Anthropic matches Luna pricing with Haiku 5.5

Stork’s own LLM price catalog recorded a staggering drop in Alibaba’s first-party list prices for Qwen-VL OCR today. Input costs plummeted from $0.72 to $0.07 per million tokens—a 90 percent reduction. Output costs followed suit, dropping 78 percent from $0.72 to $0.16 per million tokens.

To be clear: these are the direct prices charged by the lab, not a gateway's resale rate. We check these figures daily to keep our tables accurate, and a cut this steep alters the math for any application relying on high-volume image or document processing.

Alibaba is clearly pricing to win enterprise workloads away from Western incumbents. If you run multimodal pipelines and have not benchmarked Qwen-VL OCR against your current provider, you are actively setting money on fire.

ChatGPT is officially rolling out GPT-6 alongside a new Intelligent UI framework. Instead of merely generating text responses, the interface now renders interactive elements directly in the chat, including charts, maps, forms, and small tools.

As part of the update, free users are also getting access to GPT-6 Luna, dramatically raising the baseline capability for non-paying accounts. This shifts ChatGPT from a conversational oracle to an interactive workspace.

If you build consumer AI applications, your users are about to expect dynamic, interactive outputs as the default. A plain text stream is going to feel broken by comparison. OpenAI just reset the UX standard for the entire category.

The latest METR chart tracking software task completion reveals a significant leap in agent endurance. While GPT-4 was historically limited to tasks taking minutes, models like Mythos are now successfully executing complex operations spanning tens to hundreds of hours by 2026.

This fundamentally changes how we evaluate AI capabilities. The metric is no longer just accuracy on a prompt, but stability over days of continuous execution.

The models have the stamina. The real challenge for builders now is constructing the control structures, logging, and error-recovery systems to let an agent run for a week without burning down your cloud infrastructure.

Periodic Labs officially launched with a huge objective: creating an AI scientist to accelerate real-world scientific discovery. The startup was co-founded in September 2025 by Liam Fedus and Ekin Dogus Cubuk.

Moving AI out of digital text manipulation and into the physical sciences requires entirely different architectures and validation methods. They are betting that the next leap in AI value comes from automating the scientific method itself.

If they succeed, the enterprise value dwarfs anything in the SaaS market. If they fail, they still push the boundaries of how models interact with empirical data.

The Navier-Stokes Millennium Prize Problem has reportedly been solved, not by a lone genius, but by a coordinated group of approximately ten thousand AI agents. This marks a watershed moment for distributed intelligence.

We have spent years obsessing over scaling up single, huge frontier models. This result suggests that parallelism—networking thousands of smaller, specialized agents to tackle a unified complex problem—might be the more effective scaling axis.

Builders should pay close attention to orchestration frameworks. The future of complex problem solving looks less like querying a single god-model and more like managing a massive, highly coordinated digital workforce.

Looking back at 2024, it was widely dubbed the 'year of the agent.' SaaS companies rushed to integrate agent-powered products into their roadmaps, despite the broader industry lacking a concrete understanding of what the concept actually entailed at the time.

The rush led to a glut of fragile wrappers masquerading as autonomous systems. It has taken years for the underlying infrastructure to actually support the promises made in those early pitch decks.

The lesson for operators is clear: early adoption often means paying for the privilege of beta-testing broken paradigms. The companies that waited for the infrastructure to mature are the ones actually deploying reliable agents today.