OpenAI Breaks Free: Microsoft Loses Exclusive Access to Its Own AI Partner
For the past few years, Microsoft and OpenAI have been one of the most powerful alliances in tech.
Microsoft poured billions in. OpenAI got the compute, the cloud, and the credibility. Microsoft got equity, a revenue share, and exclusive access to the most powerful AI models on the planet.
It was a deal that reshaped the entire industry.
On April 27, that deal quietly changed and the implications are enormous.
OpenAI and Microsoft announced an amended partnership that ends Microsoft’s exclusive grip on OpenAI’s technology. Going forward, OpenAI can serve its products on any cloud provider. AWS. Google Cloud. Anyone.
Microsoft’s license to OpenAI’s models continues until 2032 but it’s no longer exclusive. And in a move that surprised many, Microsoft will no longer pay a revenue share to OpenAI. OpenAI will still pay Microsoft, but now with a cap, and independent of whether OpenAI ever reaches AGI.
Microsoft shares dipped nearly 3% on the news.
What triggered this? Earlier this year, OpenAI signed a $50 billion deal with Amazon. That deal created a direct conflict with Microsoft’s exclusive rights. This amendment resolves that tension and gives OpenAI the freedom it needs to scale across every major cloud platform.
Why this matters?
This isn’t just a contract update.
It signals that OpenAI has outgrown the partnership that built it. The company is reportedly eyeing an IPO as early as late 2026 and to go public, it needs to prove it isn’t locked to a single infrastructure partner.
For the broader AI ecosystem, this is huge. OpenAI’s models becoming available on AWS and Google Cloud means more competition, more choice, and likely lower prices for developers and businesses over time.
The most important AI alliance of the last five years just got a lot looser. And the race to own AI infrastructure just got a lot more open.
OpenAI Just Launched GPT-5.5 – and It’s Built to Work, Not Just Answer
One week after Anthropic released its strongest model, OpenAI responded.
On April 23, OpenAI launched GPT-5.5 internally codenamed “Spud” rolling it out immediately to paid ChatGPT and Codex subscribers.
The pitch is different this time. OpenAI isn’t leading with how smart this model is. They’re leading with how much it can do on its own, with minimal guidance.
“What is really special about this model is how much more it can do with less guidance,” said OpenAI President Greg Brockman. “It can look at an unclear problem and figure out just what needs to happen next.”
That’s not a small claim. Most AI tools today still require you to break down tasks, write detailed prompts, and babysit every step. GPT-5.5 is designed to skip that. You hand it a messy, ambiguous goal and it figures out the path forward by itself.
On benchmarks, the results back it up. GPT-5.5 leads all publicly available models across 14 benchmarks, including terminal-based task completion, cybersecurity challenges, and advanced mathematics. On Terminal-Bench 2.0 a test for navigating real computer environments it scored 82.7%, ahead of every publicly accessible rival.
API access isn’t available yet. OpenAI says it needs additional cybersecurity safeguards before opening it to developers. When it does arrive, pricing will roughly double GPT-5.4 though OpenAI argues improved efficiency means you’ll need fewer tokens to get the same output.
Why this matters?
We’re watching a shift in what AI is actually for.
The first wave of AI tools was about generating content writing emails, summarizing documents, answering questions. GPT-5.5 is designed for something different: executing multi-step work inside real software environments with minimal human hand-holding.
This is what “agentic AI” actually looks like in practice. Not a chatbot. Something closer to a digital coworker.
And the timing of this launch exactly one week after Anthropic’s Opus 4.7 tells you everything about how fast this race is moving. These companies are no longer on quarterly release cycles. They’re watching each other in real time and responding within days.
Anthropic Released Opus 4.7 – and It’s Winning Where It Counts
Before GPT-5.5 arrived, there was a new leader on the leaderboard.
On approximately April 16, Anthropic quietly released Claude Opus 4.7 its most capable model available to the public. No flashy event. No keynote. Just a release that immediately reshuffled the benchmark rankings and sent OpenAI back to the drawing board.
And the numbers are striking.
On Humanity’s Last Exam one of the hardest reasoning benchmarks in existence, designed to push models to their absolute limits Opus 4.7 scored 46.9%. GPT-5.5, released a week later, scored 43.1% on the same test.
That gap matters. Because Humanity’s Last Exam doesn’t test how fast a model types or how confidently it sounds. It tests whether the model actually knows what it’s talking about no tools, no search, no shortcuts.
But Opus 4.7 isn’t just about benchmarks. Anthropic made a specific change that developers and power users will immediately notice: they significantly reduced sycophancy the tendency of AI models to agree with you even when you’re wrong, tell you what you want to hear, and soften criticism to avoid pushback.
It also introduced extended thinking and hybrid reasoning, making it more capable for deep analysis and multi-step problem solving.
Why this matters?
For a long time, the narrative was simple: OpenAI leads, everyone else catches up.
That narrative is over.
Anthropic is no longer just a safety-focused alternative to OpenAI. It’s a direct competitor winning on capability, enterprise adoption, and developer trust all at the same time.
The week of April 16 to April 23 told the whole story. Anthropic released the best publicly available model. OpenAI responded days later with a direct counterstrike. Both companies are now so close in capability that the competition has shifted to something harder to measure: reliability, pricing, trust, and which platform developers actually want to build on.
That’s a much more interesting race than benchmarks alone.
Liked this issue? Get the next one free.
One email on Tuesday. Unsubscribe in one click.