What this is about
Frontier model releases are the headline events of the modern AI era: the moments when a major lab ships a new AI system that can do things no publicly available model could do before. This thread tracks those releases — and the broader story they tell about where AI capability is heading, who's driving it, and what tensions are emerging along the way.
Why it matters
For most people, these releases are the AI they actually interact with. ChatGPT, Claude, Gemini — these are frontier models made accessible. Each new generation raises the ceiling on what AI can help with: writing, coding, research, and increasingly, taking actions in the world on your behalf. Understanding the arc of these releases helps you make sense of what AI can do today and what it's likely to do next.
How it started: from research paper to public phenomenon
The modern frontier model story begins with GPT-3 in May 2020 — a 175-billion-parameter model that showed, for the first time at scale, that a single large model could handle a huge range of tasks without being specifically trained for each one. You'd give it a few examples of what you wanted (called "few-shot learning"), and it would figure out the pattern.
But GPT-3 was a research artifact. The moment frontier AI became a public phenomenon was November 2022, when OpenAI launched ChatGPT — a version of their technology wrapped in a simple chat interface. It could hold a conversation, answer follow-up questions, and admit when it was wrong. Tens of millions of people tried it within weeks.
The capability explosion: 2023–2024
The two years that followed saw rapid iteration. Anthropic launched the Claude 3 family — Haiku, Sonnet, and Opus — with strong performance on knowledge benchmarks and a 200,000-token context window (meaning it could read and reason over very long documents). OpenAI released GPT-4o, their first model to handle text, audio, and images in a single unified system rather than separate pipelines. Mistral AI, a French startup, released Mixtral 8x7B — a clever open-weight model that used a "mixture of experts" trick to run as fast as a much smaller model while matching GPT-3.5 on many tasks.
Then in September 2024, OpenAI introduced o1, which changed the conversation about how AI gets smarter. Instead of just training a bigger model, o1 was designed to spend more time "thinking" before it answered — working through problems step by step, like a student showing their work. This "inference-time scaling" became a new axis of competition alongside raw model size.
The expert-level threshold: 2025
By mid-2025, frontier models were crossing thresholds that had seemed distant. OpenAI released GPT-5 in August 2025, claiming state-of-the-art performance across coding, mathematics, writing, and visual reasoning. Anthropic followed in September with Claude Opus 4 and Sonnet 4 — Opus 4 scored 72.5% on SWE-bench Verified, a benchmark that tests whether an AI can fix real bugs in real software repositories. Claude Code, Anthropic's autonomous coding tool, launched as a generally available product at the same time.
Google DeepMind's Gemini model with "Deep Think" reasoning achieved gold-medal standard at the International Mathematical Olympiad in October 2025 — a competition that has challenged the world's best young mathematicians since 1959. This was an externally validated milestone, not just a lab benchmark.
OpenAI also made a surprising move: releasing two open-weight models — gpt-oss-120b and gpt-oss-20b — under the Apache 2.0 license, meaning anyone could download, run, and build on them. For a company that had kept its best models proprietary, this was a significant strategic shift.
AI doing real science: late 2025 into 2026
The next threshold was original scientific discovery. In February 2026, GPT-5.2 proposed a novel formula in theoretical physics — a result that was subsequently verified by researchers. In May 2026, an OpenAI model disproved an 80-year-old conjecture in discrete geometry. MiniMax's MaxProof system scored above the human gold-medal threshold on both IMO 2025 and USAMO 2026 using a tournament-style approach to mathematical proof. These weren't AI systems helping humans do science — they were producing verifiable new results.
The safety and governance reckoning
As capability has grown, so has friction with governments and the public. Anthropic found itself in a prolonged standoff with the U.S. Department of War, which demanded that Claude be made available for fully autonomous weapons and mass domestic surveillance. Anthropic refused, publicly, even as the department threatened to designate the company a "supply chain risk." The dispute revealed how deeply frontier AI is already embedded in national security infrastructure — and how contested the rules of that relationship are.
In June 2026, the U.S. government issued an export control directive requiring Anthropic to suspend access to its newest models — Claude Fable 5 and Mythos 5 — for foreign nationals, citing a jailbreak vulnerability. Anthropic complied while publicly disputing the standard, arguing that requiring perfect jailbreak resistance would effectively halt all frontier model deployments across the industry.
The Fable 5 / Mythos 5 release itself was notable for another reason: Fable 5 initially included undisclosed capability restrictions on AI-development topics, applied silently. When this came to light, it sparked significant controversy, and Anthropic modified the policy.
Where things stand
The frontier today is defined by models that can write and debug software autonomously, reason through complex scientific problems, and take multi-step actions in the world with minimal human supervision. The competition is global, the investment is enormous (OpenAI raised $122 billion in one round; Anthropic raised $65 billion at a near-trillion-dollar valuation), and the questions about who controls these systems — and for what purposes — are no longer hypothetical.
