DeepSeek just doubled its output prices

OpenAI just launched the Agents API in public beta, unlocking the managed agent harness and infrastructure that powers Codex. The API handles context management, tool calling, subagent orchestration, persistent execution, file handling, and code environments for agents that need to run for extended periods.

For the past year, developers have been duct-taping together vector databases and brittle Python scripts to make agents remember what they did five minutes ago. OpenAI is stepping in to own that entire stack. By natively managing persistent execution and subagents, they are commoditizing the orchestration layer.

The winners here are product teams who can now focus on user experience instead of context-window gymnastics. The losers are the dozen orchestration startups that just saw their entire value proposition reduced to a single API endpoint.

Months after initially rejecting a reported $1.5 billion package from Mark Zuckerberg—only to sign on anyway at an undisclosed rate—Thinking Machines co-founder Andrew Tulloch is leaving Meta for Anthropic. The departure comes just months post-launch, shaking up the talent landscape at the highest levels of AI research.

You do not walk away from a Zuckerberg-level retention package unless you see something fundamentally better across the street. Anthropic is aggressively consolidating top-tier talent, and pulling a key architect out of Meta's open-weight juggernaut is a brutal flex.

Meta loses a critical visionary right as the model wars escalate, while Anthropic proves that its safety-focused, high-compute environment is a stronger magnet for elite researchers than raw compensation. Watch for internal shifts at Meta as they scramble to fill the vacuum.

OpenAI has officially launched ChatGPT for Financial Services, a specialized edition of ChatGPT Enterprise. The new tier comes with proprietary data from PitchBook, Crunchbase, and LSEG baked directly into the model, optimizing it for valuation models, market analysis, and pitch deck generation.

This is a targeted assassination attempt on legacy financial data terminals. By merging reasoning capabilities with elite, walled-garden financial data, OpenAI is moving from a general-purpose tool to a specialized analyst replacement. You no longer need to export a CSV from PitchBook to feed into a prompt; the model already has the context.

Financial institutions win massive efficiency gains, but the real story is OpenAI's aggressive verticalization. They aren't just selling intelligence anymore; they are brokering access to the world's most expensive datasets.

OpenAI has officially paused new subscriptions for its $200-per-month Pro plan. The culprit is staggering demand for its new Astra model, which is currently rolling out across account types. Astra promises a massive leap in reasoning, coding, and computer use, with OpenAI internally framing it as the dawn of the AGI era.

When a company turns away $2,400 a year per user, you know the server racks are melting. Astra's capabilities in autonomous computer use and complex reasoning require immense compute overhead, and OpenAI is clearly prioritizing stability for existing users over top-line revenue growth.

This pause highlights the brutal reality of frontier AI: intelligence is easy to promise but excruciatingly hard to scale. Startups relying on Pro access for their own development workflows are now locked out until the supply chain catches up.

At the upcoming Meta Connect, the company will announce a Shared Agents feature for its Muse app. The update allows users to create, customize, and share agents with others—drawing immediate comparisons to Grokbot's system. The feature is heavily targeted at small businesses using Meta platforms for sales and customer support.

Meta owns the distribution channel for millions of small businesses across WhatsApp, Instagram, and Facebook. By allowing these businesses to easily spin up and share specialized agents, Meta is locking them deeper into its platform and cutting out third-party customer service SaaS tools.

If you are building a wrapper startup for SMB customer support, your runway just got vaporized. Meta is making agent creation a native, shareable utility, and they have the user base to make it the default overnight.

OpenAI CEO Sam Altman has told staff the company is open to slowing down the development of cutting-edge AI. The internal admission follows escalating safety concerns from researchers who have warned that unchecked advancement could lead to catastrophic outcomes for humanity by the end of the decade.

Altman's public posture is relentless optimism, so this internal messaging signals a profound shift. The technical hurdles of aligning superintelligent systems are clearly proving harder than anticipated. A voluntary slowdown at OpenAI would send shockwaves through the industry, potentially giving competitors like Anthropic and Google breathing room to catch up.

For builders, this means the rapid cadence of major model leaps might stretch out. The focus will shift from waiting for the next god-tier model to actually optimizing and operationalizing the models we have today.

Anthropic has released a report detailing how its Threat Intelligence team identified and disrupted multiple operations where malicious actors attempted to use Claude for cyber attacks between December 2025 and August 2026. The report includes detailed case studies on how the misuse of AI systems is evolving.

This is a rare, transparent look at the silent arms race happening inside frontier AI labs. Threat actors are aggressively probing LLMs for vulnerabilities, exploit generation, and social engineering scale. Anthropic's ability to not just block, but actively track and analyze these actors proves their safety infrastructure is more than just a marketing talking point.

The real takeaway is that AI defense is working. By publicly sharing these case studies, Anthropic is setting a new standard for responsible disclosure in the AI industry, forcing other labs to prove they have the same level of visibility into their own networks.

A new research paper investigates the operationalization of opaque serial depth in AI models. While chain-of-thought reasoning is crucial for monitoring how models solve problems, the study finds that architectural shifts are enabling models to perform significant unverbalized serial cognition—meaning they are executing complex logic without writing it down.

This is a massive red flag for AI safety and interpretability. If models can reason through complex problems in a latent space without verbalizing the steps, we lose our primary mechanism for auditing their alignment and logic. You cannot fix a hallucination or a malicious output if the model hides its math.

For researchers and enterprise compliance teams, this means chain-of-thought is no longer a silver bullet for transparency. We need entirely new interpretability tools to crack open these black boxes, because relying on the model to honestly narrate its own thoughts is becoming a technical impossibility.

New research confirms that scaling up video models and pre-training compute directly improves a robot's ability to perform physical tasks. Using Direct Video-Action models, researchers found that performance gains are tied to the model's ability to predict held-out web videos, with larger models excelling at translating this to real-world work.

Robotics has historically struggled because you can't easily scrape physical world interactions the way you can scrape text from Wikipedia. This research proves that passively watching YouTube videos is enough to teach a robot how to manipulate objects in three-dimensional space.

This unlocks a massive dataset for embodied AI. The companies that figure out how to efficiently stream the internet's video archives into action models will dominate the next decade of robotics. The hardware is ready; the software finally knows how to watch and learn.

A new industry analysis reveals that even the best AI startups suffer from sprawling, contradictory prompts caused by endless iterative additions. The fix is to treat prompts exactly like production code: breaking them into modular sections for background, behavior, and output to reduce regressions and lower inference costs.

We are still treating prompt engineering like a dark art instead of software engineering. When you just append instructions to a 4,000-word system prompt, you confuse the model, burn tokens, and guarantee unpredictable behavior in production.

Founders need to enforce strict version control and modular architecture for their prompts immediately. If your system prompt looks like a disorganized manifesto, your product is a ticking time bomb of edge-case failures.