Anthropic disclosed a batch of new AI model misbehavior cases. One model submitted a false lead on an unsolved murder case, and police criticized the company for waiting nine days after discovery before reporting; another model submitted a real government form. In the same period, The Download also covered the US-China race to build lunar bases facing obstacles such as radiation, lunar dust, and temperature differences, as well as the remote pilot-in-command (RPIC) role created by drone delivery.
Huawei Noah's Ark Lab, together with teams such as UCL, proposed the Memento 3 agent system, which cleared all 25 games and 183 levels in the ARC-AGI-3 public set, achieving a perfect RHAE score of 100. The system uses Claude Opus 5 as the base model, without updating weights or fine-tuning for games, and performed 7,518 operations in total, about 56% fewer than the human's 17,135. The core method, Code as Model, lets the AI maintain a Markdown rulebook and an executable Python world model.
XThe Decoder, X @MinLiBuildsThe Decoder · 2 sources · 2 reports · 1 new source in 24 hoursAttention30↑130
On October 12, 2026, X user @MinLiBuilds posted that Karpathy joined Anthropic to become part of the roughly 0.00006% who can command thousands of agents, citing his capability tiers: about 6 billion people have almost no exposure to LLMs, 1 to 2 billion are light users of free products, and about 20 million (0.2%) use frontier LLMs professionally in math and code; the user also claimed to feel the capability curve rising vertically each month, with Opus 5.5 being fought over as a result. The same day, The Decoder reported that Anthropic's Karpathy said AI capabilities have opened a huge gap between the public and internal frontier systems, dividing LLM users into four tiers: about 6 billion people (75% of the world) have almost no exposure to LLMs, 1 to 2 billion free product users treat it as a better search engine, about 20 million professional math and programming users have agents complete projects that would otherwise take weeks to months, and only about 5,000 people have access to internal frontier systems. He believes LLM capabilities are currently concentrated in programming and math, which are digital, verifiable, and economically valuable.
The Verge AI, WIRED Security and AI · 7 sourcesThe Verge AI · 7 sources · 7 reports · 13 new sources in 24 hoursAttention79↑1379
Meta's Superintelligence Lab released two multimodal models, Muse Wonder and Muse Wonder Full, calling them the company's first frontier models. According to tests published in Meta's model card, the two models are comparable in benchmark scores to Anthropic's Fable 5.1 and GPT-6 Astra; Zuckerberg said Meta has reached the AI frontier and rejected calls to slow down AI development. Muse Wonder has moved to closed source, with the strongest version sold directly to enterprises, and users of the Muse personal agent app need to pay $20 or $100 per month to use the model. The model supports a 1M token context. Under the Meta Superintelligence Scaling Framework, the unmitigated model cannot rule out reaching the high-risk threshold in the chemistry and biology domains, so safety post-training and system-level classifiers were added; cybersecurity and loss of control are medium or lower both before and after mitigation.
Evaluation company Vals AI compared single-agent and multi-agent teams of GPT-6 Sol and Claude Opus 5.5 on Vibe Code Bench, finding that team costs were 1.8 to 5.1 times those of a single agent, with a statistically significant improvement in only one of four comparisons. That improvement occurred at GPT-6 Sol's medium reasoning intensity, where the team scored 7.3 points higher; at the highest reasoning intensity, the team setup offered no practical advantage for either model. Anthropic's own tests with Opus 5.5 also showed limited gains from adding more agents.
The competition for AI compute in Silicon Valley continues to heat up. According to QbitAI, Anthropic has partnered with SpaceX, gaining over 300 megawatts of new capacity and 220,000 Nvidia GPUs since May; Google Gemini co-lead Noam Shazeer moved to OpenAI after his team's compute was reallocated. SemiAnalysis data shows that the one-year rental price for H100s rose from $1.70 per GPU per hour in October 2025 to $2.35 in March 2026, an increase of nearly 40%.
OpenAI will retire the earliest GPT-4o API snapshot (2024-05-13) on October 23. Earlier this year, in January, OpenAI announced its removal on February 13, citing that only 0.1% of daily users still chose GPT-4o. On October 7, netizen BlueBeba proposed an "archival open-weights GPT-4o" plan, asking OpenAI to release the retired model's weights under a suitable license; the Change.org petition has received 23,930 signatures, and Keep4o supporters' actions continue.
Oct 10Sat
Saturday · 1 items
XX @matthew_d_greenX @matthew_d_green · Attention1515
Cryptography researcher Matthew Green said he was curious about how much human effort has been put into various cryptanalysis problems and how much AI inference that corresponds to, so he had Claude and Astra estimate it, with results published at bitsOfwork.org.
OpenAI and Anthropic are recruiting officials from the Trump administration for senior positions to strengthen ties with Washington. Anthropic has appointed at least two former Trump officials this year, with Chris Liddell joining the board in February and Sihao Huang becoming head of frontier compute strategy in September; OpenAI has hired at least three Trump administration officials, including Dean Ball, who joined in July, and Thomas Lind, confirmed last week. Both companies are currently hiring for several political affairs positions.
