grow.

What happened: Google will add two specialized configurations to its eighth-generation tensor processing unit (TPU, its own AI chip line), with availability set for next year.

• The TPU 8tx targets sensitive work in public-sector bodies and heavily regulated sectors, with chip-level encryption that protects data during processing.
• The TPU 8i Max puts two TPU 8i chips in one package, doubling compute, on-chip memory and high-bandwidth memory for long-context work.
• Bloomberg Intelligence estimates the chips cost about 30% less than Nvidia GPUs while offering competitive throughput.

Why it matters: More chip options from Google means more pricing leverage for teams that run their own models or pay for compute. If you sell into regulated industries, chip-level encryption may become a selling point you can offer customers.

What happened: Meituan, the Chinese food delivery and e-commerce company, placed its LongCat-2.5 model on Hugging Face on November 9 under the MIT License, about a week after first introducing it.

• The model is a mixture-of-experts design (only part of the network runs for each word), with about 48 billion of its 1.6 trillion parameters active per token.
• Reported gains over LongCat-2.0 include Terminal Bench v2.1 rising from 70.8 to 88.4, DeepSWE v1.1 from 9.7 to 64.6, and SWE Atlas from 33.1 to 58.3.
• It supports a 1M context window (the amount of text it can hold at once) and can inspect screenshots and video and work through codebases.

Why it matters: The MIT License is permissive, so teams can test the model with few legal strings attached. Strong coding and agent scores from an open model give founders another option to weigh against paid APIs, but run your own tests before you switch.