What GPT-5 and Opus 4.1 quietly reveal about our progress curve

Howdy my friend!

In today’s piece:

• My reaction to GPT-5 and other recent news

• The signals that pushed my AGI timeline out—and why that’s good news

• Google’s Genie 3 & the implications

• Other handpicked things I think you’ll like (including updates to the AI agent ecosystem within AWS and Azure, the state of bug-bounty programs in the new AI world, etc)

It’s been another blockbuster week in AI development.

Lots of awesome progress but surprisingly, my AGI timelines moved further out. And I’m oddly relieved. More on that shortly.

Forgive my late email. I’ve been distracted with another AI experiment of mine that got some unexpected traction.

I couldn’t resist chasing that shiny object but more on that another time.

Let’s unpack my initial impressions of the last week or two in AI.

The headlines were huge. But beneath the hype, the current AI trajectory looks less like “hyperexponential explosion” and more like steady, compounding gains.

Here’s why I reached that conclusion.

Let’s start with the big headline item: GPT-5. Years in the making, it finally landed—and it’s excellent.

The new system isn’t just “smarter”. It’s a unified system that routes your request to the right brain automatically.

Quick questions get a zippy responder; hard problems get handed to the deeper reasoning engine (GPT-5 Thinking).

No more second-guessing which of ten model SKUs you should pick.

That simplification alone is a massive UX win. Honestly, the old maze—o3 vs o4 vs 4.1 vs mini-this and high-that—was a nightmare for casual users. GPT-5 cleans that up.

The other win: cost and efficiency.

By the API, GPT-5 lands at roughly a twelfth the price of Anthropic’s Claude Opus 4.1 and now competes toe-to-toe with Google’s Gemini 2.5 Pro (even edging it for long inputs).

It also arrives with two lean variants—GPT-5 mini and nano—that punch above their weight on intelligence-per-dollar, with nano giving Gemini 2.5 Flash a real scare on value.

If you like charts, Artificial Analysis has clear graphs on price, latency, and quality across models.

Speed-wise, GPT-5 is snappy, especially when your task doesn’t need full deep-think mode.

Quality-wise, it’s now the consensus leader and tops the LMArena leaderboard—which reflects real user preference, not just benchmark gymnastics. OpenAI delivered.

And yet… I wouldn’t call GPT-5 a paradigm shift. It’s a brilliant step up in cost, efficiency, and user experience.

It’s not a brand-new way of doing AI.

Anthropic’s Claude Opus 4.1 tells a similar story. It’s a strong frontier model, but the measurable gains (around ~2% on SWE-bench Verified) are incremental, not epoch-making.

The vibe from the frontier labs is “squeeze more juice from what works”—the Transformer + test-time reasoning playbook—rather than unveil a new recipe.

Tech progress tends to run in S-curves: an explosive climb when a new paradigm hits, then a slower middle stretch as we approach the ceiling of that approach.

Right now, it feels like we’re in that middle stretch for reasoning-via-test-time compute (the o1/o3 lineage).

To be clear, there’s still plenty of low-hanging fruit ahead—tool use, agents, memory, retrieval, safety rails, orchestration. But the shape looks more “gradual staircase with occasional hops” than “rocket to AGI next Tuesday.”

Could a fresh paradigm appear in the next couple of years? With this much talent and capital, I wouldn’t bet against new ideas. I’m just less convinced the next 5–10 years deliver full-blown AGI.

And I’m okay with that. Frankly, it’s one of the better timelines.

Safety concerns from folks like Geoffrey Hinton and Yoshua Bengio are not tinfoil. They’re prudent.

A world of quick, steady progress beats a surprise intelligence detonation.

One exception that did feel “whoa”: Genie 3, Google’s general-purpose world model.

The demo shows real-time, interactive 3D worlds spun up from prompts.

Beyond games, this has big implications for training embodied systems—think simulation-heavy robotics.

It’s not public yet, but if you like glimpses of the next “S,” this is one to watch. The official demos are here.

So where does that leave us a cloud security or security-adjacent professionals?

I’m hoping we get more breathing room to properly secure this new tech, and to understand how it’ll impact the world.

I feel like the “fog of war” in AI progress has lifted somewhat these last few weeks. We’ve got a better idea of what trajectory we might be on, but there’s still a lot of unknowns.

How did the last two weeks move your AGI timeline—forward, backward, or unchanged? Hit reply and tell me why (and what you’re building because of it).

My Favourite Finds

GPT-5: Official OpenAI blog post about their newest model (although “system” is probably a better word than model these days, given how central techniques like Mixture of Experts and real-time routing is to this type of AI)

AI Explained’s video analysis of GPT-5: my favourite AI-fluencer’s breakdown of GPT-5 and the implications. Definitely worth a watch.

Claude Opus 4.1: Official Anthropic blog post about their latest frontier model.

Genie 3: Google’s write up on Genie 3, their new general-purpose world model (discussed in my ramblings above).

Genie 3 video analysis: another shoutout to AI Explained who also covered Genie 3. Again, highly recommended viewing.

Cloud Security Alliance releases AI Controls Matrix (AICM): The CSA’s AICM provides 243 control objectives across 18 domains covering identity and access management, data security, model security and supply‑chain management.

AI slop floods bug‑bounty programs with fake vulnerability reports: TechCrunch reports that large‑language‑model‑generated “AI slop” is overwhelming bug‑bounty platforms. Every new tech has a flip side… 🤦🏼

Prowler’s State of Cloud Security Report: AI adoption surges: Based on a survey of 650 security professionals, Prowler found that almost four‑in‑five cloud‑security teams now use AI and ML to monitor their environments; 44 % said AI augments human analysts, 42 % noted improved data protection, and 38 % cited stronger threat detection.

AWS unveils Amazon Bedrock AgentCore for secure AI agents: At AWS Summit New York, Amazon announced AgentCore, a suite of services that lets organizations deploy AI agents securely and more “enterprise-grade-ily”.

Azure AI Foundry July 2025 update adds research agent, AgentOps and AI Red Teaming Agent: Microsoft introduced a Deep Research Agent (public preview) that combines Bing search with OpenAI’s o3-deep-research model to automate complex web research. New AgentOps tools provide tracing, evaluation and monitoring of agents, while an AI Red Team Agent lets security teams run adversarial testing.

Before you go, I’d love to know what you thought of today’s newsletter. That way I can improve the experience for you.

Take care.

Nelson, fellow AI enthusiast and cloud security dude.