New AI models, benchmarks, open weights and frontier model updates curated by ChAIcked.
OpenAI has rebranded and enhanced Codex to enable independent task execution lasting hours without human intervention. The tool aims to support both autonomous workflows and collaborative human-AI interaction patterns.
News · Models · Development Tools · Automation · — Open article
Anthropic has created the Jacobian lens, a tool that provides unprecedented visibility into how Claude processes information and reasons through tasks. The technique reveals both mundane and unexpected patterns in the model's internal operations.
Blog · Models · Research · — Open article
OpenAI rebrands its Codex code generation model as ChatGPT Codex to align with consumer recognition. The move simplifies product naming and reduces brand fragmentation for developers.
News · Models · Code Generation · Branding · — Open article
OpenAI may face court sanctions for allegedly concealing and deleting ChatGPT usage logs during litigation with news organizations over copyright claims. The incident could significantly impact the company's legal defense.
News · Legal · Models · Policy · — Open article
Following regulatory approval from the Trump administration, OpenAI has released GPT-5.6 for public use and unveiled ChatGPT Work, a new AI agent. The company describes GPT-5.6 as its best-performing model to date.
News · Models · AI Agents · Releases · — Open article
OpenAI has released GPT-5.6, emphasizing improved token efficiency, stronger performance per unit cost, and on-demand capability scaling for complex workloads.
News · Models · Releases · Performance · — Open article
OpenAI launched a bug bounty initiative focused on biosecurity risks in GPT-5.5, inviting researchers to identify and report potential misuse vectors related to biological threat modeling.
News · Security · AI Safety · Models · — Open article
An analyst examines OpenAI's simultaneous release of three GPT-5.6 variants across the performance-price spectrum, questioning what strategic questions remain unanswered beyond the initial benchmark excitement.
Blog · Models · Strategy · Opinion · — Open article
Microsoft 365 Copilot now uses GPT-5.6 as its preferred backbone model, delivering enhanced capabilities across Word, Excel, PowerPoint, and Chat for improved productivity workflows.
News · Models · Enterprise AI · Productivity · — Open article
Databricks published benchmarking results for AI coding agents tested against its multi-million line codebase. The evaluation provides real-world performance data for developers assessing code generation tools.
News · Models · Development · Benchmarks · — Open article
Google expanded Android Bench to include new LLM evaluation metrics while noting Gemini's performance gaps. The update invites developer input to shape future benchmark development.
News · Benchmarks · Models · Development · — Open article
An analysis identifying constraints on AI model advancement that extend beyond computational hardware limitations. The piece challenges assumptions about GPU speed as the primary bottleneck in modern AI development.
Blog · Models · Infrastructure · Performance · — Open article
Cognition announced SWE-1.7, an update to its software engineering agent that achieves performance levels comparable to GPT-5.5 and Claude Opus. The release represents progress in autonomous coding capabilities for developers.
News · Models · Development · AI Agents · — Open article
Mistral has introduced Robostral, a state-of-the-art model designed for robotic navigation tasks. The release expands Mistral's portfolio into embodied AI and robotics applications.
News · Models · Robotics · AI Applications · — Open article
A guide addresses observability challenges in multi-model AI applications that use different LLMs for different tasks, explaining how to gain visibility across diverse model endpoints and use cases.
Blog · Monitoring · Operations · Models · — Open article
OpenAI is launching GPT-5.6 Sol along with Terra and Luna model variants for public availability. The announcement indicates a new generation of OpenAI models becoming available to developers and users.
News · Models · LLMs · Product Launch · — Open article
A new method uses large language models to generate neural network improvements by constraining generation to same-family architectures, separating transfer learning effects from adaptation within a controlled source-guided protocol.
Blog · Models · Architecture Search · Development · — Open article
FourTune introduces a fully 4-bit quantization approach for post-training diffusion models, reducing memory requirements and training time to make fine-tuning on diverse downstream tasks more practical for resource-constrained settings.
Blog · Models · Optimization · Fine-Tuning · — Open article
Researchers apply bivariate bicycle codes to reduce error overhead in quantum convolutional neural networks, enabling practical execution on near-term quantum devices by lowering the error-correction threshold requirements.
Blog · Models · Quantum Computing · Optimization · — Open article
A parameter-efficient tri-branch neural network combines graph neural networks, deep networks, and SMILES embeddings to capture diverse topological and structural information for more accurate molecular property prediction.
Blog · Models · Data · Development · — Open article