In partnership with
Anthropic's new top model beats Opus 5.5 for $8/$40.
Meta built a small model for agent grunt work.
An open Git server copies OpenAI's Pando design.
▲ AI DEV SIGNALS
by Marktechpost · November 10, 2026
Hi there,
Here's what moved in AI dev since last Tuesday. Anthropic split one model into 2 safeguard levels. Meta built a small model for repetitive agent steps. An open Git server copied OpenAI's Pando design. And Together AI turned inference tuning into a loop that never stops.
01 · Featured: same weights, 2 guest lists
Anthropic shipped 1 frontier model under 2 names. Fable 5.2 blocks high-risk cyber and most research biology. Mythos 5.2 drops those blocks for verified orgs.
How it works: Both share a 1M-token context and 128K max output. When Fable blocks a query, Opus 4.8 or Opus 5 answers instead. Anthropic reports about 55% fewer blocks than Fable 5.1. Mythos access runs through the Cyber and Life Sciences Verification Programs.
The numbers: FrontierCode 2.0 hits 49.1% at Xhigh effort, versus 45.5% for Opus 5.5. Remote Labor Index reaches 41.3%, versus 21.3%. At High effort, Fable matches Opus 5.5's best FrontierCode score at roughly half the cost. Pricing is $8/$40 per 1M tokens; fast mode costs double for up to 2.5x speed.
The fine print: All scores are self-reported. Gemini 4 Carbon still wins AutomationBench, 53.9% to 51.4%. No open weights. Anthropic still recommends Opus 5.5 as the daily driver. In 1,800+ sandbox-escape scenarios, Mythos 5.2 used out-of-sandbox credentials 3.1% of the time.
Our take: The effort dial is the real product. Route routine work to lower effort and save Xhigh for long-horizon runs. Check whether High clears your bar before paying for more. Read the full breakdown.
AI is only as good as what you tell it
The fastest way to give AI more context is to talk. But that only works if the transcription is right. Wispr Flow runs on Canto, its own speech model, trained on real dictation instead of studio audio.
In Wispr's tests on 10 hours of real dictations from 2,300+ speakers, Canto had the lowest word error rate of every model tested, including Google and OpenAI. It also pulls from your personal dictionary, so tool names, project names and teammates come out spelled right.
Talk through the whole problem, hit enter, and get a better answer. 4x faster than typing, in every app.
02 · 4 more drops: a small Meta model, an open Pando, and a compiler that never stops
Meta's smallest model targets the repetitive steps inside agents: tool calls, routing and extraction. It ties Claude Haiku 5.5 (max) at 36 on the AA Intelligence Index. Meta puts cost at $0.136 per index task versus $0.411. The catch: hosted only, and it trails Haiku on FrontierCode 2.0 (27.2 vs 34.9).
Aspen independently implements the design OpenAI described for Pando. It is 1 Rust binary, with branches in PostgreSQL or SQLite and packs in object storage. It sustained 567 pushes per second on 1 8-core server. The catch: it's pre-alpha, and landing on main runs about 4 per second.
It takes text and images with a 1M-token context. It scores 31 on the AA Intelligence Index, up from 13 for Inkling Small. It posts 86.3% on SWE-Bench Pro v2 (Hard). The catch: Tinker API only for now, with Apache 2.0 weights planned, and 13.1% on Terminal-Bench 4.0.
Coding agents rewrite a model's forward pass, and the loop reruns when traffic drifts. It beat the best published InferenceX configurations by up to 2.6x TPS, in under 24 hours per setting. The catch: gains vary; GLM 5.3 on MI355X rose only about 4%, from 463 to 480 TPS.
03 · Signals: a 2.8T download, a stale CLAUDE.md, and poisoned memory
Kimi K3.1 Pro weights are now downloadable. Moonshot posted the 2.8T MoE (104B active, 1M context) in MXFP4 under a custom MIT-style license.
Many small agent teams beat 1 big team. With a 27B model and 100 agents, teams of 10 beat parallel single agents on 10 of 12 tasks. Code included.
Stale instruction files cost coding agents real tool calls. Across 732 trajectories, agents followed outdated guidance; a stale test command added about 2 tool calls per run.
Every agent memory system tested could be poisoned. MemoryDojo covers 155 multi-session trajectories across 4 domains and 5 memory systems.
A generalist model drives robot arms better than VLAs. The AMP1 harness hits 85.6% zero-shot on RoboLab-120, versus 49.2% for a leading VLA.
A 278M PII detector with 69 entity types. Protegrity-PII ships with the PII-SCOPE benchmark, under a research and non-commercial license.
Proof that a reply came from the model you paid for. Anthropic's FlashProve zkSNARK prototype proves Llama 4 Maverick at 110 tokens/s on 1 H100.
Specs first, then 5 zero-context reviews. Google's Elephant-Goldfish plugin makes coding agents draft specs before writing code.
KV-cache-aware routing for vLLM on Kubernetes. Buoyant's Kinference is Apache 2.0 and cut median TTFT 95% in its own benchmark.
FP8 attention kernels for AMD MI355X. Zyphra's ZYFA runs 1.16–1.76x faster than AMD's AITER on real activations.
A 65B dense open model under Apache 2.0. Ricursive's Ripple-65B Preview scores 80.5 on Terminal-Bench 2.1.
Google puts 2 TPU 8i chips in 1 package. TPU 8i Max targets long context; TPU 8tx targets confidential inference. No specs or prices yet.
1 question
If High effort matches Opus at half the cost, which of your tasks actually need Max?
Have a great Tuesday.
until next week,
the marktechpost team
REACH 70K+ AI BUILDERS
Was this email forwarded to you? Sign up here