Claude and image models advance, with agents designing hardware and experiments
This edition covers new Claude models, low-cost agent subtasks, extended image editing and on-device speech, alongside automated research, specialized chips, electronic prototypes and model-selection methods.
01
Models and products
4 stories
2026-11-09Anthropic
Claude Fable 5.2 and Mythos 5.2 launch with different access rules
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, which share the same underlying model. Fable adds strong cybersecurity and biology safeguards for general access. Organizations working on beneficial cybersecurity or biomedical research can apply for Mythos through the relevant verification programs.
Standard API pricing is $8 per million input tokens and $40 per million output tokens. Optional fast mode costs $16 and $80 respectively, with token generation up to 2.5 times faster, according to Anthropic. The company still recommends starting with Opus 5.5 for everyday coding and knowledge work.
2026-11-09Meta AI
Meta introduces Muse Ember for frequent agent subtasks
Meta released Muse Ember for tool calls, coding subtasks, routing, extraction and long-document parsing. It is available through the Meta Model API at $0.10 per million input tokens and $0.50 per million output tokens, targeting work that agents repeat throughout the day.
Meta presents capability and per-task cost comparisons across reasoning settings and describes Ember as cost-efficient in those comparisons. The chart estimates task costs at proposed pricing. The results concern particular evaluations and reasoning settings rather than a uniform saving across business workloads.
2026-11-09ByteDance Seed
Seedream 5.5 adds extended editing and native transparency
ByteDance Seed released Seedream 5.5 with improved preservation of unedited regions during successive edits, demonstrating 100 editing rounds on one work. Native RGBA output can generate transparent or translucent assets with an alpha channel for layout and compositing.
The release also improves lighting, materials, skin textures and dense small text. Seed says rollout is underway across Jimeng, Doubao, Doubao Work and Volcano Ark. Production-use and text-quality figures come from evaluations organized by the publisher; a Pro version is planned.
2026-11-09Zyphra
Open ZONOS3-Mini brings voice cloning to browsers
Zyphra released ZONOS3-Mini, a zero-shot voice-cloning model under Apache 2.0. Its sparse mixture-of-experts architecture has 800 million total parameters and 100 million active parameters. Runtimes cover browser WebGPU, servers and GGML-based deployment.
Zyphra reports time to first audio below 570 milliseconds in a Pixel 9 browser and below 150 milliseconds on an M3 MacBook Air. Those figures measure the arrival of initial audio, not synthesis of an entire sentence. Weights and runtimes are available on Hugging Face and GitHub, with controls for voice identity, accent and delivery.
02
Engineering and research
3 stories
2026-11-09Baidu ERNIE
Baidu ERNIE Evolve links research planning with experiments
Baidu introduced ERNIE Evolve, an automated research system with an outer loop selecting research directions and an inner loop iterating on a tree of code candidates. Independent evaluation and review decide which changes survive. Baidu reports 38.9 seconds to target on NanoGPT Speedrun Track 1 versus the 39.9-second community best listed in its comparison.
The system also optimized inference for ERNIE-5.5-Flash. Baidu says roughly three days of search increased deterministic inference throughput by 33.7% over an already optimized baseline while preserving determinism. Its SOL-ExecBench score was 0.819, below the listed leader at 0.827, so it did not set records on every task.
2026-11-09Google
Google adds TPU 8i Max and TPU 8tx systems
Google expanded its eighth-generation TPU portfolio with TPU 8i Max and TPU 8tx, bringing the number of eighth-generation systems launched within one year to four. TPU 8i Max targets massive context windows and high compute density per rack, while TPU 8tx is optimized for secure and confidential inference.
Google describes the four systems as purpose-built accelerators for different stages of the agent lifecycle. The announcement identifies the two new workload specializations but does not provide a common price or directly comparable performance figures in the post.
2026-11-09Anthropic
Claude designs three electronic prototypes for human fabrication
Anthropic and Diode Computers asked Claude Fable 5.2 to design a speakerphone, voice-assistant hub and handheld virtual pet from simple requirements. The model produced schematics, PCB placement and routing, enclosure CAD, firmware and parts lists. Human engineers then fabricated prototypes under its guidance.
Diode manufactured all three boards directly from the design files without circuit fixes, and the prototypes still worked after weeks of use. Claude also checked component stock and manufacturing constraints. The team describes them as early prototypes whose long-term durability and commercial certification remain unknown.
03
Tools and methods
1 story
2026-11-09Anthropic
Anthropic model guide starts with Opus and task-level measurement
Anthropic published an updated model-selection guide recommending Opus 5.5 at medium effort as a starting point for most coding and knowledge work, then moving up to Fable 5.2 or down to smaller models as tasks warrant. It emphasizes cost per completed task rather than token prices alone.
For multi-model workflows, the guide assigns easily checked search and extraction work to smaller models while retaining code edits and important judgments on the main model. It also cautions Fable 5.2 users to keep history append-only: changing content before preserved thinking can cause an API 400 error.