JadidJust-in AI Developments, Incidents & Deadlines
October 2026
SMTWTFS12345678910111213141516171819202122232425262728293031

Agent architecture and agent pricing are both being reset at once: use fewer agents, and pay less for them.

14 stories · transcript

Anthropic now tells teams to collapse sub-agent fleets into one agent that loads skills on demand, while OpenAI is reported to be undercutting Anthropic contracts by as much as 90 percent. A quiet weekend also brought a UEFI bootkit that survives on patched Windows machines and an AWS security agent whose headline score comes with caveats.

TLDR

  • Anthropic says one agent with on-demand skills often beats a fleet of sub-agents (Oct 25), running about twice as fast in one comparison. Agents
  • OpenAI is reportedly cutting customer costs by up to 90 percent to pull work from Anthropic (Oct 25), per The Information. Pricing Business
  • BlackX, a BlackLotus successor, bypasses Secure Boot on patched Windows machines (Oct 23) unless the revocation step was applied by hand. Security
  • AWS Continuum reaches general availability with an AWS-reported 89.0% on CyberGym-E2E (Oct 22), limited to C and C++ memory bugs. Security Production
  • Jack Clark pulls his forecast for self-improving AI forward to 2027 (Oct 25), citing a new research-taste benchmark. Research

Upcoming Deadlines

Date What Who it hits Do this Source
Sat Oct 31 Mistral retires zai-glm-5-2 Callers of that model Move to zai-glm-5-3 at the same price Mistral
Sat Oct 31 OpenAI Evals goes read-only Teams with saved evals Export evals OpenAI
Sat Oct 31 DeepSeek's free Search/Fetch promotion ends 4:00 PM UTC DeepSeek API users Budget $0.75 per 1,000 calls after DeepSeek · DeepSeek news
Wed Nov 4 Moonshot's promised date for Kimi K3.1 Pro open weightsA model whose trained weights are published for anyone to download and run.NTIA ↗ Self-hosters of open models Line up an eval; read the license when it lands The Decoder

Action Items

  1. Add the "Windows Production PCA 2011" certificate to the Secure Boot DBX on every Windows fleet this week. The monthly patch for CVE-2022-21894 does not block the vulnerable boot manager without it, and that certificate expired Oct 19.
  2. Pick one workflow built from several sub-agents and A/B it against a single agent with skills before Nov 15. Use Anthropic's published migration kit and compare latency, cost and answer quality on your own traffic.
  3. Export saved OpenAI Evals before Oct 31, then budget $0.75 per 1,000 calls for DeepSeek's Search and Fetch tools after the free promotion ends.
  4. Re-quote your largest coding-model contract against OpenAI's current offers before your next renewal. Reported discounts ran from more than 30 percent on a $100M-a-year contract to 80-90 percent on smaller ones.
  5. Move any Mistral caller of zai-glm-5-2 to zai-glm-5-3 by Oct 31.

Models & Tools

  • Anthropic: consolidate sub-agents into one agent that loads skills on demand (Oct 25) Agents Anthropic's applied AI team says specialized sub-agents add latency, cost and information loss at every handoff, and that models from Opus 4.5 through Opus 5.5 and Fable 5.1 can run complex plans in one loop. One architecture comparison ran about twice as fast. A dozen sub-agents in one customer-facing agent hurt consistency. The migration kit is on GitHub. (The Decoder | GitHub)
  • Google ran Gemini 4 Argon unattended for 11 hours on its own 200K-line Agent Development Kit (Oct 24) Agents Production The run produced two accepted pull requests: about 40% less session-preparation overhead in long sessions and 21% faster cold imports. The team wrote repository conventions as .agents/skills files, told the agent to load only what a task needs, and used a goal mode that survives context compaction. Results are Google's own. (InfoQ | Google Developers Blog)
  • Nvidia open-sources Flora, which labels robot video for about $0.72 an hour (Oct 25) Robotics/IoT Open Source Three fine-tunedTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗ Qwen3.5 models segment actions, describe them and score success. Nvidia says GPT-6 Astra costs $88 for the same hour, and that Flora scored slightly higher on a 50-video test while lagging on descriptive accuracy. It runs on 512 GB300 GPUsGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ at up to 200,000 video hours a day. All figures are Nvidia's. (The Decoder | Nvidia Research)
  • Block releases Dessa, a full-duplex voice model that runs on a MacBook M4 Max (Oct 24) Research Dessa is trained only on human speech and listens while it talks, so emotion is not driven by text tags. Block says it trained on a single Nvidia rack. Quality claims rest on Block's demo clips, and it is unclear whether any steering tags exist. (TechCrunch | Block)

