Contents

GPT-6 Pretraining Complete: 18 Months, 100k H100s, 2M Token Context

Breaking Today

Today (April 14, 2026), OpenAI officially confirmed GPT-6 pretraining completed on March 17. Pretraining lasted 18 months—far longer than GPT-5’s development cycle.

Greg Brockman (OpenAI President) subsequently confirmed the model’s existence in a podcast interview, with a single characterization: “not incremental improvement, but a major change in how we think about model development.”

Core Specifications

Metric Value
Parameters 5-6 trillion (MoE, ~10% activated)
Activated parameters ~230 billion
Context window 2 million tokens
Training duration 18 months
GPU cluster ~100,000 H100s
Training cost Over $20 billion
vs GPT-5.4 improvement 40%+ (code/reasoning/agent)

Symphony Architecture

GPT-6’s core highlight isn’t parameter count—it’s a fundamental architecture restructure, codenamed Symphony.

Design core:

  • Native multimodal: text, images, audio, video in the same embedding space from initial design—no more “text model with a vision module bolted on”
  • Dual-system reasoning: System-1 handles fast responses (fast thinking), System-2 handles logical verification and multi-step derivation (slow thinking)—echoing classic cognitive science theory

The first solves the “cobbled-together” feeling of current multimodal models. The second solves “reasoning chains breaking mid-derivation.”

2M Token Context Significance

GPT-5.4’s 1M token already solved long document processing. GPT-6 doubling that:

  • can process two novellas in a single context
  • fits complete transcript of 10 hours of audio
  • in codebases: equivalent to fitting an entire medium-sized repo into context

Pricing

Input Output
GPT-6 $2.5/M token $12/M token

Flat with GPT-5.4—no price increase despite parameter growth.

Timeline

Official schedule:

  • June 14: Public API availability (exactly 3 months after pretraining completion)
  • Enterprise early access open for application now

Brockman’s quote: “GPT-6 is the last mile to AGI.” Marketing or genuine assessment? We’ll test it the moment API opens.