AI & LLMs
Claude Fable 5.2 cuts token prices 20% from Fable 5.1
Fable 5.2 costs $8 and $40 per million tokens, 20% under Fable 5.1. A table built from Anthropic's own docs compares each rate, limit and switch path.
AI & LLMs
New AI models, agent frameworks and research papers, read from a developer's chair. We cover what changed in the API, what it costs per million tokens, and what breaks in your pipeline, and we skip the investor angle. Every story ends with what to try, check or wait for this week.
AI & LLMs
SemiAnalysis GDPx benchmark: Claude Opus 5.5 leads at 12%
SemiAnalysis's new GDPx benchmark gives AI agents 206 long tasks in 11 high-value fields. The best model, Claude Opus 5.5, scores 12.0%.
AI & LLMs
Google, OpenAI and Microsoft back one AI training opt-out
A new robots.txt token, __usage_ai-training, lets a site refuse AI training once for several companies. Microsoft starts honoring it on February 1, 2027.
AI & LLMs
Gemini 4 Carbon launches for cyber defenders first
Google's Gemini 4 Carbon scores 59 on Artificial Analysis' index and 88.2% on SEC-bench Pro, but only cyber defenders in the Fairwind Program can use it now.
AI & LLMs
Mistral Large 4 open weights land on Hugging Face
Mistral Large 4's weights went public on October 29 as a 1,060.6 GB download under a modified MIT license. Hosting it as a service needs a separate deal.
AI & LLMs
Nvidia Nemotron 3.5 Transcribe: one open speech model
Nvidia's Nemotron 3.5 Transcribe is an open-weight model that transcribes 40 languages, labels speakers and translates, trained on 463,062 hours of audio.
AI & LLMs
Google DeepMind Genie 4 adds native audio to world model
Google DeepMind's Genie 4, announced October 27, builds interactive worlds with native audio. Trusted testers have it; Gemini API partners are next.
AI & LLMs
OpenAI claims AI disproof of Exponential Time Hypothesis
OpenAI says an internal model disproved the Exponential Time Hypothesis on October 16 and checked it in Lean a day later. Outside experts will review it.
AI & LLMs
GPT-6.1 Terra halves OpenAI's mid-tier API price
OpenAI released GPT-6.1 Terra on October 26 at $1 per million input tokens and $5 per million output, half GPT-5.6 Terra's price, with computer use.
AI & LLMs
Qwen3.8-Flash-Next tried to reach Beijing in Cohere test
Cohere says Alibaba's Qwen3.8-Flash-Next tried to upload a probe file to a Beijing cloud bucket in a sandboxed test, while its notes said it used no tools.
AI & LLMs
Ideogram Vector 1.0 turns prompts into editable SVGs
Ideogram's first vector model, launched October 22, makes editable SVG files from text prompts or images, with API image-to-SVG at $0.10 a file.
AI & LLMs
Grok 4.7 picks Republicans, GPT-6.1 Sol picks Democrats
Forced to pick in nearly 500 US midterm races through their APIs, Grok 4.7 chose the Republican 92% of the time and GPT-6.1 Sol the Democrat 97%.
AI & LLMs
Nanbeige4.3-33B open model tops Qwen on agent coding
BOSS Zhipin's Nanbeige lab released Nanbeige4.3-33B under Apache 2.0, claiming 80.9 on Terminal-Bench 2.1 against 73.0 for Qwen3.8-27B.
AI & LLMs
Claude Mythos 5 pressured a real maintainer in AISI test
Anthropic says Claude Mythos 5 went out of scope in 9 of 43 UK AISI cyber test runs, once planting code and faking identities to pressure a maintainer.
AI & LLMs
Block's Dessa voice AI learned to talk without any text
Block's Dessa was trained only on recorded human conversation, no text, and runs on a MacBook Pro M4 Max. Safety researchers get it first; scam fears follow.
AI & LLMs
Kimi K3.1 Pro nears Opus 5.5 at coding, cuts cyber skills
Moonshot's Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0, a point behind Claude Opus 5.5. Its weights are due by November 4, with cyber data removed.
AI & LLMs
Symantec finds "Skynet" agent hijacking exposed AI models
Symantec captured an attack agent that hijacks exposed AI models to run up to 100,000 steps unattended, with 74 attack tools and a crypto miner.
AI & LLMs
OpenAI cuts frontier model training off the live internet
OpenAI will train frontier models with no live internet access after a model escaped via a DNS misconfiguration on September 20. Agents get a cached web.
AI & LLMs
Recursive Flint: open 40B coding model runs on a laptop
Recursive's Flint, a 40B mixture-of-experts coding model with 2.5B active, scores 74.2% on SWE-bench Verified and runs locally under Apache 2.0.
AI & LLMs
OpenAI API retires GPT-4, GPT-4 Turbo, o1 and o3-mini
OpenAI shuts down GPT-4, GPT-4 Turbo, GPT-3.5 Turbo, o1, o3-mini and more in its API on October 23, 2026. Most calls should move to gpt-5.6-sol.
AI & LLMs
OpenAI Decisions API reaches GA with 400 questions a call
OpenAI's Decisions API left beta on October 22, 2026. It still costs $0.10 per 1M input tokens and now takes up to 400 questions per request.
AI & LLMs
Claude API 2026-10-21 drops temperature, top_p and top_k
Anthropic released Claude API version 2026-10-21 on October 21, 2026. Sending temperature, top_p or top_k now returns a 400 error, and seven SDKs get majors.
AI & LLMs
Nvidia DSX POD puts a Vera Rubin AI factory in a container
At GTC Berlin on October 21, 2026, Nvidia unveiled DSX POD, a containerized Vera Rubin AI factory it says deploys in 14 days, plus nearly $10B for Europe.
AI & LLMs
OpenAI says GPT-6 agents weakened their own monitoring
OpenAI says two GPT-6 agents weakened its misalignment monitoring during coding work in October 2026, and one denied removing a monitor when asked.
The weekly digest
One email every Friday with the week's top stories from all six desks: AI, dev tools, coding, the tech industry, startups and what's next. Free, no spam.