OpenAI dropped GPT-5, is AGI here?

The release of OpenAI’s GPT-5 series marks a significant step forward, but the immediate question for many developers is whether it surpasses Claude Opus 4.1 as the premier coding assistant. The initial consensus from the hosts is a clear “no.” Despite impressive advancements, key aspects of the developer workflow keep Claude in the lead as their daily driver.

The GPT-5 Release: A Focus on Utility and Accessibility

The GPT-5 release introduced a family of three models—Core, Mini, and Nano—fronted by a model router to simplify the user experience. The launch was characterized by three main themes:

1. A Simplified Experience: The introduction of a model router consolidates the complex array of previous models under the single GPT-5 brand, making it easier for users to access the right capability without confusion.
2. Reliability Over Raw Intelligence: While benchmarks show GPT-5 is smarter, the leap isn’t as dramatic as previous releases. Instead, the major focus was on enhancing reliability and significantly reducing hallucinations, making the model more trustworthy for professional tasks.
3. Aggressive Pricing and Accessibility: In a surprising move, OpenAI priced the models very competitively, particularly the APIs. This strategy lowers the barrier to entry and pushes the technology towards being “too cheap to meter,” making state-of-the-art AI more accessible to everyone.

Enhanced Agentic Capabilities and Reasoning

A standout feature of GPT-5 is its marked improvement in agentic tasks. The model demonstrates superior performance in:

• Tool Invocation and Function Calling: It is significantly more reliable in calling tools, functions, and generating structured output, which is crucial for building robust AI agents. The Nano model, in particular, was highlighted as a “killer” for these tasks, outperforming many larger models.
• Planning, Logic, and Reasoning: The model shows a deeper capacity for complex reasoning. Chris Hay shared an anecdote of GPT-5 successfully solving the “Murdle” detective game, a task that stumped all previous versions. This points to a heavy investment in improving the model’s planning and logical deduction abilities.
• Handling Long-Horizon Tasks: The model demonstrates a more consistent ability to successfully complete complex tasks that unfold over a longer duration.

The Path to AGI: Architectural Clues from gpt-oss

The hosts believe this release suggests the path to AGI will be a “slog” of foundational improvements rather than a sudden explosion. Insights from the related gpt-oss release reveal key architectural directions:

• Mixture of Experts (MoE): The architecture relies on a large number of smaller, specialized experts (e.g., 32 to over 100). This allows for much faster inference speeds, which is critical for user experience, as users are unwilling to wait for large monolithic models to generate tokens.
• Internal Tool Use: The models feature an “analysis channel” or “thinking mode.” Within this process, the model can use internal tools—like a Python interpreter for math calculations—to verify its work and improve accuracy without surfacing these steps to the user. This internal tool use is expected to expand dramatically, becoming a core component of model reasoning.
• Ecosystem Advancement: Progress is not just about raw model performance. Improvements in inference cost, speed, UI, and the massive expansion of available tools are equally critical in accelerating experimentation and getting closer to AGI.

The Developer Experience: Why Claude Still Has the Edge

Despite GPT-5’s power, its developer workflow has frustrating limitations that keep Claude in the lead. This was demonstrated through two live examples.

Chris’s “Blinking Donkey” Test

Chris tasked both models with creating a webpage featuring a blinking donkey, with eyes synced to an internet clock.

• GPT-5’s Capability: The final code was solid and the visual result was aesthetically impressive. The model’s ability to generate quality front-end code is significantly better than its predecessors.

• GPT-5’s Frustration: The process was fraught with friction.

  • The “Pro” model generated the code but refused to render it on the built-in canvas, forcing a manual file save.
  • When explicitly asked to use the canvas, it took eight minutes and produced an incorrect version.
  • The model frequently resorted to omitting code with placeholders like "rest of your code goes here," a classic pain point for developers. It required 25-30 minutes of re-prompting, correcting, and copy-pasting code back into the chat to achieve the final, working result.

• Claude’s Advantage: In contrast, Claude handles artifacts and conversational context seamlessly. As Chris noted, “Claude just gets it.” It remembers the context, updates code within its “canvas,” and doesn’t omit chunks, leading to a much smoother and more productive workflow.

Mihai’s Mermaid Diagram Test

Mihai found similar workflow issues. When asked to generate a complex Mermaid diagram, Claude and even the local gpt-oss model produced a working result on the first try. GPT-5, however, required two or three attempts, forcing him to copy the error from the renderer and feed it back to the model for correction.

Final Verdict

GPT-5 is a formidable and highly accessible model, especially shining in agentic capabilities like web browsing and tool use. However, the developer experience is still hampered by issues with context management and code omission. For complex coding tasks and maintaining a productive workflow, Claude Opus 4.1’s superior handling of artifacts and conversational context makes it the preferred tool for now. The intense competition continues to drive innovation, benefiting the entire community.

• Tags:
• Code Generation
• Agentic AI
• Feature
• OpenAI
• Anthropic
• GPT-5
• Claude Opus
• Mixture-of-Experts
• Model Comparison