How much compute will be used to train GPT-5?

1.5kṀ1404

2027

1.2%

Fewer than 5e25 FLOP

12%

Between 5e25 FLOP and 1.5e26 FLOP

30%

Between 1.5e26 FLOP and 4.5e26 FLOP

32%

Between 4.5e26 FLOP and 1.4e27 FLOP

17%

Between 1.4e27 FLOP and 4e27 FLOP

More than 4e27 FLOP

This resolves as the bucket that contains the number of floating point operations (FLOP) used to train GPT-5

FLOP may be performed at any precision (INT8, FP16, etc.)
This resolves on the basis of the numbers reported by OpenAI, and, if they don't report this, on the basis of the first credible estimates reported in this database (details about the database may be found here)
If GPT-5 is not released by EOY, 2027, this resolves ambiguously
If the estimate is exactly on the edge of each range, we will resolve it as the largest of the two options

Technology

GPT-5 Speculation

Get

1,000

to start trading!

People are also trading

What will be true about GPT-5?

In yottaFLOPs (10^24), how much compute will GPT-4 be trained with?

GPT-5 trained with >=24k GPUs?

82% chance

What hardware will GPT-5 be trained on?

What percentage of 2025 will be left when OpenAI releases GPT-5?

32% chance

What percentage of 2025 will be left when OpenAI announces GPT-5?

32% chance

When will GPT-5 finish training?

Will a model be trained using at least as much compute as GPT-3 using AMD GPUs before Jan 1 2026?

84% chance

Will growth in the maximum MW used to train AIs, slow down by more than x2 after GPT-5-like?

55% chance

Will the ratio of inference runs to training runs on GPT5 decrease from the ratio on GPT4?

Sort by:

Epoch estimates that Grok-3 (released in February) was trained with 4.6e26 FLOPs. The trend for compute spent on frontier models has a doubling time of around five months. Since GPT-5 will likely be released 4-7 months after Grok-3, this trend predicts around 1e27 FLOPs.

a fuckton

@TamayBesiroglu What happens if it's between 1e26 and 2.5e26? It seems there are two options for this

Between 5e25 FLOP and 2.5e26 FLOP
Between 1e26 FLOP and 5e26 FLOP

Betting here based on the uneven log scale.

@BoltonBailey Yea, it's tricky not to have a non-even log-scale and round numbers. I've now edited the buckets slightly to roughly incremenent by factors of three.

@TamayBesiroglu I can see that. These buckets are much more even. I have sold my positions because the numbers on the options now no longer reflect what they were when I bought them. Luckily, no one seems to have traded in the interim.