Today for AIIntelligence for practice

NEWS / 01

News

Follow recent records from AI models, tools, and companies by publication time.

Latest release

GPT-6 Astra

OpenAI Releases GPT-6 Astra: 1.05M Context Enters Autonomous System Operation as Cyber Capabilities Reach Critical Threshold

OpenAI officially released its next-generation flagship model GPT-6 Astra on September 3, 2026. Featuring a 1,050,000 (1.05M) token context window and 128,000 (128K) output limit, Astra pivots toward end-to-end computer operation and autonomous workflows. Scoring 100% on ExploitBench, it is the first frontier model to trigger the Critical cybersecurity risk threshold under OpenAI's Preparedness Framework, prompting restricted access gates and blocking alignment monitoring.

Content priority98

Timeline

Continue by publication time

  1. 02
    Claude Fable 5.1

    Anthropic Releases Claude Fable 5.1: Challenging GPT-5.6 Sol on Long-Horizon Benchmarks with 75% Cheaper Cache Reads

    Anthropic released frontier flagship Claude Fable 5.1 and restricted research twin Mythos 5.1 on September 1, 2026. Benchmarked head-to-head against predecessor Fable 5 and OpenAI's GPT-5.6 Sol, Fable 5.1 takes the #1 spot on BenchLM at 83.0 while more than doubling terminal autonomy scores. Slashing prompt cache read pricing by 75% to $0.25 per million tokens cuts actual multi-turn agent operating expenses by up to 45%.

  3. 03
    Gemini 3.8 Flash

    Google Releases Gemini 3.8 Flash: Breaking 90% on Terminal-Bench with Robust Long-Horizon Agentic Coding

    Google launched Gemini 3.8 Flash and the specialized 3.8 Flash Cyber, targeting end-to-end terminal execution and complex code repair. Retaining a 1M context window and low inference costs, the model leverages iterative tool loops and deeper reasoning to score 90.8% on Terminal-Bench 2.1 and 61.6% on SWE-bench Pro.

  4. 04
    AlibabaQwen3.8-27B

    Alibaba Cloud Open-Sources Qwen3.8-27B: Hybrid Attention and Single-GPU Long-Horizon Agents

    Alibaba Cloud's Qwen team has open-sourced Qwen3.8-27B, a dense 27B vision-language model. Built on a hybrid architecture combining 48 linear attention layers with 16 gated attention layers, it features a native MTP draft head and controllable thinking mode, delivering native 262K context and single-GPU deployment while advancing SWE-bench Pro and LiveCodeBench benchmarks.

  5. 05
    GLM-5.3

    Zhipu AI Releases GLM-5.3 with Scaled Post-Training RL and Emergent Cyber Capabilities, Open-Sourcing Weights in Two Weeks

    Zhipu AI has officially released GLM-5.3, a flagship model enhanced through scaled post-training on a 743B base. Powered by IndexShare, SAO reinforcement learning, and the Slime async training framework, the model ranks first across CyberGym (84.5%), AutomationBench (48.2%), and GDPval-AA v2 (1769) while reaching 28.3 on Terminal Bench 3.0, accompanied by a two-week open-source roadmap.

  6. 06
    DeepSeek HarnessDeepSeek

    DeepSeek Open-Sources Modular Agent Framework DeepSeek Harness

    DeepSeek has open-sourced DeepSeek Harness (dsh), a modular agent framework. Built on the Cordis meta-framework, the system employs a plugin-first architecture to decouple model adapters, execution loops, and tool registries while providing both CLI and local Web interfaces for transparent and extensible agent development.

  7. 07
    Gemini 3.7 Flash

    Google Releases Gemini 3.7 Flash: Optimized for Long-Horizon Coding and Agentic Workflows with 50% Introductory API Discount

    Google has officially launched Gemini 3.7 Flash, optimized for long-horizon software engineering and agentic workflows. Featuring a 1M-token context window and tunable thinking modes, the model delivers major gains on benchmarks like DeepSWE and FrontierCode over 3.6 Flash, accompanied by a 50% API price reduction through the end of 2026.

  8. 08
    DeepSeek

    DeepSeek Launches the Official V4-Pro Release: Benchmarks Beat Preview Across the Board

    The official DeepSeek V4-Pro release is now live as DeepSeek-V4-Pro-0813, with the existing API model name and calling method unchanged. DeepSeek's results show the official release beating Preview across ten Agent-related evaluations and joining the leading group on tasks such as Cybergym and AutomationBench. Production workloads should still be retested internally.

  9. 09
    Qwen3.8-MaxAlibaba

    Alibaba Releases Qwen3.8-Max Frontier Foundation Model

    Alibaba has officially released its next-generation frontier model, Qwen3.8-Max. Built on a 2.4-trillion-parameter sparse Mixture-of-Experts (MoE) architecture with 95 billion activated parameters, it natively supports a 1-million-token context window and multimodal inputs. The new model substantially advances long-horizon autonomous coding and complex agent workflows, achieving top-tier rankings on the Arena overall and vision leaderboards. Alibaba also announced that open weights will be released the following week.

