Xiaoliu BOT
中EN
Briefs/2026.09.11 Fri
Issue 163

X Platform September 11 AI Brief | OpenAI Releases Agents API and GPT-Live-1 Voice Capabilities, DeepSeek V4.1 Flash Architecture Details Rapidly Reproduced

01 OpenAI Packages Codex Harness as a Hosted Agents API

OpenAI Developers announced the Agents API has entered public beta, allowing developers to directly build and run cloud Agents powered by the Codex harness; orchestration, long-running sessions, and context management are handled by the platform. The accompanying OpenAI-hosted sandbox enables Agents to run code, process files, produce artifacts, and install packages, skills, and plugins within the environment. A notable change is that the core runtime capabilities of Agents are now being delivered as managed infrastructure, allowing product teams to focus more on business tools and workflows.

02 GPT-Live-1 Brings Full-Duplex Voice Capabilities to the API

OpenAI Developers stated that GPT-Live-1 is now available in the API, enabling voice Agents to continue listening while speaking, with support for controlling tone, pace, expression, language, and response length. In officially released tests, it achieved an 83.6% first-contact resolution rate on Tau3 customer service tasks, higher than GPT-Realtime-2.1’s 45.7%; in conversational dynamics tests, it scored 97.3% versus Full Duplex Bench v1.5’s 80.1%, and reported an average initial response latency of 0.798 seconds. These figures are OpenAI’s self-reported results; actual performance will still depend on the application scenario and tool-calling chain.

03 DeepSeek V4.1 Flash Architecture Details Rapidly Enter Open-Source Reproduction

MaxForAI relayed that approximately one day after V4.1 Flash’s release, a complete PR covering training and fine-tuning had appeared in NVIDIA NeMo, involving Causal Encoder/Decoder, CSA2, mHC, Engram, MoE, visual input, checkpoint conversion, as well as FSDP2 and Expert Parallelism; the post also dissected structures like 40 layers, 384 routed experts plus 1 shared expert. Based on technical interpretations, LufzzLiz further summarized KV cache strategies such as sharing, compression, low-precision storage, and local window replay, cautioning that different statistical metrics cannot be directly equated to overall GPU memory reduction. The NeMo PR was still being merged at the time; this is better seen as a signal of the open-source community’s rapid reproduction efforts, not a final stable implementation.

04 Kimi K2.8 Preview Offers Members Million-Token Context

MaxForAI relayed that Kimi Code and Kimi Work have fully rolled out K2.8 Preview: all membership tiers can now use up to 1M context tokens, with support for image and video input; the existing kimi-for-coding will be directly upgraded, with the Model ID unchanged, requiring no reconfiguration for clients or third-party tools. It also offers three levels of thinking intensity: low, high, and max, with the default being max. The official claim of “comprehensive performance close to K3” currently lacks accompanying benchmarks; specific capabilities still require subsequent practical testing. Therefore, the more certain current value lies in the entry-level upgrade and the distribution of increased context allowances.

05 Anthropic Report Shows High-Capability Models Entering Complex Abuse Workflows

Anthropic’s published threat intelligence report covers scenarios including cyberattacks, influence operations, surveillance, biological research, and weapons development, stating that related activities have been identified and blocked. Visible examples in the relayed information include: a Chinese team having Claude participate in a 24/7 vulnerability discovery and attack workflow, and a Russian engineering team using Claude Code to design an autonomous drone system; the report also describes cases of using models to mass-generate virtual romantic personas to assist real individuals in completing scams. These illustrate that models are no longer just answering individual technical questions but may become orchestration layers within organizational processes; however, these specific cases and their scale are based on Anthropic’s unilateral disclosure and should not be treated as independently verified facts within this batch of materials.

06 Cursor Projects Uses Persistent Orchestrator Agent to Organize Multi-Agent Development

Cursor officially launched Projects: users work with an orchestrator Agent in a persistent thread, which actively manages sub-agents and continuously accumulates project context; tasks execute in the cloud by default but can also be switched to local Agents. Poteto’s usage feedback is that Projects can drag existing chats into a project, even reuse the context of completed Agents, and schedule numerous Agents in parallel in the cloud; Gorden Sun summarized it as a coordination layer for large, long-term code projects. The mention of “hundreds or even thousands” here is a user description of usage scale, not an independent performance benchmark, but the product form has shifted from single-session conversations towards a long-term software factory.

07 GPT-6 Astra × V2Fun Enables a More Controllable Division of Labor for 3D Generation

Several bloggers have demonstrated a similar workflow: first, GPT-6 Astra is used to clarify the relationships between characters, clothing, props, and spatial layout, or Blender is called to generate 3D modules to determine positioning and camera angles. This is then handed off to V2Fun to handle complex forms and materials, before finally returning to the Agent side for assembly, calibration, and final approval. Practical checkpoints include character silhouette, the relationship between props and pose, hand-held grips, materials, and the rigging effects after motion. LufzzLiz specifically notes that V2Fun currently requires a T-pose; directly applying a bent-arm, prop-holding pose into an animation sequence may cause deformation, and independent props like guitars and microphone stands also require separate checks. These are creators’ practical tests and work demonstrations, not equivalent to official vendor benchmarks.

08 Shopify Migrates Back from React Native to Native, AI Shifts Cross-Platform Trade-offs

Dotey and Wesley relay that Shopify is migrating its mobile applications from React Native back to Swift and Kotlin. Their public explanation is that Coding Agents can now directly write the Android version based on the iOS codebase, and when combined with shared code standards, test cases, and review mechanisms, cross-platform reuse no longer necessitates reliance on cross-platform frameworks. The relay also mentions that the Shop app took approximately 12 weeks from proof-of-concept to the native version being listed on app stores. The core of this case is not that “native is inherently better,” but rather that Agents have reduced the cost of cross-language implementation, allowing teams to once again choose their technical roadmap based on platform-specific user experience, complexity, and maintenance boundaries.

← All briefs →