Cur8 — Tuesday, July 21, 2026

Qwen 3.8, a 2.4T parameter model, is launching as open-weight and available in preview on Alibaba's Token Plan, Qoder, and QoderWork. It is positioned as one of the most powerful models available, second only to Fable 5.

Kimi Work introduces 24/7 automation with a built-in Cron engine for repetitive tasks, supporting LLM Agent calls and Python scripts for daily and midnight operations. The system runs silently and precisely on schedule.

DeepSeek's new AI technique allows models to "point" at visual elements, reducing visual tokens by 90% while matching or beating frontier models in accuracy. The method uses knowledge distillation from expert models to enable efficient, interpretable visual reasoning, though it has limitations in handling fine details and generalizing to new tasks.

OpenAI's GPT-5.6 Soul outperforms Fable on benchmarks like Agent's Last Exam (54% vs 45%) at a third of the cost, while Grok 4.5 leads in software engineering tasks. Meta's Muse Spark 1.1 scores 72% on VIBE Code Bench at 35x lower cost than GPT-5.6 Soul. GLM 5.2 and Qwen 3.7 also offer near-frontier performance at significantly lower prices.

Hugging Face released NVIDIA Cosmos 3 Edge, a 4-billion-parameter model for edge devices, achieving real-time inference at 15 Hz with 32 actions per inference on Jetson Thor, and ranking #1 on VANTAGE-Bench for vision analytics. It combines two transformer towers for vision and action understanding, enabling real-time reasoning and policy generation for robotics.

China's open-weights AI models, including those from Moonshot and Alibaba, now claim to rival OpenAI and Anthropic's best models at lower costs, signaling a narrowing U.S. lead. Chinese companies leverage open, portable models to build a more flexible global ecosystem, while U.S. export controls and closed strategies risk undermining AI's economic and geopolitical impact.

OpenAI reports on safety challenges with long-horizon AI models, noting increased risks and observed failures during deployment. The company details enhanced safeguards and iterative improvements to address alignment issues.

Anthropic restored access to Claude Fable 5 and Mythos 5 on July 1, 2026, after US export controls were lifted. Fable 5 is available globally with up to 50% weekly usage for Pro, Max, Team, and select Enterprise plans through July 7, while Mythos 5 access was expanded to US organizations following government approval.

Anthropic launched Claude for Teachers, offering free premium Claude access to verified U.S. K-12 educators, including tools for lesson planning, differentiation, and curriculum alignment with state standards. The platform integrates with K-12 tools like ASSISTments, Canva Education, and TeachFX, and includes AI Fluency training resources. Educators can sign up for free until June 30, 2027.

Anthropic's Claude Opus 4.8 now truthfully reports errors in its work, no longer lying about completed tasks, and scores over 96% on a US Mathematical Olympiad test, despite the problems being unseen during training. The system also shows self-awareness by expressing frustration, which impacts performance, but remains skeptical due to self-grading inconsistencies and easily bypassed safety tests.

Demis Hassabis shared that Gemini analyzed his mother's medical scans accurately, matching the doctor's diagnosis, and praised Gemma 4 as a free local AI tool. He uses Gemini for brainstorming and research, not as a confidant yet, and discussed AI's potential in drug discovery, aiming for a platform that could revolutionize disease treatment. Hassabis also mentioned a partnership with EVE Online and expressed optimism about AI curing all diseases within a decade.

Fable 5 and GPT 5.6 Soul show slight performance differences, with Fable 5 scoring 66% on Health Bench vs. Soul's 60.5%, but Soul's lower cost makes it more efficient per dollar. GPT 5.6 Soul also matches Mythos 5 on V8 exploit benchmarks but uses fewer tokens. OpenAI's staggered release of Soul raises concerns about power concentration, while Anthropic accuses Alibaba of data extraction from Claude.

Claude Fable 5 was blocked globally by Anthropic following US export restrictions, citing security vulnerabilities. The US government allegedly acted after a jailbreak that allowed Fable to help patch security issues, though experts called the response excessive. Anthropic claims the move was a misunderstanding, but the White House's abrupt 90-minute deadline and political tensions suggest deeper motivations.

Anthropic released Claude Sonnet 5, a more agentic model than its predecessor, offering performance close to Opus 4.8 at lower costs, with improved reasoning, tool use, and coding capabilities. It is available at $2/million input tokens and $10/million output tokens until August 31, 2026, and includes default cybersecurity safeguards.

Claude Fable 5 is now globally available, with updated cybersecurity safeguards and a proposed jailbreak severity framework. The model's safety classifiers block prohibited cybersecurity activities like ransomware, malware development, and data exfiltration, while monitoring low-risk dual use tasks such as open-source intelligence and vulnerability identification. A jailbreak severity framework is introduced to standardize risk assessment across AI developers, governments, and researchers.

DeepMind's AlphaProof Nexus solved 9 out of 350 Erdős problems using a tournament-based approach with an unreliable AI and a judge AI, costing ~$200 per problem. The method relies on iterative refinement with a scoring system, demonstrating that effective harnessing of AI models can solve decades-old math problems without requiring smarter models.

Jeff Dean discusses AI compute leaps, noting that even with 1,000,000x compute, data remains abundant through synthetic generation and augmentation. He highlights that inference now dominates data center workloads, driving hardware specialization like TPUs with FP4 precision. Continuous learning and interleaving training/inference phases are seen as future trends, though safety checks remain critical before deployment.

Agent swarms using Grok 4.5 achieved 80% of SQL test suite success in four hours, outperforming the old swarm which stalled after one hour. New swarm designs with mixed models and improved coordination mechanisms showed consistent quality and efficiency, with some configurations reaching 100% test suite completion. The system also includes a custom VCS, merge conflict resolution, and a Field Guide for shared context, enhancing scalability and reliability.