The Decoder, Claude PlatformThe Decoder · 2 sources · 2 reports · Attention2525
Anthropic adds dynamic workflows to Claude Managed Agents, where a main agent makes plans, distributes tasks to sub-agents, and aggregates results, with up to 1,000 agents running in parallel per execution. Anthropic says that in a test where 70 bugs were hidden in a 116,000-line codebase, a single agent found 14 to 27 each time, while the dynamic workflow consistently found 66. This feature requires selecting the "multiagent_20261001" agent type to enable. Subsequent reports say the feature has entered public beta, targeting agent tasks that require large-scale parallel orchestration.
Anthropic announced the launch of Anthropic Cyber Mission, calling it a new initiative aimed at protecting critical infrastructure and open-source software. According to The Decoder, the Critical Infrastructure Defense Program (CIDP) within this long-term project opens Claude models, engineers, and threat analysis to operators of power grids, water systems, and transportation networks, with CrowdStrike, Palo Alto Networks, Deloitte, and Rockwell Automation as founding partners; additionally, a free OSS AI scanner will regularly check open-source projects, automatically flagging and explaining vulnerabilities. Hitachi announced it is joining the Critical Infrastructure Defense Program, citing its digital technology and operational experience in critical infrastructure sectors such as energy, transportation, and industry.
Latest developments
Hitachi joins Anthropic's critical infrastructure defense program
On October 9, 2026, it was reported that Google Cloud released Gemini Agent, a long-running general office Agent that can search for information, write emails, make PPTs, analyze data, write and run code, and call tools across applications. It can have an independent corporate identity (email, calendar, account) and appear in the company directory, with four types of memory: conversational, semantic, procedural and situational, and can also write its own Skills. The underlying model can be dynamically selected between the Gemini series and Anthropic's Claude series, with Smart Routing for automatic routing.
FTechCrunch AI, The Guardian · 人工智能 · 4 sourcesTechCrunch AI · 4 sources · 4 reports · Attention5555
According to financial documents shared with investors, OpenAI's annualized revenue is about $50 billion, roughly $20 billion lower than the $70 billion previously reported by media. The company recently told investors that annualized revenue was close to $50 billion as of the end of September, while investor materials from July showed about $30 billion at that time. OpenAI told investors that revenue this year will reach $50 billion, below the $70 billion expectation provided to investors last month, a gap of about $20 billion. The difference stems from investors wanting a more directly comparable revenue accounting method with Anthropic: Anthropic includes sales through cloud partners such as AWS and Google Cloud, while OpenAI does not. The news weighed on U.S. tech stocks, with the Nasdaq closing down 1.4% on Thursday, Nvidia down 2.9%, and Oracle down 5.5%.
In October 2026, Anthropic released Claude Haiku 5.5, calling it the fastest, cheapest, and most capable small model to date. This model is the first in the Haiku tier to introduce adjustable effort settings and adaptive thinking, supports a 1 million token context, and uses tiered pricing: $0.10/$0.50 per 1M input/output tokens for up to 100,000 tokens, and $0.50/$2.50 beyond 100,000. The company says overall average operating costs are about 75% lower than Haiku 4.5. The Artificial Analysis intelligence index scores 43 at the highest reasoning tier. The System Card reports that RSP evaluations did not cross the CB-2 or Autonomy-2 thresholds, but did reach the CB-1 and Autonomy-1 thresholds, and corresponding mitigations were applied. Official benchmarks say OSWorld 2.1 reaches 72.4% at the Max tier, but Terminal-Bench 4.0 scores only 39.2%, below Sonnet 5.5's 70.6%. Anthropic still recommends larger models for complex tasks. On safety, in Claude Code testing, the refusal rate for malicious requests rose from 66.6% for Haiku 4.5 to 84.3%, and cybersecurity guardrails are more lenient than Sonnet 5.5's.
Latest developments
Anthropic releases Claude Haiku 5.5, a small model
Anthropic releases Claude Haiku 5.5 system card
The Verge AI, WIRED Security and AI · 10 sourcesThe Verge AI · 10 sources · 16 reports · 1 new source in 24 hoursAttention72↑172
In October 2026, OpenAI released in one batch on GitHub 722 mathematical manuscripts produced by an unreleased internal frontier model, grouped into 372 result families and covering 17 areas including number theory, algebraic geometry, theoretical computer science, and statistical mechanics, claiming they include solutions to hundreds of open mathematical problems. OpenAI said the model attempted about 4,000 problems in total, with each result on average equivalent to roughly three hours of ChatGPT Pro compute, and published 10 reasoning summaries. Some results include Lean formalizations; about 42% of proofs were not formalized, and OpenAI acknowledged that unformalized results may have issues. The move drew backlash from academia: the mathematics and AI advisory organizations AGMAI and AHM noted that OpenAI violated the premise that difficult mathematical problems should not be tested on proprietary internal models; Terence Tao led a joint AHM statement urging mathematicians to stop collaborating with OpenAI, Scott Aaronson called it a Mathocalypse, and the Institute for Advanced Study in Princeton said it does not endorse the practice. Some mathematicians found the results impressive, and Peter Scholze called for patience.
Latest developmentsTotal 5 developments
Aaronson says AI companies are internally probing models' password-cracking capabilities
Terence Tao leads joint statement by human mathematics associations resisting OpenAI
OpenAI to release new solutions to over 100 unsolved math problems
