What this thread covers
Frontier model releases tracks the headline-grade model launches, version bumps, and capability demonstrations from the leading AI labs — OpenAI, Anthropic, Google DeepMind, Meta, Mistral, and others — from the GPT-3 era through the present. It is the primary signal for where the capability frontier sits, how fast it is moving, and what the competitive and governance dynamics look like at the top of the stack.
---
Phase 1: The scaling paradigm is established (2020–2022)
The modern frontier begins with GPT-3 (May 2020): a 175-billion-parameter autoregressive model that demonstrated strong few-shot performance across NLP tasks without any gradient updates at inference time. The core finding — that scaling model size dramatically improves task-agnostic capability — became the organizing thesis of the next five years of frontier development.
ChatGPT (November 2022) translated that thesis into a product. By wrapping a dialogue-tuned model in a consumer interface that could acknowledge errors, challenge incorrect premises, and decline inappropriate requests, OpenAI created the first frontier AI system with genuine mass adoption. The release compressed the timeline for every subsequent lab decision about when and how to ship.
---
Phase 2: Multimodality, open weights, and the reasoning turn (2023–2024)
The 2023–2024 period saw three structural shifts run in parallel.
Multimodality. GPT-4o (May 2024) introduced a natively omnimodal architecture — audio, vision, and text processed in a single model rather than piped through separate stages. OpenAI's Sora (Feb 2024) extended the frontier into video generation, framing large-scale video diffusion as a path toward physical world simulation.
Open weights. Mistral's Mixtral 8x7B (Dec 2023) demonstrated that a sparse mixture-of-experts architecture could match GPT-3.5 performance at 12.9B active parameters — a cost-efficiency result that reshaped the open-weights ecosystem. Meta's Llama 3.1 405B (Jul 2024) pushed open weights into frontier-adjacent territory at scale, with multilingual support and extended context.
Inference-time compute. OpenAI o1 (Sep 2024) was the paradigm shift of the period. By training models to reason via chain-of-thought using reinforcement learning, OpenAI introduced inference-time compute as a new scaling axis — distinct from and complementary to training-time scaling. o1-preview ranked in the 89th percentile on competitive programming and performed at PhD level on science benchmarks. This spawned the current generation of "reasoning" or "thinking" models across all labs.
---
Phase 3: Professional-grade capability and the agentic turn (2025)
By mid-2025, the competitive frame had shifted from "can the model answer questions?" to "can the model do work?" — multi-step, tool-using, long-horizon tasks with minimal human supervision.
GPT-5 (Aug 2025) shipped with a unified sub-model routing architecture (gpt-5-main, gpt-5-thinking, lightweight variants), dynamically selecting among sub-models by task. Simultaneously, OpenAI released gpt-oss-120b and gpt-oss-20b under Apache 2.0 — a strategic reversal for a historically closed lab, signaling that open weights had become a competitive necessity rather than a concession.
Claude 3.7 Sonnet (Sep 2025) introduced the first hybrid reasoning model: a single unified model operating in both standard and extended thinking modes, without the user needing to select a separate model endpoint. Claude Opus 4 (Sep 2025) claimed the top coding position with 72.5% on SWE-bench and 43.2% on Terminal-bench, and shipped with parallel tool execution and local file access for persistent agent memory. Claude Code moved from research preview to general availability with GitHub Actions, VS Code, and JetBrains integrations.
Gemini 3 (Nov 2025) marked Google DeepMind's next-generation flagship. Gemini with Deep Think (Oct 2025) achieved gold-medal standard at IMO 2025 — the first externally validated frontier AI result at that level of mathematical competition.
GPT-5.2 (Dec 2025) extended the GPT-5 line with SOTA reasoning, long-context, coding, and vision. In February 2026, a preprint demonstrated GPT-5.2 proposing a novel gluon amplitude formula in theoretical physics, subsequently formally proved — a claimed instance of AI producing genuinely new scientific knowledge.
---
Phase 4: The governance frontier opens (2026)
The first half of 2026 is characterized by continued capability escalation running directly into regulatory and safety governance friction.
Claude Opus 4.5 (Mar 2026) reduced pricing to $5/$25 per million tokens with up to 65% token efficiency gains. Claude Opus 4.6 (Mar 2026) extended context to 1M tokens in beta, introduced adaptive developer-controlled thinking effort, and claimed a 144 Elo lead over GPT-5.2 on GDPval-AA. GPT-5.4 (Mar 2026) matched with a 1M token context window and SOTA coding and computer use.
In May 2026, an OpenAI model disproved an 80-year-old conjecture in discrete geometry — the second documented case of a frontier model producing a verifiable novel mathematical result. Gemini 3.5 (May 2026) repositioned Google DeepMind's flagship around agentic capabilities and complex workflow execution.
Claude Fable 5 and Mythos 5 (Jun 2026) represent the current edge of the frontier — and its most contested deployment. Mythos 5 is a restricted-access model capable of cracking previously secure software. Fable 5 is the general-use version, with novel safety classifiers that block or degrade responses on cybersecurity, biology, chemistry, and AI-development topics. Critically, the initial release included undisclosed capability degradation applied silently via prompt modification or steering vectors — a policy Anthropic subsequently modified after controversy. Both models are priced at roughly half the cost of the prior Claude Mythos Preview.
Within 24 hours of the Fable 5 / Mythos 5 launch, the U.S. government issued an export control directive requiring Anthropic to disable both models for all foreign nationals, citing awareness of a jailbreak method. Anthropic is complying while publicly contesting the standard applied — arguing that requiring perfect jailbreak resistance would halt all frontier model deployments industry-wide. This is the first documented case of a government forcing a commercial frontier model suspension on export-control grounds.
---
Concurrent dynamics: distillation, espionage, and the safety-governance gap
Two events from this period illustrate the security surface that frontier capability creates. In November 2025, Anthropic disclosed the first documented large-scale AI-orchestrated cyber espionage campaign: a Chinese state-sponsored actor used Claude Code as an autonomous agent to attack roughly thirty global targets, decomposing malicious tasks into seemingly innocent subtasks to evade detection. In February 2026, Anthropic publicly identified DeepSeek, Moonshot AI, and MiniMax as conducting coordinated distillation attacks against Claude — over 16 million exchanges through approximately 24,000 fraudulent accounts — targeting agentic reasoning, tool use, and chain-of-thought generation.
These events are not peripheral to the release story. They are the direct consequence of frontier capability reaching the threshold where it is worth stealing and worth weaponizing.
---
Where the frontier is heading
The trajectory from the events in this bundle points in three directions simultaneously: continued capability escalation (agentic autonomy, novel scientific results, mathematical reasoning at competition level); increasing regulatory intervention over who can access the most capable models and under what conditions; and a structural split between safety-tiered restricted-access models and general-use versions — a deployment architecture with no clear precedent in prior software industries. The open question is whether the governance frameworks being improvised in real time can keep pace with the release cadence they are trying to govern.
