New models and public filings alongside infrastructure and agent research
BunCosmos Research Lab | Research checked 26 October 2026
Developments reviewed on 26 October include a new OpenAI model release, Anthropic’s proposed public offering, and changes to how AI companies fund and distribute their products. SpaceX has announced a binding commitment for a major GPU lease, while robotics production and research into agent memory and security show the practical systems taking shape around AI. The event dates below distinguish new announcements from earlier research and reporting.
At a glance
Area | Development |
|---|---|
Model release | OpenAI releases GPT 6.1 Terra for ChatGPT Work, Codex and the API. |
Public markets | Anthropic publicly files for a proposed IPO; price and share count remain unset. |
AI funding | Meta launches planned $35 billion offerings and outlines a later $60 billion ATM programme. |
Nonprofit tools | Claude adds nonprofit-specific skills and connectors for Team and Enterprise nonprofits. |
AI infrastructure | SpaceX announces a binding $40 billion GPU lease commitment, subject to closing conditions. |
Humanoid fleets | 110 A3 robots complete a synchronized Guinness-certified dance challenge. |
Consumer robotics | Qiyuan reports its 5000th personal robot leaving production. |
AI chips | Cambricon reports Q3 revenue and profit more than doubling. |
Agent research | Small memory study and vendor security report highlight distinct operational trade-offs. |
OpenAI releases GPT 6 1 Terra
OpenAI announced GPT-6.1 Terra on 26 October, with rollout starting in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and API access through gpt-6.1-terra. The company positions it for visual reasoning, computer use and writing. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. [12]
This is an official product release, rather than the earlier appearance of the model’s name. OpenAI’s claims about speed and price-performance are company-reported comparisons; practical value will depend on the workflow, tool permissions and quality required.
Anthropic publicly files for a proposed initial public offering
On 26 October, Anthropic announced that it had filed a registration statement for a proposed initial public offering of Class A common stock. It intends to list on the Nasdaq Global Select Market under the ticker ANT. The number of shares and price range have not been determined, and the registration statement is not yet effective. [13]
The new development is the public filing, not the earlier confidential submission. It does not mean that Anthropic has completed an IPO or raised a specified amount. For readers following AI’s economics, the filing marks a step toward public-market scrutiny, while leaving the transaction’s eventual size and terms unresolved.
Meta outlines large planned offerings to fund AI investment
Meta announced planned underwritten offerings totalling $35 billion on 26 October: $22.5 billion of Class A common stock through forward-sale agreements and $12.5 billion of depositary shares representing mandatory-convertible preferred stock. It also outlined a separate Class C at-the-market programme of up to $60 billion, not expected to begin before the first quarter of 2027 and subject to the start of Class C trading. [14]
The plans support investment in AI computing and infrastructure. The combined headline figure of up to $95 billion describes proposed fundraising capacity, not money already raised. Timing, execution and market conditions still matter.
Claude expands practical tools for nonprofit organisations
Anthropic announced an expansion of Claude for Nonprofits on 26 October, including plugin skills and connectors developed with around 30 nonprofit organisations. The tools support work such as fundraising research, proposals, reporting and operations. They are installed automatically for nonprofit customers on Team and Enterprise at no additional cost. [15]
Anthropic also plans to extend ClaudeCorps through Taproot pro-bono consultants for shorter projects. That support is a planned expansion, not a completed rollout. The development is the new workflow tooling and support model, rather than a newly invented nonprofit discount programme.
