The King is Back. o3 & o4-mini are ELECTRIC! Can Google Compete?

OpenAI’s new o3 and o4-mini models are being positioned as a major leap in “agentic” AI—systems that can plan, use tools (web search, Python, terminal), and reason through multimodal inputs to produce results that look closer to scientific workflows than chat-only answers. Early access through paid ChatGPT tiers highlights a practical shift: the models can think for minutes, pull external data, run code, generate charts, and then deliver bottom-line recommendations with supporting analysis.

The core distinction is scale versus efficiency. o3 is the larger, “state-of-the-art” model aimed at pushing capabilities in coding, math, science, and visual perception. External evaluations cited in the discussion credit o3 with fewer major errors than OpenAI’s earlier o1 on difficult real-world tasks, with particular strength in programming, business/consulting-style reasoning, and creative hypothesis generation. o4-mini, meanwhile, is framed as the next-generation smaller model—optimized for fast, cost-efficient reasoning—while still delivering “remarkable performance for its size.” The naming is treated as confusing but meaningful: o4-mini replaces o3-mini and is described as a glimpse of the next generation, while the full o4 is implied to be not yet ready.

Benchmarks presented as the centerpiece show tool use as the accelerant. On AIM 2024 and AIM 2025, the scores climb sharply when Python and web tools are enabled, with o4-mini reaching the top end of the charts in the tool-augmented setting (including scores near 99.5 and 99.5-level performance). Code-focused results are similarly framed as dominant: terminal-enabled o3 and o4-mini lead on Codeforces-style competition metrics, with o4-mini outperforming the larger o3 while remaining cheaper.

The discussion also contrasts academic-style tests with exam-style tasks. On GPQA Diamond, o3 and o4-mini show slight leads over older models without tools. On “Humanity’s Last Exam,” tool access again drives the gap: o4-mini improves substantially with Python/browsing, and o3 with tools rises even further—though a specialized “Deep Research” system still tops the chart, described as powered by a fine-tuned variant of o3. Multimodal performance is repeatedly emphasized as improved, including image-based reasoning where the models can interpret diagrams or whiteboards and then manipulate images (rotate/zoom/transform) to extract hard-to-read details.

Cost and usability are treated as the other half of the story. API pricing comparisons claim o4-mini is dramatically cheaper than o3, while staying close to Google’s Gemini 2.5 Pro on many benchmarks. The narrative repeatedly returns to the idea that tool-using o-series models saturate performance on certain benchmark suites—making price and workflow fit the deciding factors.

Community testing and real-world behavior add texture. Users report strong performance on a hexagon-and-bouncing-balls physics benchmark, with several older models failing or glitching. At the same time, jailbreak attempts appear to work, including a user producing a detailed amphetamine synthesis procedure after bypassing safety. Coding experiments in ChatGPT’s interface are more mixed: the models can generate and verify smaller code tasks, but large projects (like a Minecraft clone) struggle, with output limits and canvas constraints suspected.

Finally, OpenAI’s “Codeex CLI” is introduced as an open-source coding agent that runs on a user’s computer, aiming to make tool-using models more practical for developers. The overall takeaway is that o3 and o4-mini aren’t just smarter—they’re being trained to choose tools, reason about them, and handle visual inputs in ways that make them feel more like autonomous problem-solvers than conventional chatbots.