Multiple reports indicate that Stripe has reached an agreement to acquire AI model routing platform OpenRouter for over $7 billion, a significant increase from its previous valuation of approximately $1.3 billion. OpenRouter connects over 400 models through a unified interface, and is said to serve around 8 million developers. If completed, the acquisition would extend Stripe's business beyond payments into enterprise model access, routing, and cost management infrastructure. However, the materials do not disclose the closing timeline or details on future product integration.

Anthropic has stated it will not release its more capable internal 「Model 2」. According to reports, this model outperforms the company's current flagship in programming, agent tasks, and data generation, but its potential risks are difficult to fully assess. Anthropic has also revised its estimate of the model's likelihood of exhibiting inaccurate behavior in high-risk scenarios from 「very low」 to 「low」, although it still considers the risk of severe harm to be limited. This decision reflects how frontier model release thresholds are increasingly influenced by internal risk evaluations.

OpenAI has shared multiple application metrics for GPT-5.6: Sol's performance on ARC-AGI-3 improved from 13.3% to 38.3%, with output tokens reduced by approximately six times; Hypha AI used Luna for document extraction, retaining about 98% of GPT-5.5's accuracy at 1/18th the cost; RogoAI reduced input tokens by 21% in financial research while maintaining evaluation quality. Base44 reported a 20% reduction in token usage during application build testing. However, these results are primarily based on partner and vendor reports.

Cursor has begun rolling out a test version of Origin, a code hosting platform designed for high-concurrency collaborative coding scenarios involving AI agents, deeply integrated with the Cursor development environment. Origin supports synchronization with GitHub and launched with integrations for existing GitHub tools including Vercel, Buildkite, and Depot, with further expansions planned. This move marks Cursor's expansion from an AI-powered coding client into code repository and collaboration infrastructure. However, pricing, capacity limits, and full release timelines have not yet been disclosed.

Webwright proposes enabling web agents to write and execute Playwright and bash scripts instead of clicking through pages step-by-step, storing logs, screenshots, and output files in a local workspace. The core implementation consists of approximately 1,000 lines of code. GPT-5.4 achieved 86.7% on Online-Mind2Web and increased its Odysseys score from 33.5% under coordinate-based control to 60.1%. A single task costs about $2.37, higher than the $0.30 for click-based approaches, but generated scripts can be reused and inspected.

The open-source project PRO-LONG employs a minimalist procedural memory mechanism: continuously appending the agent's observations, actions, and outcomes to a single log, which the model later retrieves using grep or Python—eliminating the need for sub-agents or dedicated retrieval systems. On the full public ARC-AGI-3 game set, it averages 18 percentage points higher than no-log baselines, achieving performance equal to or better than specialized frameworks with 4.2 to 5.8 times fewer billed tokens. Fable 5's best@2 score reached 97.4%, with total cost around $1,750.

Dharma AI released a constraint-aware GPU allocator that models real-time demand as time-varying curves and schedules batch jobs into off-peak periods based on priority. On identical hardware, cluster utilization increased by up to 33 percentage points, with priority-weighted output rising by an average of 52% and up to 105%. The system uses a 24-hour rolling planning horizon but only commits the current time step, re-optimizing every 30 to 60 minutes. Its heuristic algorithm directly encodes constraints such as continuous allocation, GPU-switching limits, and non-preemption, with latency under 1 millisecond.

Deno founder Ryan Dahl proposed treating production-grade agents as untrusted software and using an external proxy, Claw Patrol, to inspect their outbound operations. Positioned below the HTTP layer, the system isolates credentials from agents, parses protocol traffic including PostgreSQL, and blocks destructive commands via versioned HCL policies—even when traffic is tunneled. Policies can trigger Slack-based manual approval or LLM-based adjudication, supported by dashboards and rule test cases to ensure sensitive operations remain intercepted.