Mirendil, founded by former Anthropic researchers, has raised $1 billion in Series A funding to develop tools that accelerate AI R&D. Google has launched Gemini 4 Hydrogen, saying it completes tasks 2.5 times as fast as Gemini 3.8 Flash, with input and output priced at $1.50 and $7.50 per million tokens, respectively. Thinking Machines has released the weights for Inkling 2 Small, which also supports a million-token context window, but running the main version locally requires at least 500GB of total GPU memory.
According to Techmeme, citing The Wall Street Journal, Mirendil has raised $1 billion in Series A funding led by Kleiner Perkins and Clearlake Capital. Founded by former Anthropic researchers, the company aims to develop tools that accelerate AI R&D.
Read original →Million-token context windows are reaching both commercial APIs and open-weight models, making deployment requirements and weight auditability separate factors in model selection.
Google has launched Gemini 4 Hydrogen with a million-token context window across Search, the Gemini app and the API. Google says it completes tasks 2.5 times as fast as Gemini 3.8 Flash. The accompanying Orchestrator can automatically switch models within the same conversation. Input and output cost $1.50 and $7.50 per million tokens, respectively, with uniform pricing across the entire context window.
Read original →Thinking Machines has released the weights for Inkling 2 Small, which accepts text and image inputs, produces text output and supports a context window of up to one million tokens. Its mixture-of-experts architecture has 594 billion total parameters, with only 17 billion activated at a time. The weights can be used for research, fine-tuning and integration into third-party products. Local deployment already supports frameworks including vLLM and SGLang, but the main version requires at least 500GB of total GPU memory.
Read original →Will DePue has published a weight-forensics study claiming that six models from four companies, including DeepSeek and Moonshot AI, carry fingerprints of Blackwell chips. The method recovers traces of initialization from publicly available weights. The author says it matched all 24 models whose training hardware had been disclosed. However, the fingerprints identify the hardware used to create the weights; they do not prove that subsequent training used the same chips, much less establish an export-control violation.
Read original →🔍 Analysis: Gemini 4 Hydrogen and Inkling 2 Small both offer million-token context windows, but the former uses uniform pricing across the entire window, while the main version of the latter requires at least 500GB of GPU memory. Businesses are therefore comparing API costs with self-hosting resources, not just context length. Open weights also make models a subject of hardware forensics: Will DePue's Blackwell fingerprinting research extends transparency questions from model behavior to the creation of the weights. For businesses, being deployable, auditable and usable as compliance evidence are three different things. Weight fingerprints alone cannot establish what hardware was used for training or whether a violation occurred.
Developers can delegate tool orchestration to a platform or keep agent control software and session data within their own enterprise environment.
DeepSeek has enabled web search, deep-thinking search, file input, link reading and safety checks for V4.1-Flash. Tool calls and context assembly are handled server-side. Developers can also call OCR, speech recognition and speech synthesis APIs independently and use an open-source chat frontend. These extensions incur tool fees in addition to token charges. Public-facing applications must enable safety checks, and developers remain responsible for content moderation and other obligations.
Read original →Anthropic has opened a self-hosted private beta for Claude Managed Agents, allowing businesses to run the agent service in their own cloud or on-premises. Previously, only the sandbox could be self-hosted; now the software controlling agent execution, credentials and session history can also stay within the enterprise environment, while retaining the managed version's core API. Model requests can go directly to Anthropic or through a cloud provider or an enterprise's own gateway. Access to the private beta requires contacting the account team.
Read original →🔍 Analysis: DeepSeek moves tool calls and context assembly to the server, while the self-hosted private beta of Claude Managed Agents puts control software, credentials and session history back in the enterprise environment. The two approaches draw different hosting boundaries at the runtime layer. For developers, ease of integration and operational control need to be assessed separately: DeepSeek's managed tools do not transfer content-moderation responsibility to DeepSeek, and self-hosting Claude's runtime does not mean model inference must run locally. Model selection therefore involves more than API compatibility; it also requires checking where data stays, who orchestrates calls and who bears responsibility.
Shopping, code fixes and live meeting collaboration are being delegated to agents, with approval gates controlled by users, engineers and organization administrators, respectively.
OpenAI has launched ChatGPT Wallet, allowing the dots agent to shop on users' behalf, with every transaction requiring user approval through a security check. OpenAI says the agent never accesses actual card details, instead using payment credentials valid for a single transaction, while users retain their existing credit card rewards and protections. Initial access is limited to ChatGPT Pro users over 18 in the United States, with no transaction commission charged at launch.
Read original →On November 12, OpenAI launched Daybreak Red Team and Daybreak Defense Factory for web application vulnerability testing and automated fixes, respectively. Red Team is available in early access and can also validate externally reported vulnerabilities and test patches. Defense Factory scans codebases, validates vulnerabilities and generates patch pull requests with tests. Fixes still require engineer review and approval before entering production.
Read original →Anthropic has launched a beta of Claude Meetings for Team and Enterprise users, initially supporting Google Meet. Beyond post-meeting summaries, it joins as a visible participant and creates documents and whiteboards in real time, allowing teams to edit them as they discuss. Notes, action items and transcripts are saved after the meeting. By default, it does not join meetings with guests from outside the organization, though administrators can allow this. The system stores transcripts, not audio.
Read original →🔍 Analysis: ChatGPT Wallet places an approval gate before every transaction, Daybreak requires engineer review before fixes enter production, and Claude Meetings limits information sharing by defaulting to not joining meetings with guests from outside the organization. These designs govern spending, code changes and the flow of meeting information, respectively; a single blanket switch to “allow agents to act” cannot replace them. For businesses, deployment requires authorization checkpoints tailored to the consequences, rather than treating the ability to complete a task as permission to carry out the entire workflow autonomously.