Cur8 — Friday, September 11, 2026

SWE-2 achieves 50% on FrontierCode 1.1 Main, rivaling Fable 5.1 and GPT-Astra at 64% cheaper; beats SWE-1.7 and Grok 4.6 on cost-performance. Available today in Devin Desktop/CLI.

GPT-6 Astra generates excessive, low-quality code for complex tasks, preferring Python over efficient tools like patch. Despite impressive capabilities in other areas, it struggles with practical software engineering applications.

OpenAI's Agents API enables durable cloud agents with managed Codex harness, supporting code execution, file editing, and external connections; priced at standard API rates.

ChatGPT for Financial Services launched, integrating financial data with GPT-6 Astra for research, modeling, and client materials.

GPT-Live-1 enables real-time, two-way voice chats via API, featuring enhanced command adherence, personalized voices, and phone line compatibility.

Agents API enables cloud agent development with Codex-powered orchestration, long-running sessions, and tool integration.

Formal verification of Navier-Stokes fluid dynamics equations in Lean 4, enhancing reliability and safety for AI applications.

Trust in OpenAI for unpublished math research is questioned. Concerns raised by multiple researchers.

Codex and ChatGPT identify antimicrobial candidates from genomes, targeting drug-resistant infections.

Claude now restricts access to users over 18, requiring age verification via Yoti for suspected minors. Account reinstatement follows successful verification using selfies, IDs, or Yoti's Digital ID app; Anthropic receives only pass/fail results.

AI misuse disrupted across seven harm areas; Claude Haiku, Sonnet, Opus models used. Cyber operations automated by AI, reducing attacker sophistication needed; state-sponsored groups, criminals, and individuals involved. GTG-20006, linked to Russian espionage, used AI-driven workflows for tool development, phishing, and data exfiltration, targeting Ukrainian and European governments, drone technology providers.

Amazon SageMaker Inference's prefix-aware routing reduces Llama 3.1 70B P50 TTFT by up to 77%, increases throughput by 16%, and boosts KV cache hit rates from ~25% to over 80%.

Amazon SageMaker HyperPod's new model caching cuts inference cold starts from 30+ minutes to seconds for large models like DeepSeek-R1 (600+ GB), with benchmarks showing ~60% faster scale-out and up to 97% reduction in image-pull time.

Marengo 3.0 now available in Amazon Bedrock Knowledge Bases, enabling natural language search across video, audio, and image content with a 512-dimensional vector space. Available in us-east-1 and us-west-1 regions.

Enterprise AI assistant Amazon Quick generally available for desktop (macOS, Windows) and mobile (iOS, Android). Cuts time-to-insight significantly while maintaining governance and accuracy; supports HIPAA, FedRAMP, SOC 2, ISO 27001. Used by Southwest Airlines, LabCorp, PGA TOUR.

Automate RFI questionnaire workflows with Amazon Quick Automate, reducing errors and response times; integrates with Amazon S3 for data storage and retrieval.

AWS introduces a configurable, instruction-driven PII detector for LLMs on Amazon Bedrock, evaluated against five public corpora and eight other detectors, including OpenAI PrivacyFilter. The detector is model-agnostic, supports multilingual contexts, and requires no retraining for new entity types.

AEM metric introduced for multi-turn agent evaluation; pinpoints root cause turns, separates from cascading effects. Measures correctness via truthfulness and completeness sub-metrics, with binary pass/fail per turn.