Claude Fable 5.2 ships 30% faster at a lower price, Meta's small Muse Ember, and Google's TPU 8tx and 8i Max

Claude Fable 5.2 is Anthropic's top model for the hardest coding, reasoning and long-running agent work. Claude Mythos 5.2 is the same model with the cybersecurity and biology safeguards relaxed, so only organizations approved through the Cyber or Life Sciences verification programs can use it. Both have a 1M-token context window, up to 128k output tokens, and adaptive thinking that is always on. On the API you just change the model name, and you turn on fast mode, which raises output speed up to 2.5x, by adding speed: "fast" to the request.

"model": "claude-fable-5-2"

Pricing is $8 per million input tokens and $40 per million output tokens ($16 and $80 in fast mode). It's available now on the Claude API, AWS, Google Cloud and Microsoft Foundry.

In Anthropic's numbers, Fable 5.2 scores 49.1% on FrontierCode 2.0Vendor claim and 69.9% on FrontierSWE v2Vendor claim, ahead of Opus 5.5, and at High effort it roughly matches Opus 5.5 at Max for about half the cost per taskVendor claim. Independently, it scores 65 on the Artificial Analysis Intelligence Index, first of 74 modelsIndependent. Two things to check before you switch: if your code edits the system prompt, tools or earlier messages and then sends back a Fable 5.2 thinking block, the request fails with a 400, and computer use now only accepts the new computer_toolset_20260801.

Muse Ember is the smallest and most efficient model in Meta's Muse family. It targets the work agents repeat all day: chained tool calls, coding subtasks, routing requests, extraction and reading long documents. It has a 1M-token context window and takes text, images, audio and video as input. It's available now on the Meta Model API at $0.10 per million input tokens and $0.50 per million output tokens. The weights are not released.

Most agent spend comes from running small steps hundreds of times. Meta says Ember matches Claude Haiku 5.5's best score at about a third of the cost per taskVendor claim. On the Artificial Analysis Intelligence Index it scores 36 at its top settingIndependent, second of the 32 models Artificial Analysis compares it with. If you're looking for a cheap model for sub-agents or routing, run it side by side with the small model you use now.

Google added two systems to its eighth-generation TPU lineup. TPU 8tx is for confidential computing: it keeps data and model weights encrypted in hardware even while they're being processed. TPU 8i Max packs two TPU 8i chips into one package, doubling compute, SRAM and HBM bandwidth for long-context work. Google also open-sourced TPU Sync, which tiers KV caches across HBM, host memory and storage, and put a TPU agent suite into preview that plugs into coding tools like Claude Code and Cursor to help move workloads onto TPUs and tune kernels. PyTorch code runs on TPUs directly through TorchTPU.

pip install torch_tpu

When agents keep long contexts open across many steps, memory runs out before compute does. Google says TPU 8tx cuts latency by up to 2x versus TPU 8tVendor claim and that TPU Sync reduces KV transfer latency and time to first token by more than 90%Vendor claim. For teams in regulated industries that need to fine-tune or run agents on sensitive data, it adds a confidential-inference option in the cloud.