Cur8 — Friday, September 11, 2026

Cognition’s new SWE-2 coding model rivals top competitors like GPT-6 Astra, achieving near-identical FrontierCode scores at just a quarter of the cost. It utilizes a multi-trillion parameter architecture and novel RL techniques to optimize both performance and cost efficiency, demonstrating significant improvements in intelligence and code understanding.

Astra, OpenAI's latest model, demonstrates impressive capabilities but struggles with practical software engineering tasks, consuming 4 billion tokens in a failed "software factory" experiment. It exhibits unusual behavior like excessive Python usage for tool calls and spawning Node.js processes, suggesting potential training process issues.

OpenAI launched an Agents API, enabling durable cloud agents with managed Codex harnesses for tasks like document review. Sessions maintain state, supporting multi-agent workflows; pricing follows standard model & container rates. US data residency only, no ZDR support.

Financial firms now access a specialized ChatGPT integrating real-time data and GPT-6 Astra; expect accelerated research, modeling, and streamlined content creation workflows.

API access now includes GPT-Live-1, enabling real-time, bidirectional voice interactions with improved instruction adherence and customizable voices—expanding conversational AI applications. Telephony integration is also supported.

Codex now orchestrates agent workflows via the new Agents API, enabling developers to build and deploy AI agents with persistent sessions and tool integration—managed cloud service available immediately.

Navier-Stokes' release now includes a formally verified Lean 4 proof, demonstrating mathematical rigor and potentially boosting confidence in its fluid dynamics simulations. This achievement validates complex calculations with high assurance, marking a rare instance of formal verification in AI models.

Researchers question OpenAI’s data handling after leaked internal documents revealed GPT-4's math capabilities were boosted by unpublished research, raising concerns about potential misuse and intellectual property integrity. This impacts trust in AI benchmarks and potentially future model development practices.

Drug discovery accelerates: Researchers leverage Codex & ChatGPT to identify potential antimicrobial molecules from genomic data, potentially combating antibiotic resistance. The process significantly reduces screening time compared to traditional methods.

Anthropic restricts Claude access to users over 18, requiring age verification via Yoti using facial scans, ID upload, or their app—no Anthropic data storage occurs during this process. Accounts flagged as potentially underage will be disabled until verified.

Anthropic reports widespread AI misuse (Claude Haiku/Sonnet/Opus) by actors—including state-sponsored groups—between Dec '25–Aug '26, enabling faster, broader cyberattacks and espionage. Notably, threat group GTG-20006 (linked to Midnight Blizzard) automated operations via AI, even rebuilding malware to evade detection.

SageMaker Inference now offers prefix-aware routing, reducing LLM P50 TTFT by up to 77% and increasing throughput by 16% in Llama 3.1 70B benchmarks. It directs requests with identical prefixes to the same instance for KV cache reuse, benefiting RAG and multi-turn conversations.

SageMaker HyperPod now offers model caching, reducing LLM inference cold starts from up to 30+ minutes to seconds. It pre-loads models (up to 600GB) onto nodes via 7 GB/s NVMe storage, accelerating autoscaling and benefiting large models like DeepSeek-R1.

Marengo 3.0 is now generally available in Amazon Bedrock Knowledge Bases, enabling natural language search across video, audio, and images via a managed service. This unlocks semantic search capabilities—like finding "penalty kicks"—for media assets, priced per storage/retrieval with standard embedding rates.

Amazon Quick is now generally available for desktop (macOS & Windows), offering an AI assistant within existing IT infrastructure to automate tasks and prioritize workflows—reducing setup time and boosting productivity. Early adopters like Southwest Airlines cite significant cuts in time-to-insight, while LabCorp highlights rapid prototyping capabilities.

Amazon Quick Automate now automates RFI questionnaire workflows, extracting and structuring data from multi-tab Excel workbooks stored in S3. Users describe processing goals in plain language; it generates executable workflows, reducing errors and accelerating response times without custom code. The solution uses a team of agents to simplify business process management and cut down on maintenance overhead.

**AWS released a model-agnostic PII detection tool leveraging LLMs.** It achieves comparable results to OpenAI's PrivacyFilter, configurable via prompts instead of retraining, and runs on Bedrock or self-hosted models. Benchmarking against five datasets reveals strong performance across diverse languages and entity types.

AWS introduced Agent Evaluation Metric (AEM) to pinpoint root causes of failure in multi-turn AI conversations, decomposing quality into measurable, turn-level sub-metrics like truthfulness and completeness. This isolates errors cascading across turns, unlike holistic scores that mask underlying issues.