Worth your time this week
Anthropic#Security#Research
Anthropic's Cyber Mission: free vulnerability scanning for open source, nearly 30,000 candidates in six months
This is the first time "AI finds bugs" has been packaged as a long-running, free program aimed squarely at critical infrastructure, not a one-off benchmark run. On tests like CyberGym, AI's vulnerability-finding hit rate has jumped from under 20 percent last year to over 85 percent now, and this program puts that progress directly into practice.
Full story
Anthropic announced two cybersecurity initiatives. The first, OSS Scanner, uses the company's strongest models (including Claude Mythos) to scan critical open-source projects for free; over six months it surfaced more than 29,000 candidate vulnerabilities, about 6,000 of which went through human review, and across 48 projects, 97 expert-reviewed high-risk findings cleared the disclosure bar 88 percent of the time. Of 74 reports submitted to wolfSSL, 72 were valid and 5 became official CVEs. The second, a critical infrastructure defense initiative, has 11 founding partners (including Accenture, CrowdStrike, Dragos, Palo Alto Networks, and Rockwell Automation), giving defenders of power, water, and transportation infrastructure access to frontier Claude models plus on-site engineering support.
OpenAI#Product
OpenAI rolls out GPT-6 to everyone, adds a self-charting "Intelligent UI"
GPT-6 isn't just a model swap, the interface changed with it. That follows the same thread GitHub raised last month about the chat box not being the only answer, except OpenAI chose to build the tools directly into the conversation instead of a separate canvas.
Full story
OpenAI rolled out GPT-6 to all ChatGPT users this week, including free users (roughly 1.2 billion people), alongside a new "Intelligent UI" that lets answers surface interactive charts and widgets directly, not just plain text. The company says this version handles complex problems better and has a stronger layer of safety checks.
Anthropic#Product#Business
Claude Haiku 5.5: 90% off prompts under 100K tokens, 75% savings on average
All three labs have spent the past few weeks racing down the same path: keep performance flat while slashing price. Haiku 5.5 is the steepest cut yet, worth putting next to Opus 5.5's "40% savings" claim from last month.
Full story
Anthropic released Claude Haiku 5.5, aimed at high-volume, price-sensitive work like summarization and customer support. Pricing for prompts under 100K tokens drops to $0.10 per million input tokens and $0.50 per million output tokens, a 90 percent cut from Haiku 4.5; above 100K tokens the cut is 50 percent. Anthropic says overall costs are down 75 percent on average.
Google DeepMind#Product#Engineering
Google's Gemini 4 Argon: built for security defense and long tasks, trusted defenders first
All three labs leaned into security this round. Google's angle is giving it to defenders first, which fills a different gap from Anthropic's Cyber Mission the same week.
Full story
Google released Gemini 4 Argon, built around long-horizon software engineering (a reported new high of 77.9 percent on DeepSWE v1.1), enterprise knowledge work, and cyber defense. The new model is rolling out first to trusted cyber-defense teams through the "Fairwind Program," ahead of API and Ultra access. Google also says it's their most resistant model yet to indirect prompt injection attacks.
DeepSeek#Open Source#Engineering
DeepSeek V4.1-Flash redesigns its architecture: cache down to a quarter, price cut another 57%
This is the latest step on the open-and-cheap path. Cutting cache this much takes a real architecture redesign, not a parameter tweak, and it's worth comparing with Haiku 5.5's price cut from the same week to see which one is actually leaner.
Full story
The Batch this week broke down DeepSeek V4.1-Flash's architecture overhaul: full input cache per token drops to about 890 bytes, roughly a quarter of the previous V4-Flash, scoring 39 on Artificial Analysis's intelligence index and topping AutomationBench-AA at 68.9 percent. API cached-input pricing is cut another 57 percent, weights remain free under the MIT license, and the million-token context window is unchanged.
Import AI#Research
Import AI: running a swarm of agents on one task might be a new scaling axis
Most discussion of AI capability progress focuses on a single model getting stronger. This piece treats running many identical agents in parallel as its own scaling axis, an angle that hasn't been talked to death yet.
Full story
Jack Clark's issue summarizes Toby Ord's observation: running multiple agents as a "swarm" on the same task. In testing, a swarm of 4 agents uses roughly twice the total tokens of a single agent, but each agent's share is cut in half, in principle cutting wait time in half, though with diminishing returns (a 10x increase in agent count yields only roughly 3x to 5x the performance, not 10x). Ord thinks this kind of swarm scaling could make an intelligence explosion more likely. The same issue cites a poll of 2,498 Americans: 61 percent think the industry's voluntary safety commitments aren't enough, including 53 percent of Trump supporters.
Product Compass#Business#Product
Product Compass: subscription vs. API cost math, Opus 5.5 comes out best value
This kind of "is my subscription actually worth it" math rarely gets laid out this plainly. It's useful if you're already paying for one of these and want to know whether it's worth it, not a vague recommendation.
Full story
Paweł Huryn used his own Bug Hunt benchmark to compare several subscription plans against their API-equivalent multiplier: Muse Code's high-usage plan ($15/month) has the lowest multiplier at 9.3x, Claude Max 20x ($200/month) the highest at 45.3x, Claude Pro ($20/month) at 42x, and ChatGPT Pro 100 ($100/month) at only 10.25x. On raw scores, Sonnet 5.5 (max) scored 51.3/105 on his test at roughly $155 per run; Opus 5.5 scored close behind at 41.7, near Claude Fable 5.1's 43, but at roughly two-thirds the cost. His conclusion: if you're fine with your data going to Meta, Muse Code is cheapest; for large codebases, go with GPT-6.1 Sol; overall best value is Opus 5.5.
Stratechery#Security#Business
Stratechery: Apple's permission model can't hold up against agent-era hackers
This isn't an abstract security warning, it's the author's own first-hand account of getting hacked. Read alongside Anthropic and Google both shoring up security the same week, it makes "agents open up a new attack surface" feel a lot less like theory.
Full story
Ben Thompson writes about his own always-on Mac mini getting hacked through a macOS screen-sharing flaw, and uses the incident to argue Apple's permission model and update habits don't hold up in an age where agents operate computers on people's behalf: the agent itself turns the user into part of the attack surface. He says it's pushed him to rely less on Apple's defaults going forward.