GPT-5: OpenAI Unifies the Stack
GPT-5 shipped in August 2025 and did something OpenAI had been hinting at for months: it collapsed the GPT-4/o-series split into a single adaptive model. One model, scales to the task. It's the most significant OpenAI release since GPT-4.
The o-series (o1, o3) was OpenAI's answer to the reasoning question. But having two separate model families, one for "regular" tasks and one for "hard" tasks, created friction. Users had to know which model to pick. Developers had to route queries. It was a workaround, not a solution.
GPT-5 is the solution.
The Idea
One model. Adaptive compute. When you ask GPT-5 a simple question, it answers quickly. When you ask it something hard, it thinks longer. You don't choose. The model calibrates based on the difficulty of what you asked.
This is how reasoning should work. And at launch, GPT-5 executes on it well enough that switching between "thinking" and "non-thinking" modes feels like a relic of the last generation.
What's New
- Unified architecture. No more GPT-4o vs. o3 routing decision. One endpoint, adaptive reasoning depth.
- Best across general benchmarks at launch. Leads GPQA Diamond, MMLU-Pro, and most science/math evaluations at the time of release. Strong on coding.
- Broad multimodal. Text, images, audio, code, function calls: all native, all improved.
- Better long-context performance. 128K context window, used more reliably than previous generations.
- Stronger tool use. Function calling, code interpreter, web search: all more reliable than GPT-4o.
How It Compares at Launch
| Model | GPQA Diamond | Reasoning | Multimodal |
|---|---|---|---|
| GPT-5 | top at launch | ★★★★★ | ★★★★★ |
| Claude 4 Opus | top 3 | ★★★★★ | ★★★★☆ |
| Gemini 2.5 Pro | top 3 | ★★★★★ | ★★★★★ |
| Llama 4 Maverick | open-weight leader | ★★★★☆ | ★★★★☆ |
| Grok 3 Thinking | top 5 | ★★★★★ | ★★★☆☆ |
GPT-5 is the best general-purpose model at launch. Claude 4 Opus retains the coding agent crown. Gemini 2.5 Pro remains the best on long-context and video.
Best For
- Everything you used GPT-4o for, but better
- Tasks that benefit from adaptive reasoning without manual mode-switching
- Multimodal workflows involving mixed inputs
- Broad enterprise deployments where a single model needs to handle diverse task types
Not For
- Coding agents specifically (Claude Code + Opus 4 remains the best stack here)
- Ultra-long context (Gemini 2.5 Pro's 1M window is still ahead)
- Self-hosting (still closed, still API-only)
The Real Story
GPT-5 is an architecture signal as much as a model release. The direction is clear: one adaptive model that scales effort to the task, not a menu of specialized models you have to pick between. Every lab is moving this way. OpenAI shipped it first.
What comes next is whether they can hold the lead. Gemini 3 Pro is already in development, and the competitive pressure has never been higher.
Part of our Model Watch series. Next: Gemini 3 Pro →