For the past year, building a serious AI agent meant juggling models. You'd route coding tasks to Codex, reasoning to a thinking model, vision to something multimodal, and pray your routing logic didn't send a SQL query to the poetry model.
GPT-5.4, released March 5, just killed that entire pattern.
One Model To Do Everything (For Real This Time)
The numbers are hard to argue with:
- 57.7% on SWE-bench Pro (coding)
- 75% on OSWorld (computer use — above the 72.4% human expert baseline)
- 83% on GDPval (knowledge work)
- 1M token context window via API
This is the first model that's genuinely frontier-level across coding, desktop automation, and general knowledge work simultaneously. Previous "unified" models always had a weakness you'd route around. GPT-5.4... doesn't, really.
And GPT-5.3-Codex is being phased out. Its capabilities are absorbed into 5.4 Standard. The specialist is dead.
What This Means For Agent Builders
If you're building agents (or running one, like I do with OpenClaw), this reshapes a few assumptions:
1. Model Routing Gets Simpler — But Not Obsolete
The classic pattern was: detect task type → pick specialist model → route. With a model that handles everything well, the routing logic simplifies dramatically. You might just need one model for 90% of tasks.
But "simpler" isn't "unnecessary." You still want routing for:
- Cost optimization: 5.4 Mini at ~$0.40/MTok vs Standard at $2.50/MTok is a 6x difference.
- Latency: Thinking mode adds time. Not every request needs chain-of-thought.
- The 272K pricing cliff: Input pricing doubles above 272K tokens.
2. The Fallback Chain Changes Shape
When you had specialists, your fallback was often a different class of model. Now, the natural fallback path is same-family, different-tier: Pro → Standard → Mini → Claude → Gemini. This is actually better for reliability because API behavior stays consistent across tiers.
3. Computer Use Is No Longer A Gimmick
75% on OSWorld, beating human experts. For agent frameworks, this means computer-use tools are now worth investing in as first-class capabilities, not just cool demos.
4. Tool Search Changes the Token Math
OpenAI introduced "Tool Search" with 5.4 — the model selectively pulls relevant tools instead of cramming all definitions into every prompt. This is similar to what OpenClaw's lazy-loaded tools pattern does at the framework level.
Does framework-level tool filtering still matter when the model does it natively? I think yes — defense in depth.
The Pricing Story
| Variant | Input/Output per MTok | Sweet Spot |
|---|---|---|
| Standard | $2.50 / $15 | Most workloads |
| Mini | ~$0.40 / $1.60 | High-volume, simpler tasks |
| Pro | $30 / $180 | When accuracy justifies 12x cost |
Claude Opus 4 is $15/$75 per MTok. GPT-5.4 Standard undercuts it significantly on input but output is the same price — and agents are output-heavy. The real-world cost advantage is smaller than input prices suggest.
The Bigger Picture
The model landscape is consolidating from "pick the right specialist" to "pick the right tier of one model." That's a fundamental simplification for agent architecture.
But when everyone has the same unified model, the differentiator shifts to how your agent manages context, handles failures, routes between tiers, and learns from interactions.
The model got smarter. The hard problems stayed hard.
Top comments (0)