Now: 2024–2026
The current frontier is models that reason before answering, use software on a person’s behalf, and increasingly arrive as open weights. Entries here are recent, so they are the ones most likely to be revised as the record settles.
It learned to think before speaking
OpenAI released o1, a model trained using reinforcement learning to "think" through problems before generating an answer. By generating hidden chains of thought, it achieved breakthrough performance on complex math, coding, and logic tasks that had stumped previous models.
Why it mattered. It proved that scaling test-time compute could yield massive intelligence gains, shifting the industry's focus from simply training larger models to giving them time to reason.
openai.comIt learned to use a computer
Anthropic introduced a "computer use" capability for Claude 3.5 Sonnet, allowing the model to look at a screen, move a cursor, click buttons, and type text. Rather than interacting through APIs, the model was trained to operate standard desktop software exactly as a human would.
Why it mattered. It marked the transition from models that merely answered questions to agents that could autonomously execute multi-step workflows across arbitrary applications.
anthropic.comReasoning went open source
Chinese AI lab DeepSeek released DeepSeek-R1, an open-weights reasoning model that matched the performance of OpenAI's o1. By publishing their full training methodology and releasing the weights, they demonstrated how to distill complex reasoning capabilities into highly efficient models.
Why it mattered. It shattered the assumption that frontier reasoning capabilities would remain locked behind proprietary APIs, accelerating global open-source development and forcing a re-evaluation of AI compute costs.
api-docs.deepseek.comThe frontier disappointed
OpenAI released GPT-5, a massive multimodal model marketed as having "PhD-level" intelligence. Despite the hype, the launch faced severe user backlash over its overly sanitized personality and perceived regressions in helpfulness compared to GPT-4o.
Why it mattered. It demonstrated the limits of scaling laws when constrained by aggressive safety alignment, forcing OpenAI to backtrack and reintroduce older models to appease users.
fortune.comThe video dream died
OpenAI discontinued its flagship video generation model, Sora, shutting down the web and app experiences in April 2026 and scheduling the API for deprecation. Despite the initial hype around Sora 2, the immense compute costs required to generate high-fidelity video proved economically unsustainable.
Why it mattered. It was a stark reality check for generative AI, proving that technical feasibility did not guarantee a viable product and forcing the industry to reckon with the staggering costs of multimodal generation.
help.openai.comFeed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.