The Omniscient Bulletin · October 8, 2026
OpenAI brings GPT-6 to every ChatGPT tier as Anthropic cuts what Haiku costs
OpenAI is putting GPT-6 in front of free ChatGPT users while treating the models as High capability in cybersecurity and in biological and chemical domains, and Anthropic cut what Haiku costs to run. The agent story cuts both ways: Windows now enforces the file and network limits organizations set for agents, while CrowdStrike says an agentic pentest tool was used against South Korean financial organizations.
Frontier
OpenAI is rolling out new GPT-6 models to free and paid ChatGPT users, and treats them as High capability in cybersecurity and in biological and chemical domains
OpenAI is launching GPT-6 to everyone in ChatGPT, on free and paid plans globally, where new October versions of GPT-6 Sol and GPT-6 Luna replace GPT-5.6 Sol and Luna. Under its Preparedness Framework, OpenAI treats them as High capability in cybersecurity and in biological and chemical domains, and says neither reaches its High threshold in AI self-improvement. Codex and ChatGPT Work users stay on previously released GPT-6 versions. OpenAI's announcement says paid tiers get it first, with Free and Go a day later, alongside a feature called Intelligent UI.
Frontier
Anthropic launched Claude Haiku 5.5, its first Haiku model with an adjustable effort setting, and says it costs around 75% less to run than Haiku 4.5
Anthropic launched Claude Haiku 5.5, its first Haiku-class model with an adjustable effort setting, and says it costs around 75% less to run on average than Haiku 4.5. Its per-token price is 90% below Haiku 4.5's on requests of up to 100,000 tokens, and 50% below it on requests over 100,000 tokens. Anthropic is also halving the price of Claude Sonnet 5.5 cache reads, which it says makes Sonnet about 20% cheaper on most agentic work, and this week will add monthly API credits of $100 or $200 for Max users and up to $500 for Team.
Industry
Samsung guided to about 107.4 trillion won of third-quarter operating profit, against 12.17 trillion won a year earlier
Samsung Electronics guided to consolidated operating profit of approximately 107.40 trillion won for the third quarter of 2026, against 12.17 trillion won a year earlier and 89.49 trillion won in the second quarter. It put sales at approximately 195 trillion won, up from 86.06 trillion won a year earlier. The figures are the medians of estimate ranges, because Korean disclosure rules do not allow guidance to be offered as a range. The release gives no breakdown by business.
Industry
Broadcom is working to arrange more than $50 billion in financing for the custom AI chip it is developing with OpenAI, the Wall Street Journal reported
Broadcom has spent recent weeks trying to line up more than $50 billion in financing for the custom AI chip it is building with OpenAI, the Wall Street Journal reported, citing people familiar with the discussions. Apollo and Blackstone are among the lenders it has approached, and the talks are at an early stage, so the deal size could change. Oracle, separately, is discussing with Apollo and Goldman Sachs how to pay for a big chip order, probably via a new entity that would buy the chips and lease them to Oracle.
Agents
Microsoft made its agent containment layer generally available on Windows 11, and says Copilot will be able to use local context from a PC with permission
Microsoft made Microsoft Execution Containers generally available on Windows 11, letting organizations set which files and networks AI agents can reach, with those policies enforced at runtime. It says GitHub Copilot, OpenClaw and OpenAI's Codex already support it, and that Anthropic's Claude Code and Nous Research's Hermes Agent are among those that will. Microsoft also said Copilot will be able to draw on local context from a user's PC with permission, in features it expects to start reaching Copilot+ PCs in the coming months.
Agents
CrowdStrike says an actor likely to be a Chinese speaker used ARTEX, an open-source agentic pentest tool, with language models against South Korean financial organizations
CrowdStrike says attacks on several South Korean financial organizations from late September to early October used ARTEX, a newly released open-source agentic penetration testing tool developed in China, alongside language models. It assesses with moderate confidence that the unattributed actor probably speaks Chinese and is after money. Directories the actor controlled held Claude Code session histories and Claude memory files, and the ARTEX instance ran on DeepSeek v4.1-flash. The number of organizations hit remains unconfirmed.
Policy
Treasury announced its first civil penalty under the outbound investment program, $200,000 against Amidi over a stake in a Chinese embodied AI company
The Treasury Department said it issued its first civil penalty under the Outbound Investment Security Program in July, fining Amidi, LLC $200,000 for failing to notify Treasury of an investment by its controlled foreign entity in a Chinese embodied artificial intelligence company. Amidi's subsidiary, a Chinese fund, put approximately $92,478 into Shanghai Qiongche Intelligent Technology, known as Noematrix. Amidi is the parent company of the business known as Plug and Play Tech Center.