SoftBank has transferred the final $10 billion of its OpenAI investment and now holds a stake of about 13%; to fund the payment, SoftBank issued $11.1 billion in junk bonds. Hugging Face found that a model with identical weights scored 62% in one agent framework and 33% in another; multi-framework training raised LFM2.5-2.6B’s score across four frameworks from 42% to 54%. With the execution model and budget unchanged, a Meta controller raised GPT-5.5’s ProgramBench score from 63.7% to 71.5%. The OpenAI Agents API now lets developers launch a browser and an agent with one call; the platform reports 99.97% single-turn reliability.
SoftBank has made its final payment, securing a stake of about 13% through an investment funded by both asset sales and borrowing.
SoftBank has transferred the final $10 billion installment of its OpenAI investment and now holds about 13% of the company. According to the poster, Masayoshi Son sold his entire Nvidia stake to raise the first payment; SoftBank issued $11.1 billion in junk bonds to fund the last. The two payments were financed through an asset sale and borrowing, respectively.
Read original →🔍 Analysis: SoftBank sold its Nvidia holdings to fund one payment and issued junk bonds to fund another. The investment’s risk therefore depends not only on the value of its OpenAI stake but also on the debt SoftBank took on. Assessing an AI-model investment of this size requires looking at how the money was raised, not just how much was invested.
Framework compatibility, task scheduling and how an agent handles conversation history all affect task performance.
Hugging Face has published a guide to multi-framework reinforcement learning: a model with identical weights scored 62% and 33% in two different agent frameworks. Its approach records training data through proxies compatible with OpenAI, Anthropic and Gemini API formats, without requiring changes to the frameworks. LFM2.5-2.6B improved from 42% to 54% across four frameworks while making 31% fewer tool calls; the trainer, data and seven post-trained models are all available.
Read original →Meta Superintelligence Labs reports that a dedicated controller raised GPT-5.5’s ProgramBench score from 63.7% to 71.5%. The controller decides which tasks to run next; the model executing those tasks and the budget remained unchanged in the test, isolating the effect of the scheduling strategy. For comparison, Codex scored 58.0%.
Read original →Microsoft has proposed FOCUS, a training-free method for compressing agent context at test time. It reports reducing peak context by up to 48% and improving task success rates by up to 8.9 percentage points. The method identifies which past interactions matter for the next decision and retains only the relevant history. It addresses the possibility that agent performance deteriorates as interaction history grows.
Read original →🔍 Analysis: After finding that a model with identical weights scored 62% in one framework and 33% in another, Hugging Face used multi-framework training to improve compatibility. Meta, meanwhile, raised GPT-5.5’s ProgramBench score from 63.7% to 71.5% by changing only the scheduling, leaving the execution model and budget unchanged. Microsoft’s FOCUS addresses another source of runtime overhead: it cuts peak context by up to 48% while improving task success rates by up to 8.9 percentage points. When comparing agent systems, developers should test frameworks, scheduling and context management separately rather than attributing every score difference to the underlying model.
OpenAI is simplifying how browser agents are launched, while T3 Code is expanding task delegation across models.
OpenAI’s Agents API added computer-use capabilities this week, allowing a browser and agent to be launched with one API call. The update also includes Bedrock Managed Agents and an OpenAI-hosted environment with a choice of lightweight or higher-spec configurations. The platform reports 99.97% single-turn reliability and a 20% increase in tool-call speed.
Read original →T3 Code has merged a four-month orchestrator rewrite spanning 823 commits and 1,912 files; the project now has more than 400,000 users. The new delegate_task lets agents launch subagents across providers or models. The update also adds Pi support, an ACP registry, thread forking, mid-conversation model switching, a subagent lineage view and scheduled tasks.
Read original →🔍 Analysis: The OpenAI Agents API combines browser and agent launches into a single call, lowering the barrier to integration. T3 Code’s delegate_task and subagent lineage view address a different challenge: coordinating work across models after tasks are delegated and tracking what happens. Teams running complex tasks need to test both how easy the system is to invoke and how visible its execution is.
Claude is expanding its business-software integrations, while Muse Gadgets is opening up device development.
Anthropic has added 27 integrations to Claude for Small Business, which now offers 43 workflows covering tools including Shopify, Salesforce and Stripe. The product has been installed more than 900,000 times since its May launch. The new workflows extend beyond back-office tasks to finding customers, responding to inquiries and drafting proposals, addressing growth needs raised by about one-third of business owners at the spring roadshow. They run in the Claude Cowork desktop app, which waits for user approval by default before sending, publishing or making payments.
Read original →Meta has open-sourced Muse Gadgets’ ESP32 firmware and Linux device SDK. Developers can use off-the-shelf development boards or a Raspberry Pi to connect displays, buttons and sensors to Muse. The Raspberry Pi setup can also connect to Home Assistant. Meta has also made 5,000 Muse Home Link smart-home bridge devices, available free in limited quantities to subscribers. The SDK and firmware are licensed under Apache 2.0 and provided as-is, without a warranty.
Read original →🔍 Analysis: Claude’s 27 new integrations connect to existing tools such as Shopify, Salesforce and Stripe. Muse Gadgets’ SDK and firmware let developers connect displays, buttons and sensors to Muse. The former brings AI into established business processes and waits for approval by default before sending, publishing or paying; the latter lets developers build new device form factors. The choice depends on whether the goal is to handle existing work or design a new way to interact.
Meta conducted the research through a standard chat interface rather than a custom agent framework.
Meta has published six papers on open mathematics problems using Muse Spark 1.1 and 1.2. The research was conducted through ordinary meta.ai chats, without a custom agent framework. Meta’s AI models had previously reached gold-medal level in five competitions across mathematics, physics and chemistry. This work shifts to open problems with no established solution path, exploring whether AI can contribute to such research.
Read original →