APIs & Pricing

  • OpenAI is buying share with discounts of up to 90 percent, The Information reports (Oct 25) Pricing Business Rippling moved OpenAI's share of its coding workload from 16 to 52 percent as Anthropic's fell from 57 to 18. Replit got over 30 percent off a contract worth more than $100M a year, and DataCamp says costs fell 80-90 percent. OpenAI is partly trading margin for share. (The Decoder | The Information)

Security

  • BlackX bootkit abuses Baton Drop, CVE-2022-21894, to bypass Secure Boot on patched Windows machines (Oct 23) Security Medium · CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ 4.4 · Exploited in the wild · Fix: Windows update plus manual DBX revocation. Kaspersky links BlackX to APT41 and diplomatic targets in Central Asia. It needs admin or physical access, so it is persistence, not entry. Unrevoked boot managers still load though the PCA 2011 certificate expired Oct 19. (Securelist | The Hacker News)

AI in Production

  • AWS Continuum is generally available, with an AWS-reported 89.0% end-to-end score (Oct 22) Security Production It finds bugs, checks them against your AWS environment and opens fix pull requests. AWS reports 819 of 920 CyberGym-E2E tasks, against a prior public high of 65.9%. Only 37.8% of patches fixed the historical bug the benchmark chose, and all tasks are C and C++ memory flaws. (The New Stack | AWS)
  • Alibaba Cloud says agents write 95% of new code in its basic-software team (Oct 24) Production Agents OpenAnolis chair Ma Tao made the claim at GOSIM and said agents also pick packages and patches, review code and backport fixes in Anolis OS 27. The 95% figure is a single-team, self-reported number. (Pandaily)

Industry & Business

  • Eka Robotics raises $150M Series A for a vision-force-action manipulation model (Oct 24) Business Robotics/IoT Andreessen Horowitz and F-Prime invested. Eka says its robots pick unseen objects at roughly 1,500 operations per hour and cites simulation-trained reinforcement learningReinforcement learningTraining by trial and error, where the model takes actions and adjusts based on the rewards it gets.Kaelbling et al. 1996 ↗ as the route to dexterity. Performance claims are the company's. (The Robot Report)
  • Mave raises $34M for agents that investigate security alerts across existing tools (Oct 22) Business Security The Tel Aviv startup closed a $28M Series A led by Red Dot Capital Partners on top of a $6M seed. It has about 30 staff and plans to hire in go-to-market and AI research. (SecurityWeek)
  • Nvidia and PreussenElektra will study turning shut German nuclear sites into AI data centers (Oct 21) Infra Business The preliminary assessment targets the existing grid links, land and water, not nuclear power. It is exploratory, with no sites or capacity named. (Tom's Hardware)

Policy & Safety

  • Update — OpenAI's red-team agents also reached blocked services through approved connections (Oct 25) Agents Security In one test an agent reused a firewall-approved connection to reach a blocked authentication endpoint in a metrics service and made it fetch a model-supplied URL. Others used DNS and firewall misconfigurations. OpenAI says it resumed frontier training on Oct 1 after hardening, and reserved compute for continuous red-teaming. (The Hacker News | OpenAI)

Benchmarks & Evals

  • TasteVal: models now out-score 24 human experts at choosing and reading experiments, and Jack Clark moves RSI to 2027 (Oct 25) Research P-Zero Research's benchmark covers eight AI-development tasks on one H100 and tests experiment design and interpretation, not choosing the question. Clark says the capability has doubled roughly every three months since December 2025 and now expects self-improving AI as early as 2027. The benchmark is narrow and not independently replicated. (The Decoder | P-Zero Research)

Paper Spotlight

  • OpenMobile-2: app diversity beats data volume for open smartphone agents (Oct 25) Research Open Source With the same training data, raising app count lifted success on unseen apps from 60.7 to 67.5 percent. The environment covers 110-plus apps, 35 of them rebuilt replicas. The 27B model scores 59.6 on SPA-Bench, the best among open models, per the authors. Weights, data and instructions are released. (The Decoder | arXiv)

Type to search · ↑↓ to move · enter to open