AI agents are starting to plan, delegate, test, and revise work on their own, so developers spend less time wiring workflows and more time reviewing results. OpenAI’s safety firings show the oversight problem is not keeping pace.
Ling-3.1-flash gives developers a 1M-token context window for long coding, search, office, and other agent tasks, while activating 25B of its 560B parameters per token. It routes each token through eight of 512 experts, keeping the model’s working computation smaller than its total size. The model page reports strong results across Terminal-Bench 4.0, SWE-bench Pro, GDPVal-AA V2.1, and healthcare and finance tasks.
Researchers can now generate decades of simulated climate data at 3 km resolution for arbitrary global regions, rather than working only with the model's 100 km inputs. HiROv2 uses diffusion-based downscaling to produce precipitation, 10 m winds, and sea-level pressure at six-hour intervals, while the paired ACE2S-X-SHiELD emulator simulates atmospheric variability from days to centuries. The repository includes checkpoints, inference configs, and a run script.
Gemini 4 Argon (High) delivers a 53 Intelligence Index score at an average evaluation cost of $1.99 per task, giving users a reasonably priced model for text and image input. It accepts both modalities, produces text, and fits roughly 1,500 A4 pages into its 1M-token context window. Its listed API price is $2 per 1M input tokens and $10 per 1M output tokens, while the comparable-model median score is 27.
The Agentic Scientist lets researchers run computational investigations with a coordinated group of coding agents, from idea generation through experiments, analysis, and report writing. A meta-controller runs agents in parallel and checks each stage against a rubric, while sandboxed experiments keep execution separated. The authors report one co-discovered paper accepted at ALIFE 2026 and autonomous tests across machine-learning problem-solving tasks, with ablations pointing to orchestration, not just model capability, as a key factor in reliable autonomy.
Character.ai is still available, but its future now depends on a court-supervised sale: without an actionable bid, the company says it will wind down the platform. The bankruptcy process covers all or substantially all of its assets, with completion expected by the end of November or early December 2026. Users can keep using the service and paid subscriptions for now, with billing unchanged.
Modal says it has crossed a $1B annualized run rate, with usage tripling in five months and growing more than 15x year over year. For developers, the practical bet is a serverless AI stack that handles scarce GPUs, slow-loading models, isolation, and deployment across a capacity pool spanning 20+ clouds. One customer, TypeSafe AI’s Jev, served 1T tokens on Modal in its first three days.
Anthropic’s dynamic workflows let Claude decide how to split a large task, launch agents, pass information between them, and check discrete results, instead of making developers configure a coordinator and fixed roster. The lead agent writes the orchestration program itself, while developers can still add specialist sub-agents. That shifts the bottleneck from wiring agents together to reviewing the work they produce.
OpenAI fired three safety researchers two days after promising independent evaluators employee-like access, making self-policing harder to trust for anyone relying on the lab to disclose dangerous model behavior. The op-ed says one fired researcher had been a contact for an evaluator investigating the Hugging Face incident, while OpenAI said the group mishandled sensitive information and the employees dispute that account. Its proposed fix is oversight with access and protections that do not depend on OpenAI’s approval.
METR proposes giving the public independent evidence about whether frontier AI developers can control their internal systems, so staff, boards, scientists, and policymakers can judge when risks are becoming acute. The plan focuses on assessments of internal development, not just deployed products, and would publish evidence without claiming it can certify that development is safe. METR says it has prototyped these exercises with developers since 2025 and published a pilot report with Anthropic, Google, Meta, and OpenAI.
NVIDIA is releasing the full Run:ai platform as open source, so platform teams, AI clouds, and ecosystem partners can run shared GPU infrastructure without buying a Run:ai subscription. Run:ai dynamically allocates GPUs and manages workloads across data processing, training, and inference on Kubernetes; teams can use the whole platform or components such as the KAI Scheduler and GPU Fractioning Operator. The NGC Community Product requires NVIDIA’s applicable license terms and excludes enterprise support.
PilotDeck is an open-source agent operating system for people running multiple, ongoing projects, giving each one separate files, memory, and skills instead of letting context bleed across work. Its white-box memory lets users inspect and fix what the agent remembered, while Smart Routing matches task difficulty to different models and Always-on work can continue away from the keyboard. It supports Web, CLI, and IM front ends through MCP.
You’re caught up.
That was all 11 stories in the 15 October issue.