Google has released Gemini 4 Argon, saying it increased internal agent code commits by over 35% and reduced rollbacks by nearly a third. API pricing is $2 per million input tokens and $10 per million output tokens. Anthropic's daily code-change merges have risen from about 2,500 to over 30,000, and it has moved its entire monorepo from Git to its in-house Scion system. Meta has introduced Muse Wonder and Wonder Full, supporting sustained agent work with cross-conversation memory and the ability to try alternatives when blocked. Wonder is now available via API. Character.AI has filed for bankruptcy and agreed to sell some assets to Disney for $20 million.
Gemini 4 Argon reports gains in output and fewer rollbacks, while Scion addresses infrastructure pressure from the growing volume of code changes.
Google has released Gemini 4 Argon, making its API available to developers and enterprises through platforms including AI Studio. Google says internal teams using the model increased agent code commits by over 35% and reduced code rollbacks by nearly a third. The Chrome team shortened an architectural refactor from months to weeks. Launch API pricing is $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached inputs.
Read original →Anthropic's daily code-change merges rose from about 2,500 in May 2026 to over 30,000 in September. The company spent five months building its own version control system, Scion, and moved its entire monorepo off Git. Unlike Git's design, which supports offline work, Scion targets always-online agents and tasks: the server performs merges and history queries, while clients fetch data on demand. Anthropic emphasizes that its workload is unusual and expects most organizations will not face the same problems in the near term.
Read original →🔍 Analysis: Gemini 4 Argon's internal usage figures pair a more than 35% increase in code commits with a nearly one-third reduction in rollbacks, showing that agent output cannot be measured by commit counts alone. Anthropic's rise from about 2,500 daily merges to over 30,000 shifts the question to how much version control systems can handle. Beyond choosing a model, development teams also need to assess whether their collaboration systems can absorb the additional changes. But Scion is designed for a specialized, always-online workload; this does not imply that ordinary teams also need to leave Git.
Muse Wonder addresses memory and obstacles during tasks, while QwenGyre tackles resource consumption in training on long trajectories.
Meta has released Muse Wonder and Wonder Full for personal and workplace agents. Active Memory remembers and retrieves relevant information across conversations, while Active Focus lets agents try alternatives when blocked and keep tasks moving. Wonder is available through the API and Muse Code; Full is available only to Muse Code subscribers.
Read original →A research team has introduced QwenGyre, an online reinforcement learning framework that achieves speedups of up to 1.85× over colocated approaches and 1.78× over asynchronous approaches. It reallocates GPUs between rollouts and training without interrupting execution, deduplicates branching trajectories, and evaluates partial progress to reduce GPU idle time and redundant training. In experiments with 700,000 tokens per rollout, its NL2RepoBench score rose from 52.5% to 58.5% after 48 training steps.
Read original →🔍 Analysis: Muse Wonder's Active Memory and Active Focus address how agents keep working, while QwenGyre tackles the training costs of such extended execution. At 700,000 tokens per rollout, duplicate trajectories and idle GPUs magnify resource consumption. The two address the execution and training sides of long tasks, respectively. This means developers evaluating these agents need to consider both whether they can carry tasks forward and the compute costs of sustained operation, rather than comparing only single-turn answers.
Maia 300 uses tokens per megawatt and rack-level throughput as performance metrics.
Microsoft has released its next-generation accelerator, Maia 300, an upgrade over Maia 200. Microsoft says it produces 40% more inference tokens per megawatt than leading GPU systems. At rack scale, token throughput can reach up to 7 times that of Maia 200.
Read original →Character.AI enters bankruptcy proceedings, while Cursor joins a subscription system with shared usage allowances across products.
Character.AI has filed for bankruptcy. The company has also agreed to sell some assets to Disney for $20 million. The sale covers only a portion of its assets.
Read original →Cursor has announced it is joining The X Subscription, with paid plans using a shared usage pool. The pool covers Cursor, Grok Build, and other Grok products, allowing the same allowance to be used across products. Existing Cursor subscribers can keep their current plans or switch to the X subscription.
Read original →Arena uses real-world execution trajectories to evaluate safety and reliability, while Anthropic limits extraction of full reasoning traces.
Arena announced last week that it had completed a $200 million Series B round at a valuation of $3.1 billion. It also launched the Arena Alignment Index, which evaluates AI safety and reliability. The index is based on real-world agent execution trajectories, using records of actual behavior for its assessments.
Read original →Anthropic has published research on distillation defenses, saying early evidence shows its new measures reduced the amount of data collected by attackers by about 90%–99%. The defenses focus on limiting extraction of full reasoning traces, because this data is more useful than final answers for replicating model capabilities. The company also alleges that Zhipu's GLM-5.3 distilled cybersecurity capabilities from leading U.S. models and says its safeguards are easy to bypass. These measures also have costs: Anthropic has rejected substantial paid traffic, and some legitimate users need to update their integrations.
Read original →🔍 Analysis: Arena Alignment Index bases its evaluations on real-world agent execution trajectories, while Anthropic's distillation defenses emphasize that full reasoning traces are more useful than final answers for replicating capabilities. Behavioral data can both help assess systems and expose their capabilities. The data used in these two cases is not equivalent, but both bring the boundaries of data access to the forefront. Anthropic's rejection of substantial paid traffic, along with integration updates required of some legitimate users, shows that protecting capabilities is already directly affecting commercial API use.