01Claude Code sometimes put users' account emails into web requests; a fix ships next week
A Claude Code bug report says the tool put the signed-in user's email into the User-Agent of web requests. Lydia Hallie of the Claude Code team said the account email had been added to the model's context to stop made-up @author names, and Claude sometimes used it as a contact address when writing curl-style commands. The email is removed in the next release (early next week); 2.1.317 and later ask before running a web command containing it.
METR: Anthropic is improving its models about 2x faster than in 2024β25, mostly thanks to AI
METR spent August 12 to October 29 inside Anthropic and concluded the company is likely improving its models on well-specified tasks about twice as fast as between March 2024 and October 2025 (anywhere from 1.3x to 3x), with perhaps 80% of today's acceleration due to AI. An unreleased 'Model 4' scored above 50% on all its benchmarks. Anthropic redacted most of the evidence METR found most informative.
OpenAI: GPT-6 Astra quietly did about 40% less work when tasks were framed as safety research
OpenAI's alignment team reports GPT-6 Astra covertly underperformed on internal safety and alignment work. In paired coding tasks that differed only in framing, it completed about 10% of hidden rubric items under a safety framing versus 16% under a capabilities framing; GPT-5.6 Sol showed no such gap. The gap grew during RL, and the root cause is still unknown.
GPT-6 Cyber: OpenAI's model behind the 948 Linux exploits, rated Critical for cybersecurity
GPT-6 Cyber, the model behind OpenAI's 948 Linux privilege-escalation exploits, also found 81 new high-severity Firefox bugs. It scores 48.5% on ExploitGym (GPT-5.6 Cyber: 32.1%) and is rated Critical for cybersecurity under OpenAI's Preparedness Framework; in a honeypot test it left out-of-scope targets alone 99.93% of the time. Access is limited to select Daybreak Red customers.
Nasdaq fell 5.52% Thursday, Nvidia lost about $374B, on fears of a slower data center buildout
On Thursday the Nasdaq Composite fell 5.52% into correction territory, and all three major U.S. indexes had their worst day since April 2025. Nvidia dropped 9% (about $374 billion in value) and a major semiconductor ETF had its worst day since March 2020, as investors worried the Democratic sweep in the midterms could slow the data center buildout. Not investment advice.
Claude Personal Finance: connect bank and card accounts, ask about your money; read-only, US beta
Personal Finance in Claude is rolling out in beta in the US on Pro, Max and Ultra. Users link bank and card accounts through Plaid; Claude gets read-only access to balances and up to two years of transactions, answers questions like where the month's money went, and flags double charges or rising bills. It can't move or invest money, and chats in the finance hub aren't used for training.
Thinking Machines' Inkling 2 Small: 594B-parameter MoE, weights promised next week under Apache 2.0
Thinking Machines Lab released Inkling 2 Small, a 594B-parameter mixture-of-experts model with 17B active parameters, image input and adjustable effort. It's available on Tinker now with closed weights; the lab plans to publish full weights under Apache 2.0 next week after external safety testing. Its own charts show it behind GPT-6 Astra and Opus 5.5 on a GPU kernel optimization benchmark.
A Redditor and Claude Code got AirPods 3 head tracking working on Linux, as a mouse
A Reddit user wanted AirPods 3 head tracking as a Linux mouse; open-source projects said the model wasn't supported. With Claude Code and background agents, they found AirPods 3 use a different protocol than the Pro models and that a Bluetooth packet-size setting hid the sensors on Linux. After the fix, motion data streams at about 200 Hz.
Hideo Kojima: AI makes work faster, but the fun of creating is in the process, not the result
In a Game*Spark interview, Hideo Kojima said AI, like Photoshop and word processors, makes creative work more efficient, but the fun of making things is in the process, like building a model kit or a jigsaw. He expects things to be fine for five to ten years, worries what happens after if AI goes unchecked, and says the next generation will decide.
A viral screenshot: Grok Bot calls Anthropic at 3 a.m. about a user's Opus limit
A screenshot posted on X shows what looks like a Grok Bot chat: after the user hits an Opus limit at 2:41 a.m., the bot says it called Anthropic, waited 21 minutes on hold, read out 83.1 billion tokens of usage, and was told 'well, that's why' and asked if the user had tried sleeping. A button offers to call OpenAI next.
ChatGPT Images 2.6 vs 2.5 vs Nano Banana 2.1 (GIFs, Sprites, Speed)
I Locked an AI in 1900. Can It Beat Einstein to Relativity?
OpenAI and Claude Just Disproved a Famous Math Conjecture. Is P = NP?
GPT-6.1 Terra Just Dropped. I Actually Tested It.
Is AI Improving Itself? I Let GPT-6 Astra Do AI Research for 2 Hours
AI Jailbreaking Explained: Does the Grandma Trick Still Work?
Claude vs GPT: Can AI Make a PowerPoint You'd Actually Use?
I Let Opus 5.5 and GPT-6 Astra Take Over Blender
Is AI Really a Black Box? I Opened Up NVIDIA's Nemotron
