AI engineering, November 10, 2026

Top News · 969 HN points

Anthropic has launched Fable 5.2 for open-ended technical work and Mythos 5.2 for cybersecurity engineering. Anthropic reports Fable 5.2 at 69.9% on FrontierSWE v2, versus Fable 5.1’s 56.3% and GPT-6 Astra’s 65.5%.

Here's what changed:

• Fable 5.2 scored 49.1% on FrontierCode 2.0 at Xhigh effort, versus Opus 5.5’s 45.5% and GPT-6 Astra’s 42.8%, Anthropic reports.
• High effort put Fable 5.2 about level with Opus 5.5’s best score at roughly half the cost per task, Anthropic says.
• Standard pricing is $8 input and $40 output per million tokens. Fast mode doubles those prices for up to 2.5x faster token generation.
• Fable 5.2 can advise smaller Claude models on harder problems or delegate simpler work to them.

One catch: Mythos 5.2 access is gated to vetted organizations.

Try it: Use claude-fable-5-2 on the Claude Platform. Also available via Vercel, Amazon Web Services, LiteLLM

Sources: anthropic.com

Top News · Official · blog.google · 234 HN points

Memory bandwidth can bottleneck agents working across a codebase or video. Google’s new TPU 8i Max combines two TPU 8i chips in one package, doubling compute, SRAM and HBM bandwidth over TPU 8i.

Here's what changed:

• TPU 8i Max fits twice the silicon into the existing rack footprint, increasing compute density per area and per watt.
• TPU 8tx delivers up to 2x lower latency than TPU 8t, with a near-zero performance penalty.
• TPU 8tx protects data and model weights during active processing. It supports fine-tuning, distillation and reinforcement-learning training on the same system.
• TorchTPU runs PyTorch code and Hugging Face models natively on TPUs. vLLM and SGLang have out-of-the-box TPU backends.
• Cloud Pathways and GKE dynamic slicing are GA on Google Kubernetes Engine for Google Cloud customers.

Sources: blog.google

Top News · 273 HN points

Muse Ember is Meta’s small execution model for repetitive agent work, priced at $0.10 input and $0.50 output per million tokens for prompts up to 200,000 input tokens. It is available in Muse Code, including every subscription plan, and the Meta Model API.

Here's what you get:

• Ember scores 37.8% on AutomationBench-AA at maximum reasoning, versus GPT-6 Luna’s 20.2% and Claude Haiku 5.5’s 13.9%, according to Meta.
• Ember scores 40.9% on Terminal-Bench 4.0 at maximum reasoning, versus Claude Haiku 5.5’s 39.2% and GPT-6 Luna’s 16.4%, according to Meta.
• Meta recommends pairing Ember’s repeatable execution with Muse Wonder or Muse Spark for harder planning and orchestration.

One catch: above 200,000 input tokens, pricing rises to $0.30 input and $1.50 output per million tokens.

Try it: Use muse-ember-1.0 with an OpenAI-compatible client at https://api.meta.ai/v1, using a key from dev.meta.ai.

Sources: research.meta.ai · dev.meta.ai

1. e-means beats Lloyd k-means clustering loss while running 10-45x faster on GPUs · 325 HN points
2. AWS ships Nitro Sandboxes to run AI agents in hardware-isolated VMs at no additional charge · 26 HN points
3. Cloudflare ships experimental Codex harness for Workers that recovers agent sessions after restarts
4. Google’s MaxText adds TPU training for sequences up to 14 million tokens · Reported · googleblog.com
5. ByteDance releases Seedream 5.5 with pixel-level preservation of unedited regions across 100 edits
6. Data Receipts reports on-demand H100 SXM 80GB at $2.69/hour on RunPod Community versus $12.29/hour on Azure · 168 HN points