> SPOTLIGHT
WHAT MATTERS TODAY
Stripe has acquired OpenRouter for more than $7 billion. The deal puts Stripe in a new position within the AI ecosystem: not just providing payment infrastructure for businesses, but sitting between them and the AI models they use.
OpenRouter is an AI gateway that lets customers choose the right model for each task and budget instead of being locked into a single provider. The platform offers one access point to more than 400 models and says it has 8 million users worldwide. CEO Alex Atallah once described OpenRouter as the “Stripe of AI.”
What Stripe bought, then, is not just another AI product. It is access to the distribution layer where businesses decide which model to use, how to route requests, and how to optimize costs. As models become easier to replace, that routing and distribution layer could become the most valuable part of the ecosystem because it determines who gets access to what, at what price, and with how much vendor lock-in.
Dario Amodei made the argument after investor Gavin Baker said that Amodei's warnings about AI risks were helping fuel a backlash, particularly against data centers. Amodei rejected that view, but acknowledged that the public has a negative perception of AI.
Amodei says the deeper problem is a crisis of trust that has built up over decades. Promises such as “AI will cure cancer” are not enough to rebuild that trust. The public wants to see real results.
On regulation, Amodei argues that AI should not be viewed through only two options: leaving it unregulated or concentrating power in a few large companies. He supports rules of the road that reduce cyber, bio, and alignment risks, constrain frontier labs, and still leave room for open-weight models. Open weights can broaden access, he says, but they do not solve concentration because compute and chips remain concentrated among the best-resourced players.
The Financial Times warns that the next China shock could come from open-weight AI. China is building a self-reinforcing ecosystem spanning models and compute, technical standards, and governance, then taking that ecosystem abroad as a global public good.
“Open source” in this context is more accurately described as open weight: the model weights are public, but the training data may not be. That approach lowers adoption costs, fits emerging markets, and helps China find more customers as domestic demand weakens.
The trade-off is that once weights can be downloaded and modified, the original developer has less ability to update safety protections. China's AI influence could therefore spread through models, technical standards, and governance, not just through products.
> SIGNAL HEADLINES
CAPTURE THE SHIFT
DeepSeek Harness turns agent orchestration into replaceable modules: Rohan Paul highlights DeepSeek Harness, where models, tools, storage, scheduling, UI, and sub-agents can be swapped as modules. The pattern makes the orchestration layer a composable part of the agent stack.
Coding harnesses should not default to the terminal: Patrick Collison argues that the terminal is precise but has low information density and limited UI affordances. He favors self-contained desktop or web-based coding harnesses where agents can work with richer context.
Q2.5 2026 AI timelines update: The AI Futures Project says AI timelines have shortened slightly and its latest update includes coding uplift and revenue anchors. It estimates real-world progress at roughly 70% to 90% of the AI 2027 scenario.
ChatGPT Computer History remembers computer activity: The Verge reports that ChatGPT's opt-in Computer History feature uses activity events to understand a user's workflow, with app and site exclusions and deletion controls. It does not record images, video, or audio, but it creates a trade-off between better context and more shared activity data.
The US asks partners to choose sides in the AI race with China: Reuters reports that the US is expected to ask partners to choose sides in the AI race with China. If implemented, access to compute, models, and technical standards would become a direct geopolitical issue for countries and companies.
Investors hunt for the next AI winners as capex anxiety fades: Reuters describes a market shifting from “Does AI capex work?” to “Who will keep the profit growth once capacity constraints ease?” Attention is expanding from hyperscalers to neoclouds, the deployment layer, and companies that can turn compute into durable profit.
Regulation could erase Anthropic's six-month lead: The All-In Podcast argues that if model releases were regulated like aviation, Anthropic's roughly six-month lead could be eroded, affecting its token premium and valuation. The commentary raises a market question about how long frontier labs can maintain pricing power if regulation slows releases across the industry.
AI makes previously uneconomical markets serviceable: Aaron Levie points to security reviews, contract analysis, and customer upsell as valuable work that was previously too expensive to perform comprehensively. As processing costs fall, many small markets can become large enough to support their own automated workflows.
> ONE PRACTICAL USE OF AI TODAY
Try a 1M-token context window in Codex
Tibo explains how to enable a 1M-token context window in Codex for GPT-5.6 Sol. A larger context lets Codex retain more code, tool output, and conversation history before summarizing older material. The feature requires a supported model, and Codex's defaults have already been tuned for performance and cost.
Open ~/.codex/config.toml and add or update these three lines at the top level, before the other sections:
model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000The first line selects the model. The second sets a one-million-token context budget. The third starts automatic compaction at around 900,000 tokens, leaving some headroom. Save the file, restart the Codex client, and start a new session.
To try the configuration for a single CLI session without changing your defaults:
codex -m gpt-5.6-sol \
-c model_context_window=1000000 \
-c model_auto_compact_token_limit=900000Check which model and version are available in your environment before applying it. Tibo also notes that the default configuration was tuned carefully, so a larger context window is not automatically better for every task.
> WORTH READING
ANALYSIS & THESIS
Simon Willison tests Qwen 3.8 27B, an Apache-2 model with a roughly 17GB quantized version that supports vision and coding agents. He also notes that xhigh can overthink and that local speed remains a bottleneck. The piece is a practical look at the gap between a model being runnable locally and being fast enough for daily use.
MIT FutureTech summarizes more than 60,000 worker evaluations across more than 6,000 tasks and reports that AI reached a minimally sufficient level on roughly 50% to 75% of text tasks in the sample. The piece offers a useful counterweight to both instant-automation hype and the idea that workflow redesign can wait.
Lenny Rachitsky previews a conversation with OpenAI's Head of Design about designer frustration, the AI product designer role, two-speed design, and the continuing role of craft and taste. It is an external reading context for how product teams are redefining design work in the AI era.
