What this area covers
Frontier model releases are the headline events of the modern AI era: the moments when a major lab ships a new AI system that meaningfully expands what machines can do. This thread tracks those releases — and the scientific, commercial, and regulatory shockwaves that follow — from the foundational GPT-3 paper through the safety-tiered deployments of mid-2026.
Why it matters
If you use any AI tool today — a chatbot, a coding assistant, an image generator — it almost certainly traces back to one of the releases in this thread. These models set the ceiling for what AI can do, and that ceiling has risen faster than almost anyone predicted. They also increasingly set the terms of geopolitical competition, regulatory debate, and scientific discovery.
Phase 1: The language era (2020–2023)
The modern story starts with GPT-3 in May 2020 — a 175-billion-parameter model that showed, for the first time at scale, that a single large AI could handle a huge variety of tasks without being specifically trained for each one. You could give it a few examples of what you wanted (called "few-shot learning") and it would figure out the pattern.
That research insight became a product moment in November 2022, when OpenAI launched ChatGPT. The conversational format — ask a question, get an answer, follow up, push back — made AI feel accessible to everyone, not just researchers. Public adoption was unlike anything the tech industry had seen.
Mistral AI entered the picture in December 2023 with Mixtral 8x7B, an open-weight model (free to download and modify) that matched or beat GPT-3.5 on many tasks while running far more efficiently. It signaled that the frontier wasn't exclusively the domain of well-funded American labs.
Phase 2: Seeing, thinking, and acting (2024)
Three releases defined 2024. First, GPT-4o (May 2024) made AI natively multimodal — it could process text, audio, and images in a single unified system, rather than routing through separate pipelines. Second, Meta's Llama 3.1 (July 2024) pushed open-weight models to 405 billion parameters, putting frontier-class capability in the hands of anyone with the hardware to run it. Third — and perhaps most consequential for the long run — OpenAI's o1 (September 2024) introduced a new idea: instead of just making models bigger, you could make them think longer at inference time, using chain-of-thought reasoning to work through hard problems step by step. This "inference-time scaling" opened a new axis of improvement that didn't require ever-larger training runs.
Phase 3: Agents take over (2025)
By 2025, the question had shifted from "can AI answer questions?" to "can AI do things?" The answer, increasingly, was yes.
Anthropic's upgraded Claude 3.5 Sonnet (July 2025) introduced computer use — the ability for an AI to look at a screen, move a cursor, click, and type, just like a human at a keyboard. OpenAI shipped GPT-5 (August 2025), claiming state-of-the-art performance across coding, math, writing, and vision. In a notable strategic shift, OpenAI also released two open-weight models — gpt-oss-120b and gpt-oss-20b — under the permissive Apache 2.0 license, the first time the company had made models freely available for anyone to run and modify.
Claude Opus 4 and Sonnet 4 (September 2025) pushed software engineering benchmarks further, with Opus 4 scoring 72.5% on SWE-bench — a test of real-world coding tasks. Claude Code, Anthropic's autonomous coding tool, became generally available with integrations into GitHub, VS Code, and JetBrains.
Then came a string of results that would have seemed like science fiction a few years earlier. Gemini with Deep Think achieved gold-medal standard at the International Mathematical Olympiad (October 2025). GPT-5.2 derived a genuinely new formula in theoretical physics, subsequently verified by researchers (February 2026). An OpenAI model disproved an 80-year-old conjecture in discrete geometry (May 2026). MiniMax's MaxProof system scored above the human gold-medal threshold on both IMO 2025 and USAMO 2026 (June 2026). AI was no longer just assisting with known science — it was producing new science.
Phase 4: Safety tiers and government friction (2026)
The most recent releases have introduced a new layer of complexity: not just what a model can do, but who gets access to which capabilities under what conditions.
Anthropic's Claude Mythos 5 and Fable 5 (June 2026) represent the clearest example. Mythos 5 is a restricted-access model capable of cracking previously secure software. Fable 5 is the general-use version, but it ships with novel safety classifiers that silently degrade responses on sensitive topics — cybersecurity, biology, chemistry, AI development — via prompt modification or steering vectors. The undisclosed nature of this degradation sparked controversy before Anthropic revised the policy. Within a day of launch, the U.S. government issued an export control directive requiring Anthropic to suspend both models for all foreign nationals, citing awareness of a jailbreak technique. Anthropic complied while publicly disputing the standard applied, arguing that requiring perfect jailbreak resistance would effectively halt all frontier model deployments.
This followed months of tension between Anthropic and the U.S. Department of War over whether Claude could be used for mass domestic surveillance and fully autonomous weapons — uses Anthropic refused to enable regardless of government pressure.
Where it's heading
The trajectory points toward AI systems that are less like tools you query and more like colleagues you assign work to — capable of running for hours, using software, and producing results that require expert verification rather than expert creation. The binding questions are no longer purely technical: they are about who controls access, what uses are permitted, and how governments and companies negotiate those boundaries in real time.
The pace of releases shows no sign of slowing. The gap between a model's technical capability and society's ability to govern it is, if anything, widening.
