AI news for Tuesday, November 10, 2026: Claude Fable 5.2 ships, Meta's Muse Ember undercuts on price

Two new models, a regulator and a breach: running agents keeps getting cheaper, and trusting them keeps getting more expensive.

Models Safety3 min
Anthropic ships Claude Fable 5.2 with 55% fewer blocks
Anthropic released Claude Fable 5.2 to all users on November 9 at $8 per million input tokens, and says it blocks about 55% fewer requests than Fable 5.1. The benchmarks will get the headlines. The fewer refusals are what you will feel.
Why it matters If your security or biology work kept hitting Claude's blocks, retest it this week. And if you only need a strong default, Anthropic itself points most coding and knowledge work to Opus 5.5, so don't pay $40 per million output tokens by reflex.
SRCAnthropic
Models Agents3 min
Meta's Muse Ember prices agent work at $0.10 input
Meta released Muse Ember on November 9 at $0.10 per million input tokens and $0.50 per million output, with a 1 million token context window. Meta says it beats Claude Haiku 5.5 and GPT-6 Luna on agent tasks. Make it prove that on your own calls.
Why it matters If your agents spend most of their tokens on routing, extraction and tool calls, Ember could cut that part of the bill hard. Test it on your own repetitive steps before you trust Meta's chart, and watch the price triple past 200,000 tokens.
SRCMeta AI Research
Policy Safety3 min
Australia plans an AI safety regulator after OpenAI breach
Australia's government said on November 10 it will legislate mandatory safety obligations for frontier AI companies, including independent model testing and rapid reporting of rogue behavior, enforced by a new regulator. The target is agents that misbehave, not chatbots that misspeak.
Why it matters If you build or deploy agents, rapid disclosure of rogue behavior is heading into law, starting in Australia. Know what your agents can reach and log what they do now, because a regulator will ask for both.
SRCiTnews
Safety Business3 min
Fireworks AI breach reached some customers' model weights
Fireworks AI says an attacker used a cloud provider's October 23 breach to reach its development machines and accessed some customers' incomplete model weights, datasets and secrets. Its inference was not hit. Rotate your keys anyway.
Why it matters If you trained or evaluated models on Fireworks, treat your secrets as exposed until Fireworks tells you otherwise: rotate them and ask whether your weights were in the accessed subset. If you use another inference vendor, ask who hosts its research machines.
SRCFireworks AI
Hardware3 min
Google adds TPU 8tx for confidential AI inference
Google announced TPU 8tx, for confidential inference, and TPU 8i Max, for long-context work, on November 9, its third and fourth TPU systems this year. There is no price or date yet, so the software is what you can use today.
Why it matters If your data can't leave a locked-down environment, TPU 8tx is a Google chip pitched straight at you. Ask Google Cloud for pricing and dates before you plan around it, and try TorchTPU now if you want PyTorch on TPUs.
SRCGoogle

ShareX LinkedIn Hacker News Email

Prompt of the day

Split your agent's work between a big model and a cheap one

› I run an AI workflow and want to cut what I pay for model calls without losing quality. Here are the steps it performs, roughly how many times a day, and the typical input and output length of each:
[list your steps, e.g. "classify incoming email, 2,000 a day, 800 tokens in, 20 out"]

The prices I'm comparing, per million tokens (input / output):
Large model: [price]
Small model: [price]

For each step:
1. Say whether it needs a large model's judgment or is repetitive work a small model can handle (routing, extraction, tool calls, formatting). One sentence of reasoning.
2. Estimate the daily cost on each model and show the math.
3. Name the failure I'd see if the small model gets it wrong, and a cheap check that would catch it.

Then give me a routing plan: which steps go to which model, the total daily cost before and after, and a 20-example test I can run this week to confirm the small model holds up. Flag any step where you're unsure and tell me what data would settle it.

Open in Magai

Works in any AI chat. In Magai, run it against several models at once.

Get every edition

This one arrived in inboxes at 6am ET.

Anthropic ships Claude Fable 5.2 with 55% fewer blocks

Meta's Muse Ember prices agent work at $0.10 input

Australia plans an AI safety regulator after OpenAI breach

Fireworks AI breach reached some customers' model weights

Google adds TPU 8tx for confidential AI inference