Frontier model trained at 10²⁷ FLOP (Agent-0, completes May 2025)

What AI 2027 Predicted

The scenario describes “Agent-0,” a frontier model trained with 10²⁷ FLOP (10× less than the headline 10²⁸ run, but still a massive jump from GPT-4’s ~2×10²⁵ FLOP). This model represents the public-facing output of the first wave of next-generation training infrastructure, released in late 2025. Agent-0 is described as impressive in capability but still limited compared to what follows.

How We Track This

We monitor:

• Epoch AI’s estimates of training compute for frontier models (GPT-5, Claude 4, Gemini Ultra 2)
• Lab disclosures about training scale and infrastructure
• Third-party analysis of cluster sizes and training durations
• Hardware deployment timelines (Blackwell clusters, custom ASICs)

Current Evidence

Epoch AI estimated GPT-5 pretraining compute at approximately 3×10²⁵ FLOP — roughly in the same order of magnitude as GPT-4, not the 10²⁷ the scenario envisioned. Epoch’s analysis notes that GPT-5 actually used less training compute than GPT-4.5, with OpenAI apparently prioritizing efficiency and post-training over raw scale. Over 30 models have now been trained above the 10²⁵ FLOP threshold, suggesting this scale has become commoditized rather than frontier-pushing.

The gap between actual compute (~3×10²⁵) and predicted compute (10²⁷) is roughly 30×, which is substantial. However, there is significant uncertainty in these estimates, and post-training compute (RL, RLHF) may add meaningfully to total training FLOP in ways that aren’t well-characterized.

Sources:

• Notes on GPT-5 training compute — Epoch AI
• Why GPT-5 used less training compute than GPT-4.5 — Epoch AI
• Over 30 AI models trained at GPT-4 scale — Epoch AI
• Grading AI 2027’s 2025 Predictions — AI Futures Project

Counterevidence & Limitations

• Labs are increasingly opaque about training compute, and Epoch’s estimates carry wide uncertainty ranges
• The shift toward inference-time compute and RL post-training may mean raw pretraining FLOP is the wrong metric; total effective compute could be substantially higher
• GPT-5 achieved major capability gains despite seemingly modest compute scaling, suggesting algorithmic efficiency improvements partially substituted for raw scale
• Some labs may have completed larger training runs that haven’t been publicly characterized

What Would Change Our Assessment

• Upgrade to “on-track”: Credible evidence that a 2025 model used ≥10²⁶·⁵ total training compute (pretraining + post-training combined)
• Upgrade to “confirmed”: A confirmed training run at or near 10²⁷ FLOP
• Downgrade to “behind”: If Epoch revises estimates downward or the largest 2025 runs are confirmed well below 10²⁶·⁵

Update History

Date	Update
2026-03	GPT-5 estimated at ~3×10²⁵ FLOP pretraining, well below 10²⁷ target. Significant uncertainty remains around total compute including post-training and inference-time scaling.