SpaceX secures a binding commitment for GPU financing
On 26 October, SpaceX announced a binding commitment to enter into an approximately $40 billion master finance lease for GPUs and related equipment with an entity managed by Apollo. The company said the equipment would include NVIDIA Vera Rubin chips for its terrestrial computing infrastructure. This updates the previous day’s account based on Elon Musk’s statement that financing had closed. [1]
Entry into the lease remains subject to specified closing conditions and is expected before the end of 2026. Lease schedules are expected to begin when equipment arrives in the first half of 2027. The financing commitments are led by Apollo and Goldman Sachs, with anchor commitments from funds and accounts managed by BlackRock, Goldman Sachs Asset Management and capital accounts advised by KKR. A committed transaction is not the same as a completed closing. [1]
The broader point is not whether every data-centre company follows SpaceX’s path. It is the scale of capital now required to compete for advanced chips and compute capacity. Better models increasingly depend on very expensive physical infrastructure: processors, networking, electricity, cooling, buildings and financing.
Zhiyuan puts 110 humanoids under one fleet-control system
On 24 October in Jinan, Zhiyuan Robotics used 110 full-size Yuanzheng A3 humanoid robots to set a Guinness World Record for the most humanoid robots dancing simultaneously. Chinese reporting says a Guinness adjudicator witnessed the challenge and certified the result. [2][3]
The demonstration used Zhiyuan’s LinkBots fleet-control platform and centimetre-level UWB positioning to coordinate synchronized movement and formation changes. Zhiyuan said the robots were production models rather than one-off machines built only for the record attempt. [2][3]
A synchronized dance does not prove that 110 humanoids can work autonomously in an unpredictable factory or public environment. But the experiment tests something commercially important: whether many physical robots can be manufactured consistently enough to receive coordinated commands, maintain position and change formation together. Robotics is slowly shifting from “what can one impressive robot do?” toward “can fleets of ordinary production robots operate reliably together?”
Qiyuan says its 5,000th personal robot has left the production line
Qiyuan Robotics, a consumer-robot brand controlled by Shangwei New Material, announced that its 5,000th personal robot rolled off the production line on 25 October. The company says it reached that number about one month after the Q1 and T1 models went on sale on 20 September. [4][5]
Qiyuan says production has increased from a little over ten units per day to more than 200, with one robot leaving the line roughly every 2.5 minutes. A Shanghai store visit reported by Chinese financial media also found most models sold out, with some new orders pushed into December. That store-level observation should not be treated as a complete national demand measurement, but it is a useful early commercial signal. [4][6]
The products are still far from general-purpose humanoids capable of doing every household task. Their early use cases include language practice, programming education, entertainment, following users and basic assistance. The more important development is manufacturing: consumer embodied-AI products are beginning to produce real volume, which will generate the usage data and reliability problems that laboratory prototypes cannot reveal.
Cambricon’s third-quarter numbers show how quickly China’s AI-chip market is scaling
Chinese AI-chip company Cambricon reported third-quarter revenue of 3.531 billion yuan, up 104.47% from a year earlier, and net profit attributable to shareholders of 1.404 billion yuan, up 147.74%. Operating cash flow for the quarter was slightly negative at about 83.7 million yuan. [7]
The results do not tell us that Cambricon has matched Nvidia technologically. Financial growth measures sales and business momentum, not model-training performance or chip efficiency. But the numbers are strong evidence that demand for domestic AI computing hardware is translating into large commercial revenue inside China.
Cambricon also scheduled a third-quarter investor briefing for 26 October. The useful question for the next stage is not simply whether Chinese chip companies can sell more hardware, but how their systems perform at scale in training and inference once customers move beyond procurement targets and begin comparing reliability, software support and total cost of ownership. [8]
A small study asks whether strong AI agents need complicated memory systems at all
The new preprint MemoryTax: The Bitter Lesson of Agent Memory examines a practical question for long-running AI agents: should important information be extracted into a structured memory system in advance, or should a strong model search the raw conversation history when it needs something? [9]
The study compared four memory approaches and three coding-agent systems using GPT-4o-mini, GPT-6 Luna and Claude Opus 5.5 on one 419-turn conversation containing 152 questions. The author reports that structured systems such as Mem0 were most useful for the weakest model, while an agent using Claude Opus 5.5 with direct search tools reached 99% accuracy at less than half the cost of feeding the entire conversation as context. [9]
This is an interesting result, not a final verdict on agent memory. The experiment used only one conversation and one run per configuration, and some of the apparent advantage came from image-caption access that was not identical across systems. The paper is valuable because it identifies a real trade-off: every memory-writing system decides what to keep before it knows which future question will matter. If the writer leaves something out, a stronger model reading that memory later cannot recover information that was never stored.
