Is gpt 5.4 so much costlier compared to gpt 5.3?

Discussions
,
1

I see a roughly 10x cost difference between the gpt5.4 model and gpt 5.3 model. Are others seeing this as well? I am running both in MAX mode.

image

3

max mode depends on use context size (>200k price twices). On non-max they are pretty similar in prices

4

Hey, what you’re seeing is expected, but the main reason is Max Mode, not just the base rates.

GPT-5.4 has a long-context surcharge. When the input goes over 272k tokens, the input price doubles and the output price becomes 1.5x. In Max Mode with -high reasoning, it’s easy to hit that limit. GPT-5.3 Codex starts with lower base rates, and its surcharge works differently, so the price gap grows fast.

In non-Max mode, the prices are pretty similar. If you don’t need the extra-long context for every task, switching GPT-5.4 to regular mode will reduce that gap a lot.

A couple of options:

  • Use GPT-5.4 without Max Mode when you don’t need the full context
  • Try -med instead of -high on GPT-5.4. You’ll get fewer thinking tokens with similar quality for most tasks
  • GPT-5.3 Codex with -high is strong on its own. It’s close to Opus 4.6 on our benchmarks for a lot less cost

Pricing details: Models & Pricing | Cursor Docs

2 Likes
5

Thanks for the usage tip on GPT-5.4 and I agree, 5.3-Codex-High is excellent.

By the way I have rolled forward again to Cursor Version 2.6.21 and am enjoying trouble free agent terminals. The latest UI is excellent, and my session productivity has been thru the roof recently.