OpenAI ships GPT-5.4—three variants and a 1.05M token context

What is new

GPT-5.4 is OpenAI’s most capable and most efficient model to date. It comes in three variants:

• GPT-5.4 Standard — balanced cost / performance for everyday workloads
• GPT-5.4 Thinking — reasoning-first variant for complex tasks
• GPT-5.4 Pro — maximum capability for the most demanding work

1.05 million tokens

A context window of 1,050,000 tokens lets you fit into a single call:

• An entire mid-size codebase
• Thousands of pages of documentation
• Years of email or meeting notes

For dev workflows this means the AI can see the full repo context without chunking or RAG.

Benchmarks

GPT-5.4 hit a record 83% on OpenAI’s internal GDPval test and set new bests on computer-use benchmarks (OSWorld-Verified, WebArena Verified).

My take

For clients who were on the fence about deploying an AI agent on complex tasks, this is the inflection point:

1. Reasoning — the Thinking variant unlocks task-decomposition workflows that previously failed at planning
2. Long context — eliminates a lot of RAG feature engineering for 80% of use cases
3. Three pricing tiers — start on Standard and switch production-critical paths to Pro

I expect the competition (Claude, Gemini) to ship upgrades within a month—prices and context up, entry barriers down.