Ghost agents may remain dangerous after people stop using them
Arrakis Security’s 2026 Autonomous Agents Threat Report raises a less glamorous but very practical agent-security problem: abandoned software can keep its permissions. The company says that, across 12 weeks of anonymized telemetry from thousands of agents in customer environments, 76% of enabled agents had not run in 60 days and 44% had not run in 180 days, while their credentials and triggers remained active. [10][11]
Arrakis also reports that 24% of dormant agents could be started by an unsigned external webhook, 70% of all measured agents could reach sensitive systems, and 82% of agents carried at least one exposure. These figures come from a security vendor analysing its own customer telemetry, not from an independent population-wide audit, so they should not be generalized to every enterprise agent deployment. [10][11]
Even with that limitation, the underlying security lesson is straightforward. Organisations already know how to disable employee accounts when someone leaves. They need an equivalent lifecycle for agents: an owner, an expiry date, restricted credentials, logs, a decommissioning process and a regular check that old automations cannot still be triggered from the outside.
What these developments establish
New releases, public-market plans, chip financing, robot production and agent research all expose different constraints on AI adoption. Better models matter, but so do access, cost, reliable operation and the ability to govern systems after deployment.
A model launch is not independent proof of performance, a proposed offering is not completed financing, and a robot demonstration is not a field trial. The useful question is what each development establishes now—and what still depends on future delivery, testing or adoption.
Sources
[1] SpaceX. GPU financing announcement. 26 October 2026. Source
[2] 21st Century Business Herald / Eastmoney. Zhiyuan A3 humanoid robots achieve Guinness World Record. 26 October 2026. Source
[3] ITHome. Zhiyuan organizes 110 humanoid robots to dance simultaneously in Jinan. 24 October 2026. Source
[4] ITHome. Qiyuan Robotics’ 5,000th personal robot leaves production line. 25 October 2026. Source
[5] Securities Daily / China Enterprise Report. Qiyuan Robotics scales to 5,000 personal robots in one month. 26 October 2026. Source
[6] Caiwen / Sina Finance. Daily capacity rises above 200 units; Shanghai store reports limited stock. 25 October 2026. Source — Store-level observation, not a nationwide sell-through audit.
[7] ITHome. Cambricon Q3 2026 results: revenue and net profit more than double. 23 October 2026. Source
[8] Cambricon / CNINFO. Cambricon notice of Q3 2026 investor briefing. 17 October 2026. Source
[9] arXiv. MemoryTax: The Bitter Lesson of Agent Memory. 21 October 2026. Source — Preprint; one conversation and one run per experimental cell.
[10] Arrakis Security. 2026 Autonomous Agents Threat Report. October 2026. Source — Vendor research based on its customer telemetry.
[11] GlobeNewswire / Arrakis Security. Arrakis Security report finds 76% of AI agents dormant but credentials remain active. 22 October 2026. Source
[12] OpenAI. Introducing GPT-6.1 Terra. 26 October 2026. Source
[13] Anthropic. S-1 registration statement announcement. 26 October 2026. Source
[14] Meta. Offering announcement distributed by PR Newswire. 26 October 2026. Source
[15] Claude. Expanding Claude for Nonprofits. 26 October 2026. Source
Evidence note. Company announcements and vendor reports are identified as such. Research preprints are not treated as peer-reviewed conclusions. Where a source reports an observation from one store, one benchmark, one customer population or one experimental setup, the article keeps that limitation attached to the claim.