Every week in AI brings a launch. This week brought something rarer. It brought measurement, and the measurement was not flattering.
I read the week the way I read my own work. Claims are cheap. Shipped things can be checked. Here is what actually happened, and what I think it means for people who build.
A benchmark for work that pays
On November 2, SemiAnalysis launched GDPx, a benchmark built to score AI models on work that people get paid to do. It covers eleven domains, including chip design, cybersecurity, legal work, quantitative trading, and software engineering.
By SemiAnalysis scoring, the best model overall was Claude Opus 5.5, and it scored 12 percent.
Read that number the right way. It does not mean agents are useless. I use them every day, and they save me real hours. It means the distance between a good demo and dependable paid work is still large, and now there is a public scoreboard that says so in plain numbers.
That is healthy. For years the argument about agents ran on anecdotes. A benchmark tied to economic work turns the argument into evidence. Builders can see where models fail, buyers can ask harder questions, and labs have to compete on outcomes instead of applause.
An agent that cannot be measured cannot be trusted with work that matters.
Open weights closed another gap
On November 4, Moonshot AI published the full weights of Kimi K3.1 Pro on Hugging Face, as it had promised. The release uses Moonshot own licence, and businesses selling model access above a revenue threshold need a separate deal. Artificial Analysis puts the model first on Terminal Bench 4.0, which backs the coding claims Moonshot made at launch.
The same day, Tencent launched Hy4. It adds image reading and computer use, and it is available through Tencent own apps and API and on outside platforms. Tencent says the weights will follow in the coming days.
Put those two releases together and the pattern is clear. Capable models are no longer a closed club. A builder who wants control over cost, data, and deployment now has serious options to run and adapt on their own terms. That pressure is good for everyone who builds on top of models, including me. I build Manifest Anything on the Manifest Engine and our own AI models, and widely available strength raises the bar for what a platform has to add on top.
The sharpest tools went to defenders first
On November 5, OpenAI released GPT-6 Cyber, a security focused version of GPT-6 Astra. Access is limited to a select group of approved Daybreak Red customers through the API and Codex. OpenAI rates the model Critical for cybersecurity capability. Its own system card notes that an outside tester, Irregular, found it broadly comparable to the base model, with better results on long attack tasks.
I like the shape of that release. A model that can find vulnerabilities and build exploits is dual use by nature. Giving it first to vetted defenders, in a controlled channel, treats capability as something to steward. The rest of the industry should copy the pattern even when they do not copy the product.
CrowdStrike made a matching move on the oversight side. On November 3 it announced a definitive agreement to acquire Fiddler AI, and it plans to sell the technology as Falcon Observe, covering evaluation before deployment, agent tracing, behaviour monitoring, and audit evidence. Governance is becoming a product category, not a slide in a pitch deck.
Video keeps getting cheaper to make
Google DeepMind released Gemini Omni 1.2 Flash on November 5. Google says it can generate and edit continuous video clips up to 15 seconds, with improved scene consistency and character speech, and that drafting at 360p or 720p costs $0.07 per second, with the option to upscale finished work to 1080p or 4K.
Fifteen seconds sounds small until you price a campaign. Short clips are the unit of modern marketing and product demos. When the draft costs cents, the bottleneck moves from budget to taste. The people who win will know what to make, not just how to make it.
Code and control consolidated
On November 4, OpenAI announced it will acquire Pierre Computer Company, the firm behind Code Storage. Pierre open source tools, Trees and Diffs, already power file browsing and change review inside Codex. The deal has not closed, and it remains subject to regulatory approval.
The signal matters more than the size. OpenAI is buying the layer where humans review what agents write. As agents produce more code, the review surface becomes the product. Whoever owns the place where a person reads a diff and approves it owns the trust point of the workflow.
What I take from this week
Three lessons stay with me. First, measure agents against paid work, not against applause. Second, expect open weights to keep closing the gap, and build your value above the model layer. Third, expect the most capable tools to arrive gated, monitored, and audited, and treat that as a feature.
I build in public, so my own numbers stay open to check. My Threads posts drew 716,310 views across the trailing 30 days, and my Instagram audience stands at 77,412 followers. Proof beats promises for a founder too.
Next week will bring more launches. I will be asking the same questions. What shipped. Who verified it. What work can it actually do.
If your conference or team is working through what these shifts mean in practice, I speak on building with AI in the open. Book me for a keynote here.
