It’s been a busy week at OpenAI. A model update, a funding push, a court order, a cross-cloud pivot, a 10-hour outage, and a blog post where Sam Altman outlines the future as he sees it. Let's get into it:
(1) The o3‑pro release: reasoning becomes a billable unit
This week, OpenAI continued its chaotic-neutral naming streak and launched o3‑pro, now billed as its most capable model yet.
The core pitch here is that o3‑pro is more reasoning-capable than o3, more accurate, and more stable under load - i.e., the kind of model you’d want writing code or summarizing legal briefs, not drafting wedding toasts or pretending to be a cat on Reddit
So what is o3‑pro?
Technically, it’s a higher-compute sibling of o3: slower, smarter, and selectively available (currently for ChatGPT Pro and Teams, with Enterprise and Edu coming soon). It handles tools - browsing, file uploads, code execution, vision input, and memory - and scores competitively on key benchmarks, outperforming Claude 4 Opus on GPQA Diamond (PhD-level science) and beating Gemini 2.5 Pro on AIME 2024 (advanced math), per OpenAI’s internal testing.
Human reviewers consistently preferred it over o3 across domains like science, education, business, and writing - scoring it higher on clarity, instruction-following, and factual accuracy. In other words: fewer hallucinations, more hand-holding.
But the model isn’t flawless: it’s slower than o1‑pro, it can’t generate images or work with Canvas, OpenAI’s new collaborative workspace. Also, temporary chats are disabled due to technical issues. Which is to say: still in beta.
But the model itself is only half the story.
This release draws a deliberate line between consumer-facing AI and production-grade AI.
The base o3 has now been repriced, 80% cheaper for API access
o3‑pro, by contrast, comes with significantly higher pricing ($20 per million input tokens, $80 per million output)
It’s a clear signal of differentiated positioning and a not-so-subtle nudge to enterprises: if you want predictable output under real-world conditions, you’ll need the premium SKU.
By splitting the model line, OpenAI is doing what every infrastructure company eventually does: Establish pricing tiers tied to value and use case. Segmentation.
This release implicitly communicates: If you want cheap, fast, and “good enough,” use o3. If you’re doing complex workflows, multi-agent orchestration, or internal tooling - you need o3‑pro.
You’re no longer buying “the best AI.” You’re buying predictability, latency trade-offs, and contractual stability. In other words, AI is being packaged like cloud compute: tiered SKUs, enterprise billing, and a tacit understanding that if you want the good stuff, you’re paying for priority lane access.
OpenAI is building the billing system for a world in which “thinking” is metered.
(2) The Google Cloud deal: friendship is a GPU abstraction
OpenAI, famously backed by Microsoft to the tune of $13 billion, is now also… working with Google Cloud.
At first glance, this reads like a violation of some exclusivity clause or at least a test of corporate loyalty. But the reality is simpler: they need GPUs. As Sam Altman put it, their infrastructure was “melting.” So they did the rational thing: they asked their frenemy for help.
This partnership - alongside existing deals with Oracle, CoreWeave, and SoftBank’s Stargate - cements something important: you can’t afford tribal loyalty in AI.
OpenAI’s alliance with Microsoft might govern its default hosting stack, but compute scarcity overrides loyalty. In this market, your primary cloud is whoever can get you online today. The bottleneck isn’t ideology, it’s latency and access to H100s.
What’s also happening here is the quiet normalization of multi-cloud AI. Not in the nice, hybrid-cloud marketing-deck sense, but in the “survival-mode workload routing” sense.
And for Google? The move is strategically ironic. The company that built TPU pods to train its own models is now selling compute to the one startup threatening its dominance in search and LLMs. But this isn’t charity - it’s leverage. Every cycle OpenAI runs on Google infrastructure gives them a seat, however small, inside the most important AI company in the world.
In this phase of AI, everyone works with everyone until they can afford not to.
(3) The fundraise: OpenAI’s cap table is now a map
For the second installment of its $40 billion mega-round led by SoftBank, OpenAI is in talks with a who's-who of sovereign and strategic capital: Saudi Arabia’s Public Investment Fund, Reliance Industries in India, Abu Dhabi’s MGX (already a shareholder), SoftBank, Coatue, Founders Fund, and others.
The motivations are strategic, not just financial. SoftBank wants exposure to the most consequential AI company on the planet. PIF wants a direct seat at the frontier after years of investing indirectly through U.S. funds. Reliance wants to distribute OpenAI’s tech through India’s largest telecom and enterprise network. And OpenAI? It wants capital and control over the physical substrate of intelligence.
This is no longer venture capital. It’s geo-economic capital allocation.
(4) The outage: 10 hours of silence
On June 10, ChatGPT went dark.
For roughly 10 hours, users experienced widespread issues across ChatGPT, the API, and other OpenAI services. Web access failed. Latency spiked. Temporary chats disappeared. It was the longest and most disruptive outage OpenAI has had in months.
The timing couldn’t be worse- coming right as OpenAI was onboarding enterprise use cases and selling the reliability of o3‑pro. If you’re asking to be embedded into workflows, you can’t afford to drop the ball like a consumer product.
The root cause hasn’t been publicly detailed, but it doesn’t really matter. The lesson isn’t in the logs. It’s in the optics. This was a reminder that we’re building on sand.
OpenAI is quickly becoming a default cognitive layer for consumers, students, and enterprises alike. When it goes down, so do workflows. So do expectations.
(5) The blog post: building softly into the singularity
Amid the model upgrades, GPU deals, and capital maneuvers, Sam Altman also dropped a blog post this week titled “The Gentle Singularity” It reads like a calm, almost meditative forecast. But beneath the tone is a radical thesis about where we’re headed - and how fast.
Some key claims:
2025: AI agents capable of real work - coding, reasoning, planning
2026: Systems that contribute original scientific insights
2027: Embodied robots performing complex real-world tasks
By 2030: A single person could do the work of many across decades
The point isn’t just that AI is improving. It’s that productivity itself is being detached from headcount - and reattached to capital, compute, and model access.
Altman describes this as a “gentle” singularity. Not a runaway intelligence explosion, but a steady inversion of limits. Energy and intelligence, once bottlenecks, become abundant. Cost curves collapse. AI becomes infrastructure.
And to counter growing concerns about AI’s energy footprint, he throws in this stat:
The average query uses about 0.34 watt-hours, about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes. It also uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon
In fact, his broader claim is that energy and intelligence are converging as scalable, near-infinite resources. The real bottleneck becomes how we organize around them - politically, economically, socially.
The singularity may be gentle. The power dynamics won’t be.
Notable mentions to events that didn’t make the top 5.
OpenAI is appealing a court order in the New York Times copyright suit that would force it to indefinitely preserve every ChatGPT output. The legal system wants a paper trail; OpenAI argues it compromises user privacy. What’s really at stake is who controls AI memory: the company, the courts, or the user? The subtext: what you type into ChatGPT could someday become legal evidence - unless policies, or precedents, change. In hindsight, “Dear ChatGPT, am I the problem?” was a risky opener.
Mattel is integrating OpenAI’s tech - like ChatGPT Enterprise- into its toys and internal workflows, aiming to launch an “AI-powered” product by late 2025 that brings smarter, interactive play to Barbie, Hot Wheels, UNO, and more. ChatGPT Barbie was not on my 2025 bingo card, but here we are.
And the week’s not even over.
For a company trying to align models, OpenAI remarkably good at keeping us off balance.