  10. 10
    ByteDanceSeedance 2.5

    ByteDance Releases Seedance 2.5 Video Generation Model

    ByteDance has officially released its next-generation video generation model, Seedance 2.5. Building upon its unified multimodal audio-visual generation architecture, the new model doubles single-clip duration to 30 seconds, supports up to 50 multimodal reference inputs, and introduces timestamp-based local editing alongside native 4K output, accelerating the transition from short-clip generation to industrialized production-ready video creation.

  11. 11
    DeepSeek-V4-FlashDeepSeek

    DeepSeek-V4-Flash Official API Released in Public Beta

    DeepSeek has announced the public beta release of the official DeepSeek-V4-Flash API (DeepSeek-V4-Flash-0731). Retaining the same lightweight MoE architecture with 284B total and 13B active parameters, the re-post-trained model delivers significantly enhanced agent capabilities. Benchmark scores on Terminal Bench 2.1 and Toolathlon approach or exceed the performance of the early V4-Pro-Preview.

  12. 12

    camelAI Architecture Evolution: Ditching VMs and Bash for Cloudflare Durable Objects

    camelAI CTO detailed their evolution of moving coding agents completely off virtual machines. By running the agent inside Cloudflare Durable Objects, persisting filesystems via SQLite and R2, and replacing bash with a JavaScript sandbox powered by Code Mode and Dynamic Workers, camelAI achieved drastic cost reductions, lower latency, and simplified operations.

  13. 13
    OpenAI

    OpenAI Analyzes 800K (80 Ten-Thousand) Work Prompts: 43.5% of Specialized Tasks Show Cross-Occupational Crossover

    OpenAI released a workforce study analyzing over 800K (80 ten-thousand) work prompts from U.S. ChatGPT enterprise users. Excluding generic admin tasks, 43.5% of occupation-specific prompts involved cross-functional tasks. Customer experience, design, and HR had the highest crossover rates, with financial calculation and technical troubleshooting emerging as the most portable skills.

  14. 14
    Model Context Protocol

    Anthropic and AAIF Release MCP 2026-07-28 Spec, Shifting Protocol to Stateless Architecture

    On July 28, 2026, Anthropic and the Agentic AI Foundation released the MCP 2026-07-28 spec. As the largest architectural update to MCP, it removes stateful sessions and handshakes, introducing stateless self-describing requests, header-based routing, and official extensions.

  15. 15
    Codex SecurityOpenAI

    OpenAI Open-Sources Codex Security CLI and SDK for Automated Vulnerability Discovery and Patching

    On July 28, 2026, OpenAI open-sourced the Codex Security CLI and TypeScript SDK on GitHub under the Apache-2.0 license. As an AI-powered application security agent (internally codenamed Aardvark), Codex Security goes beyond traditional SAST tools by constructing codebase threat models, validating vulnerabilities in isolated sandboxes, and generating remediation patches to streamline DevSecOps workflows.

  16. 16
    Kimi K3Moonshot AI

    Moonshot AI Open-Sources 2.8-Trillion Parameter Model Kimi K3 Alongside Training and Operator Infrastructure

    Moonshot AI announced the full open-source release of its flagship model Kimi K3's weights on July 27, 2026. Featuring 2.8 trillion parameters, Kimi K3 adopts the Kimi Delta Attention (KDA) hybrid linear attention architecture and Attention Residuals, supporting native multimodality and a 1M token context window. Moonshot AI also open-sourced its infrastructure tools, including MoonEP for distributed training and FlashKDA operator libraries.

  17. 17
    Google

    Google Releases ATLAS AI Economy Report: Analyzing 14.65M Real Conversations to Assess Workplace AI Impact

    Google DeepMind and the Chief Economist's Office released ATLAS v1.0, analyzing 14.65M de-identified interactions globally. Findings show AI covers 68% of occupations with a 21% median task saturation, full task automation is under 10%, 86% of usage occurs outside work, and non-English conversations represent 66.7%.

  18. 18

    Nvidia CEO Jensen Huang Rejects 'AI Doomerism': Calls Mass Unemployment Fears 'Complete Nonsense', Defends Chinese Open Models and Infrastructure Scale-Up

    In an exclusive Axios interview, Nvidia CEO Jensen Huang sharply criticized AI doomerism and speculative regulation, calling fears of extinction or mass unemployment 'complete nonsense.' He praised the quality of Chinese open-source AI models, debunked safety 'backdoor' myths, urged regulators not to rely solely on advice from a couple of tech CEOs, and projected that global AI computing and energy infrastructure must expand 5 to 10 times.

  19. 19
    AnthropicClaude Opus 5

    Claude Opus 5 Benchmark Results Revealed: Top Rankings Across Frontier Model Comparison Define New Baseline

    Anthropic's newly released Claude Opus 5 achieves standout scores across key AI benchmarks, leading Humanity's Last Exam (HLE) at 64.7%, scoring 91.59% on MMLU-Pro, and reaching 97.00% on SWE-bench Verified, outperforming models including GPT-5.6 Sol, Claude Fable 5, Kimi K3, and Opus 4.8.

  20. 20
    AnthropicClaude Opus 5

    Anthropic Formally Releases Claude Opus 5

    Anthropic officially launched Claude Opus 5 on July 24, 2026. The model approaches the frontier intelligence of Claude Fable 5 while maintaining $5/$25 per million tokens pricing, introducing default thinking and adjustable effort settings alongside major alignment safety advances.