OpenAI

GPT 4.5 Preview

Type
language model
Context
128K tokens
Max output
16K
API string
azure/gpt-4.5-preview

Prices by provider

Provider Input, $ per 1M Output, $ per 1M In our data since
Microsoft Foundry $75 $150 31 Jul 2026

Only the base tier and only per-token prices are shown. Batch, discounted and cached rates, as well as prices per image or per second of video, do not go into this table: they cannot stand in the same column as a price per million tokens.

The date in the last column is the day this price first entered our collection. The price may well be older: before that day we simply were not recording it. It has not changed since — otherwise a new row with a new date would stand in its place.

Measurement results

Task set Result Run conditions Measured by
Arena Score, multi-turn dialogue 1,443.92 1,430–1,457
Arena Score in English 1,419.49 1,412–1,427
Arena Score in Russian 1,418.24 1,403–1,434
Arena Score in French 1,418.01 1,362–1,474
Arena Score, overall 1,417.47 1,412–1,423
Arena Score, mathematics 1,411.89 1,397–1,427
Arena Score, long queries 1,406 1,392–1,420
Arena Score, instruction following 1,404.35 1,396–1,413
Arena Score, hard prompts 1,403.65 1,394–1,414
Arena Score, programming 1,397.05 1,384–1,410
Arena Score, creative writing 1,394.29 1,382–1,406
Arena Score, expert questions 1,393.91 1,371–1,417
Arena Score, working with images 1,195.13 1,183–1,207
MATH, difficulty level five 78.63 % · with a tuned harness Epoch evaluations
Creative writing (Lech Mazur’s evaluation) 75.6 % lechmazur/writing Github repository
Fiction.LiveBench — holding a long context 63.9 % Fiction.live leaderboard
GPQA Diamond — graduate-level questions 58.25 % · with a tuned harness Epoch evaluations
Aider Polyglot — code edits in six languages 44.9 % · with a tuned harness Aider LLM Leaderboards
WeirdML — unusual machine learning tasks 39.37 % WeirdML Leaderboard
Mock AIME 2024–2025 — olympiad problems 37.72 % · with a tuned harness Epoch evaluations
SimpleBench — trick questions 21.4 % SimpleBench Leaderboard
Cybench — cybersecurity tasks 17.5 % · with a tuned harness Cybench leaderboard
ARC-AGI — generalising to unseen patterns 10.3 % · with a tuned harness ARC Prize Leaderboard
ARC-AGI-2 0.8 %
Humanity’s Last Exam — expert-level questions 0.67 %