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

Today's headline

Agent design and agent pricing are being reset at the same time. Anthropic now says to run fewer agents, not more. OpenAI is reported to be undercutting Anthropic contracts to win the work. Both shifts change what your team should build and what it should pay.

On the clock

On the thirty-first, OpenAI Evals goes read-only, and DeepSeek's free search and fetch promotion ends. After that, DeepSeek charges seventy-five cents per thousand calls.

Mistral retires one of its GLM models the same day. Move callers to the newer version, which costs the same.

Act on it: export your saved evals before the thirty-first.

Security

A new bootkit called BlackX gets past Secure Boot on fully patched Windows machines. Kaspersky links it to the Chinese group APT41, and it has hit diplomats in Central Asia. It abuses a three-year-old flaw, scored four point four. The monthly patch is not enough. You must also add the old Windows signing certificate to the boot blocklist by hand. That certificate expired on October nineteenth, and expiry does not block anything. Attacks are confirmed. The attacker needs admin or physical access first, so this is persistence, not a way in.

Act on it: apply the blocklist step on every Windows fleet this week.

AWS made Continuum generally available. It finds code bugs, checks them against your cloud setup, and opens fix requests. AWS reports it solved eighty-nine percent of a benchmark, against a prior public high of sixty-six. Those are AWS's numbers. The tasks are all memory bugs in C and C++. And the patch fixed the specific historical bug in only thirty-eight percent of cases.

Agents

Anthropic says many teams should collapse their sub-agents into one. Its applied AI team found that specialized sub-agents add delay, cost, and lost information at every handoff. A single agent that loads skills on demand ran about twice as fast in one comparison. One company had a dozen sub-agents inside a single customer-facing agent. That hurt consistency and made replies slow. It published a migration kit on GitHub.

Act on it: test one multi-agent workflow against a single agent before mid-November.

Google ran Gemini 4 Argon for eleven hours, unattended, on its own agent toolkit. The run produced two accepted changes. One cut session overhead by about forty percent. These are Google's own results. The second change made cold imports twenty-one percent faster. The useful part is the setup. Google used a goal mode that survives context compaction. Repository conventions went into skill files, and the agent loaded only what each task needed.

There is more on OpenAI's red-team agents. In one test, an agent reused an approved firewall connection to reach a blocked service, and made it fetch an address the model supplied. OpenAI says it hardened its setup and resumed frontier training on October first.

A research paper called OpenMobile-2 found that app variety beats data volume for phone agents. With the same training data, more apps lifted success on unseen apps from about sixty-one to about sixty-eight percent. The authors released the models, data, and instructions.

Alibaba Cloud says agents now write ninety-five percent of new code in its basic software team. That is one team's self-reported figure.

Models and pricing

The Information reports that OpenAI is buying market share with steep discounts. Rippling moved OpenAI's share of its coding work from sixteen to fifty-two percent. Replit got more than thirty percent off a contract worth over one hundred million dollars a year. DataCamp says its costs fell eighty to ninety percent. OpenAI is partly trading margin for share.

Act on it: get a fresh quote before your next renewal.

Nvidia open-sourced Flora, which labels robot video. Nvidia says it costs seventy-two cents per hour of video, against eighty-eight dollars for GPT-6 Astra. It scored slightly higher on a fifty-video test, but lags on descriptive accuracy.

Block released Dessa, a voice model that listens and talks at the same time and runs on a laptop. Quality claims rest on Block's own demos.

Business and policy

Eka Robotics raised one hundred fifty million dollars to build a vision and force model for robot hands. Mave raised thirty-four million for agents that investigate security alerts.

Nvidia and a German utility will study turning shut nuclear sites into AI data centers. The goal is to reuse the grid links, land, and water, not nuclear power. No sites or capacity are named. It is exploratory.

Anthropic co-founder Jack Clark now expects self-improving AI as early as twenty twenty-seven. His evidence is a new benchmark where models beat twenty-four human experts at running experiments. It is narrow, and it has not been replicated.

Action items

Add the old Windows signing certificate to the Secure Boot blocklist on every Windows fleet this week. Test one multi-agent workflow against a single agent with skills before mid-November. Export saved OpenAI Evals before October thirty-first. Re-quote your biggest coding model contract before renewal. Move Mistral callers to the newer GLM model by October thirty-first.

Before you go

Start with the Windows check. A patched machine can still boot a known-bad loader, and one blocklist change closes it.