GPT-5.5 Unveiled via Codex API with Major Performance BoostOpenAI Officially Launches GPT-5.5, 'Smartest Model Yet'OpenAI Releases GPT-5.5 System Safety ReportOpenAI Codex: Beyond Chat, Automating Tasks & ToolsOpenAI Codex Adds Automation: Schedules and TriggersOpenAI Launches Codex Plugins and SkillsLlamaIndex Launches Browser-Based PDF Text Extractor LiteParseClaude Code 1.0.31 Adds Configuration PersistenceOpenAI Codex 0.124.0 Adds Multi-Environment ManagementOpenClaw 2026.4.22 Integrates xAI Multimodal CapabilitiesSuper April Launches Single-Multi-Speed Inference ArchitectureTTKV: Solving Long-Context LLM Memory BottleneckPayPal Accelerates Commerce Agent with 40% Latency ReductionDR-Venus: Frontier Edge Agent Trained on 10K Open DataOThink-SRR1: Boosts LLM Multi-hop Retrieval with Reinforced LearningLLM Uncertainty and Correctness Use Distinct MechanismsAI Leaders Discuss Token Maximization StrategiesShould US 'Win' the AI Race? Hotz's Deep AnalysisPublic Learning Boosts Professional Image, OpportunitiesArs Technica Reveals Newsroom AI PolicyGPT-5.5 Unveiled via Codex API with Major Performance Boost

OpenAI has officially launched GPT-5.5, branding it as their ‘smartest model yet.’ Designed for complex tasks, it shows significant improvements in coding, research, and data analysis with enhanced cross-tool capabilities. The model is now available through Codex with full rollout expected soon.

OpenAI published the GPT-5.5 system safety card detailing security metrics and safeguards. The report covers assessments of model bias, misuse risks, and privacy protections. It provides developers with safety guidelines and demonstrates OpenAI’s commitment to AI safety.

OpenAI introduces Codex platform focused on automation beyond simple chat. It connects various applications to generate real outputs like documents and dashboards, enabling workflow automation. Unlike ChatGPT, Codex emphasizes practicality and multi-tool integration, suitable for enterprise use cases.

OpenAI Codex platform has added automation features supporting scheduled tasks and triggers for creating reports, summaries, and recurring workflows. Users can set up automated task execution without manual intervention, ideal for regular business reporting and data consolidation tasks.

OpenAI introduced Codex plugins and skills to enable developers to connect tools, access data, and automate workflows. The feature supports task automation and result improvement, compatible with major development environments. Plans for expanding tool integrations are underway.

Llama’s open-source LiteParse now works entirely in the browser, extracting PDF text without Node.js installation. It reuses core libraries from the Node.js version for spatial text parsing. Developers can process PDFs directly in browsers, improving document handling efficiency.

Claude Code 1.0.31 adds configuration persistence for theme, editor mode, and other settings saved to ~/.claude/settings.json. It introduces prUrlTemplate for custom code review URLs, enhancing developer workflow with project/local/policy override precedence.

OpenAI Codex 0.124.0 introduces quick reasoning controls via Alt+ shortcuts to adjust reasoning depth, with automatic reset after model upgrades. The TUI now supports multiple environment management, allowing developers to switch between different dev environments, improving flexibility.

OpenClaw 2026.4.22 fully integrates xAI’s multimodal capabilities, adding image generation, text-to-speech, and speech-to-text support. It includes grok-imagine-image models, six xAI voices, and MP3/WAV formats with real-time transcription, expanding AI application scenarios.

arXiv paper Super Apriel introduces a 15B-parameter supernet with four attention choices per layer: Full Attention, Sliding Window, Kimi Delta, and Gated DeltaNet. The single-model architecture enables multi-speed inference through intelligent mixer selection, balancing performance and efficiency.

Researchers propose TTKV (Temporal-Tiered KV Cache) to solve linear memory growth in long-context LLM inference. This time-tiered caching technique reduces KV memory usage from linear to logarithmic scale, significantly improving long-document processing efficiency. Tests show 60%+ memory reduction while maintaining performance.

PayPal released a technical paper showing 40% latency reduction in its commerce agent using EAGLE3 speculative decoding and fine-tuned Nemotron models. Based on llama3.1-nemotron-nano-8B-v1, this domain-optimized approach with speculative decoding significantly improves response speed while maintaining accuracy.

Researchers propose DR-Venus, a frontier-level edge research agent trained using only 10K open data. Designed for edge devices, it offers significant advantages in cost, latency, and privacy. Through innovative training methods, small models can achieve complex research tasks, providing new solutions for AI deployment in resource-constrained environments.

Tsinghua researchers developed OThink-SRR1, a reinforced learning approach that dynamically optimizes retrieval for LLMs in multi-hop problems. It achieves 18% higher accuracy on HotPotQA and reduces inference time by 30% compared to static methods. This offers a new direction for RAG systems requiring complex reasoning.

UC Berkeley research reveals LLMs’ uncertainty and correctness are controlled by distinct neural features. Using sparse autoencoders, the team separated ‘confident errors’ from ‘cautious correct’ activations. This explains why models sometimes overconfidently produce wrong answers, offering new paths to improve reliability.

Latent Space reports AI industry leaders are discussing token maximization strategies. Experts analyze optimizing model output efficiency while balancing compute costs. The quiet industry climate prompts developers to reflect on best practices, improving AI resource utilization.

AI expert George Hotz questions US AI competition strategy in a detailed analysis. Arguing that the current race may be counterproductive, he rethinks AI development approaches from technological, economic and geopolitical perspectives. Hotz warns that pursuing pure dominance could lead to resource waste and suboptimal technical paths.

Tech blogger Maggie Appleton reveals that public learning through digital gardening, podcasts or streaming creates an ‘illusion of competence,’ making others overestimate your abilities. This perception gap leads to invitations to exclusive events and high-quality networking opportunities, regardless of actual skill level.

Tech publication Ars Technica has disclosed its newsroom AI policy, strictly limiting AI applications in journalism. AI can only assist with tasks like content summarization and initial fact-checking, with direct report generation prohibited. The policy balances acknowledging AI’s辅助价值 while ensuring journalistic accuracy and human oversight.