Claude streamlines long conversations; LongCat 2.5 launches
Claude is reducing repeated uploads in long conversations, LongCat 2.5 has launched, and Hunyuan Hyra 2.0 is turning research into reusable models and tools. Today’s 12 stories also cover configurable audio questions, frontier-model evaluation, and new plans for electricity and optical connectivity.
01
Models and Developer Tools
4 stories
2026-11-02Anthropic
Claude message threads reduce repeated conversation uploads
Anthropic has launched message threads in public beta on Claude Platform. Developers send the full conversation once, then provide only new messages and the previous response ID. That cuts the overhead of repeatedly uploading long histories. Anthropic reports roughly 35% lower median time to first token for Claude Code requests containing 500,000 to one million tokens.
Threads are encrypted in memory, never written to disk, and expire five minutes after the last request. Clients must resend the history after expiry. The model still reads the full context, so token usage and standard pricing are unchanged. Availability includes the native platform and corresponding Claude services on AWS, Microsoft Foundry, and Google Cloud.
2026-11-02Tencent Hunyuan
Hunyuan Hyra 2.0 turns research work into reusable tools
Tencent Hunyuan has introduced Hyra 2.0, extending its research-agent work from solving individual problems to developing and improving scientific models and tools. Tencent describes an Overall, Mission, and Worker hierarchy connected to WeCom, computing platforms, and code repositories, placing research tasks inside an existing development environment.
The public results repository added concrete releases on November 2: the HyInv inverse-folding model, the HyPLM-Flash protein language model, and the HyFF family of force fields for molecular simulation. The releases include weights, inference code, or technical reports. HyPLM-Flash also includes an INT8 inference engine, giving researchers reusable artifacts beyond a written answer.
2026-11-02Meituan
LongCat 2.5 launches for long-horizon agents and visual coding
Meituan has released LongCat 2.5, a natively multimodal model with 1.6 trillion total parameters. The company highlights visual understanding, agentic coding, and long-horizon tasks, and says training ran entirely on domestic computing infrastructure. This is the full release, distinct from the earlier LongCat 2.5 Preview.
The model is available through an API and a web interface. The platform changelog lists integrations with Claude Code, Codex, and Hermes, alongside visual coding, creative 3D, and end-to-end engineering use cases. The announcement describes million-token training sequences; that statement does not establish an identical context limit for every API endpoint.
2026-11-02Hume AI
Hume introduces configurable questions about audio
Hume has introduced Audio Decision API, which takes audio, a developer-defined question, and candidate answers, then returns a probability for each answer. Questions can concern a laugh, a perceived speaking style, or the location of a sound. Multiple questions can be asked about the same recording without training a separate classifier for each task.
Developers can add context, set confidence thresholds, and process an audio library in batches. An interactive playground is available, while API access is through a waitlist. For questions about emotion or delivery, the output estimates how a voice may be perceived; it is not a definitive measurement of the speaker’s internal psychological state.
02
Research and Security
3 stories
2026-11-02OpenAI
OpenAI monitors generation with activation classifiers
OpenAI describes activation classifiers that inspect internal model signals and score generation token by token. When a threshold is reached, the system can pause a response or block a tool call and send the case to a reasoning model for review. OpenAI says deployment began with GPT-5.6 across internal and external traffic for its most capable models.
In the reported experiments, these classifiers were more robust against the tested jailbreaks than the text classifiers used for comparison. They supplement existing text-based safeguards. Their coverage still depends on explicitly labeled, trained-for categories of unwanted behavior, and a new base model requires retraining; the results do not establish detection of every undefined risk.
2026-11-02SemiAnalysis
GDPx tests frontier models on 206 tasks across 11 domains
SemiAnalysis has launched GDPx, combining 206 tasks from 14 specialist data partners across 11 domains, including chip design, software engineering, biology, and law. At launch, the leaderboard placed Claude Opus 5.5 first with an aggregate score of 12.0%, followed by Gemini 4 Carbon and GPT-6 Astra.
