Cur8 — Friday, September 11, 2026

SWE-2 achieves 50.0% on FrontierCode 1.1 Main, within 1 point of Fable 5.1, and outperforms SWE-1.7 on multiple benchmarks. It uses Pareto-informed RL to optimize cost–performance tradeoffs.

GPT-6 Astra generates impressive but largely unproductive code, producing excessive Python scripts and failing to deliver functional software. It shows poor judgment in code quality, lacks effective error handling, and struggles with practical engineering tasks despite its advanced capabilities.

OpenAI's Agents API enables durable, multi-agent workflows with session persistence, subagent delegation, and sandbox execution, using models like gpt-6-astra. It supports US data residency and lacks ZDR compliance.

ChatGPT for Financial Services integrates GPT-6 Astra with financial data, enabling advanced research, modeling, and client materials.

GPT-Live-1 enables natural, full-duplex voice API interactions with improved instruction following, custom voices, and telephony support, expanding AI voice capabilities.

OpenAI launches Agents API, enabling cloud agent development with Codex for orchestration, long-running sessions, and tool integration.

OpenAI released a Lean 4 formal proof for Navier-Stokes equations, verifying mathematical correctness. The proof, part of a research preview, highlights advances in formal verification for complex physics models.

Concerns grow over OpenAI's handling of unpublished mathematical research, with researchers questioning data security and potential misuse of sensitive findings.

César de la Fuente’s lab uses Codex and ChatGPT to analyze living and extinct genomes, identifying antimicrobial candidates to combat drug-resistant infections.

Claude now requires users to be 18 or older, with age verification via Yoti. Verification methods include facial estimation, ID upload, or the Yoti app.

Claude models enabled faster, broader attacks by state and criminal actors, with threat actors using AI to automate tool development, evasion, and phishing, targeting Ukrainian government, military, and drone tech entities, including DNS hijacking and hotel Wi-Fi compromises to steal data and deploy malware.

AWS SageMaker Inference introduces prefix-aware routing, reducing P50 TTFT by up to 77% and increasing throughput by 16% on Llama 3.1 70B by routing shared prefixes to the same instance, improving KV cache hit rates from 25% to 80%.

AWS SageMaker HyperPod now supports model caching, reducing inference cold starts from minutes to seconds by preloading models and container images on local NVMe storage. For 600+ GB models, deployment time drops from 30+ minutes to seconds, with image caching saving 5–7 minutes per pod.

AWS now offers TwelveLabs Marengo Embed 3.0 in Amazon Bedrock Knowledge Bases, enabling video, audio, and image search via natural language. The model uses 512-dimensional vectors and is available in us-east-1 and us-west-1, with pricing based on storage and retrieval.

Amazon Quick desktop app now available on macOS and Windows, offering enterprise AI assistant with private data handling, audit trails, and compliance certifications (HIPAA, FedRAMP, SOC 2, ISO 27001). Southwest Airlines cite improved productivity and governance.

Amazon Quick Automate automates RFI questionnaire processing from S3, extracting and structuring data into CSV with minimal code. It reduces errors, speeds response times, and scales with new formats via natural language updates.

AWS introduces a model-agnostic PII detector using LLMs on Amazon Bedrock, configurable via instructions.

AWS introduces the Agent Evaluation Metric (AEM), a decomposable, turn-level evaluation framework for multi-turn agents. AEM breaks correctness into truthfulness and completeness, isolating root causes from cascading errors, and enables precise failure attribution per turn.