The Signal — July 25, 2026
Anthropic put a new frontier model into everyone's editor the same day it announced it, Nvidia and SK Group attached half a trillion dollars to the memory bottleneck, and a rare US–China document agreed on something. Here's what moved.
Anthropic ships Claude Opus 5
Anthropic released Claude Opus 5 on July 24, and unusually, you could use it that afternoon. The model showed up in GitHub Copilot and on AWS Bedrock the same day, not weeks later. Anthropic pitches it as a "thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price," aimed at agentic coding and computer-use tasks, and it keeps the existing Opus token pricing rather than charging a premium for the jump.
Worth separating what's verified from what's marketing. The availability and pricing are confirmed across Anthropic, GitHub, AWS, and independent coverage. The performance claims, including the "near Fable 5 at half the cost" framing and the prompt-injection resistance numbers, are Anthropic's own evaluations. Those tend to hold up better than most vendor benchmarks, but they haven't been independently reproduced yet. If the price-to-capability ratio is anywhere near what Anthropic says, the interesting part isn't the leaderboard, it's that frontier-class agentic work gets meaningfully cheaper to run at volume.
Sources: Anthropic · GitHub Copilot · The Decoder · AWS Bedrock
Nvidia and SK Group put $500 billion behind the memory bottleneck
Nvidia and South Korea's SK Group announced a more than $500 billion AI initiative on Friday, and the shape of it says a lot about where the constraint has moved. Alongside large-scale data centers, the centerpiece is an SK Hynix partnership to lock in next-generation memory, with joint development of future HBM and a planned 2GW cloud buildout whose first facility is targeted for 2027. Compute has not been the only ceiling for a while now; high-bandwidth memory supply is, and this is Nvidia trying to buy certainty on that front years ahead.
The number deserves a caveat. This is announced and planned capacity, not delivered infrastructure, and the headline figure spans a multi-year initiative rather than a check being written this quarter. Mega-partnerships like this have a way of getting quietly rescaled once the build actually starts. Still, the direction is clear: the companies that control memory supply are becoming as strategically important as the ones designing the chips.
The US and China agreed on open models, sort of
The US, China, and other APEC economies released a joint digital and AI ministerial statement on July 23 backing trusted open-source approaches, while carving out the usual caveats around security, data protection, and intellectual property. Getting Washington and Beijing to co-sign anything on AI is notable on its own, and the fact that the common ground is open models rather than restriction cuts against the assumption that both capitals are drifting toward tighter control.
It's a nonbinding ministerial statement, not law and not a treaty. Nobody is obligated to do anything, and the security and IP caveats leave plenty of room for either government to act however it wants later. But these statements shape the language that harder policy gets built on, and right now that language leans toward keeping open models legitimate.
On the Editor's Desk
We held a few things. Amazon closed its San Francisco AGI Lab this week amid layoffs in its AGI unit, but the story is a few days old and the company says top-model work continues elsewhere, so we're sitting on it rather than leading with a stale layoff item. OpenAI's ChatGPT Health rollout and Anthropic's Claude voice-mode update both ran yesterday. A UK and US evaluation of Kimi K3's cyber capabilities looked interesting, but we ran a very similar frontier cyber-eval story two days ago and it read as a repeat. Several research papers on agent training and long-context inference were solid but too narrow for today.