The aggregate weights domains equally, combines partners within each domain, and averages three attempts per task. Some tasks use continuous rather than binary grading. The authors also note that safety refusals affect results in several domains. The 12% figure is therefore an aggregate score on this demanding test, not the share of all work a model can perform or a measure of economy-wide job automation.
2026-10-28Parameter Labs / Simon Willison
Researchers disclose a fixed ChatGPT shared-page vulnerability
Parameter Labs reports that a specially crafted public ChatGPT share could execute code in a signed-in visitor’s web session, retrieve private conversations, and call account APIs. The report was published on October 28. OpenAI fixed the issue on October 26 and classified it as P1.
The chain crossed isolation boundaries involving HTML Preview, a third-party component used by interactive graphs, and cross-window message validation. Opening the shared link was sufficient for the demonstrated attack; no browser extension installation was required. Simon Willison’s subsequent analysis discusses the sandbox boundary. This is a disclosure of a repaired vulnerability, not a claim that the same hole remains open.
03
Infrastructure and Industry
3 stories
2026-11-02OpenAI
OpenAI plans up to one gigawatt of grid demand response
OpenAI has launched Grid Protectors, a program with utilities and regulators to adjust future computing-campus electricity demand when grids are under stress. The first pilot is planned for Effingham County, Georgia. An arrangement with Georgia Power provides for up to 1,000 megawatts of demand response once the campus is operating.
The campus is still in development, so that figure is a planned ceiling rather than capacity already delivered. Response speed, duration, and any on-site generation or storage depend on local arrangements. OpenAI also says it will pay the full cost of electricity service and supporting infrastructure. Flexible demand is intended to accompany, rather than remove the need for, new generation and transmission.
2026-11-02The Decoder
Anthropic’s reported revenue surge comes with a $23.3 billion loss
According to The Decoder’s account of reporting by Bloomberg and The Information, Anthropic generated nearly $17 billion in quarterly revenue, roughly twelve times the year-earlier level. Operating costs were about $17.5 billion, while the net loss reached $23.3 billion, compared with $12.1 billion a year earlier.
The report attributes part of the spending increase to short-term additional compute and employee donation matching. It also cites roughly $895 billion in compute and data-center lease commitments in updated filings. Those are obligations across future periods, not costs already paid in one quarter. Likewise, the reported net loss is not simply the difference between the revenue and operating-cost figures.
2026-11-02Marvell / Lumentum
Marvell agrees to acquire Lumentum in an all-stock transaction
Marvell and Lumentum have signed an all-stock acquisition agreement under which each Lumentum common share would receive 4.5 Marvell shares. The companies aim to combine Marvell’s interconnect, switching, and custom-chip capabilities with Lumentum’s lasers, optical transceivers, and optical circuit switching for AI data-center connectivity.
They say closer co-design of electrical and optical components could support more integrated optical solutions. Closing is expected in the first half of 2028, subject to both companies’ shareholder approvals, regulatory clearances, and other conditions. The agreement has been signed, but the businesses have not yet combined; anticipated operating benefits remain projections rather than achieved results.
04
Workflows and Learning
2 stories
2026-11-02Kazz
Kazz describes an event-driven personal agent workflow
Creator Kazz has described a personal agent-automation system developed over roughly a year. In his account, events initiate work, a central automation layer uses rules.json for routing and scheduling, and agents run on a local Mac mini before handing delivery work to a separate CI layer.
The author describes five delivery steps: review, checks, merging, publishing, and verification. Decisions involving data deletion, spending, or a change in direction are sent to him through Feishu cards. This is an account of his own implementation, separating events, execution, delivery, and human decisions; it is not evidence of a generally available automation product that works unchanged in other environments.
2026-11-02Google
Google launches a free skills hub for Africa
Google has launched Skilling Africa with Google, a free hub combining more than 15 training initiatives. Its stated goal is to help three million African learners build skills by 2030. Resources include Google Career Certificates, Hustle Academy, and Cloud Skills.
The platform supports Google-account sign-in and is designed for mobile devices and low-bandwidth connections. Learners can download lessons over Wi-Fi and study offline. An AI study assistant called SCOUT can answer questions, recommend learning paths, and help organize a schedule. The three-million figure is a future training target, not a count of people already trained through the newly launched hub.