GPT-4.5 is live in the API!

[nikunj]

Today we’re excited to announce gpt-4.5-preview, our largest model yet, as a research preview. It has deeper world knowledge, with a better understanding of user intent. GPT-4.5 is designed for more natural conversation. It excels at tasks requiring creativity, empathy, and broad general knowledge, including writing help, coaching, brainstorming, and nuanced communication. We’ve also seen it perform well at agentic planning and execution.

GPT-4.5 supports function calling, Structured Outputs, vision, streaming, system messages, evals, and prompt caching. It’s available via our Chat Completions, Assistants, and Batch APIs today.

GPT-4.5 is very large and compute-intensive, so it’s not a replacement for GPT-4o. A typical query costs on average $68 / 1M tokens, with cache discounts ($75 / 1M input tokens, $37.5 /1M cached input, $150 / 1M output). Batch jobs are discounted 50% and cached input is discounted 50%.

We’re evaluating whether to continue serving it in the API long-term as we balance supporting current capabilities with building future models. If GPT-4.5 plays a unique role for your use-case, let us know in this topic.

You can read more about GPT-4.5 in our blog post. We can’t wait to see the creative applications you’ll build!

[PaulBellow]

Introduction to GPT-4.5 OpenAI Livestream 2/27/25

There is an arms race ramping up this year, Gemini Flash eating up a lot of the low end, Claude 3.7 / Grok 3 / DeepSeek V3 as direct competitors to GPT-4o. IMO we will see A LOT of rushed moves this year cause everyone is feeling tremendous amount of competition.

We had a little discussion during the live event, but feel free to put feedback here in this Announcement thread, so OpenAI can evaluate how they want to serve it in the API in the long-term.

[mat.eo]

@nikunj

We’re evaluating whether to continue serving it in the API long-term

Is this not the last non-reasoning model?

Well… I imagine there’s no answer for this. But I’m taken back by this. I’m also interested in seeing what use-cases are found with this model

I think there’s a lot of value in “non-reasoning” models, but not like this.

[etdilley]

I’m using ChatGPT (apparently 4.5 now) as the LLM in my Home Assistant, and simply responding to the email that said “if you use it let us know” … I use it, I like it a lot, I pay for it, please keep it in the API.

[sire]

Obviously haven’t had more than a couple hours to test the new 4.5 model but I can already tell it’s a big improvement in terms of sales related automation (requiring a «human touch»). Please keep supporting it in the API

[mat.eo]

Do you mind elaborating on the differences to gpt4o you found?
Especially in the human touch - which can be considered invaluable

[PaulBellow]

Welcome to all the newcomers! We’ve got a great community here with a wealth of knowledge if you do a bit of digging. Thanks for stopping by to talk about GPT-4.5, but we hope you stick around!

[2034-SWE]

GPT-4.5 is by far the MOST exceptional model I have ever used in my entire life (thousands of hours using AI) for a very specific, high-demand, nuanced, profitable use-case.

I think your API is UNDERPRICED.

Actually INSANE how good this model is. I’m mind blown. And I’m only a few minutes in.

Respectfully, I ask that you keep serving this model via API long-term.

Truly, I anticipate >$100k monthly usage in API credits with this (type of) model alone, long-term. Maybe even >$1M monthly by 2026.

And I think my use-case may end up being a very popular case study as to why/how GPT-4.5 is immediately profitable in a very meaningful, scalable, high-demand domain/industry. Stay tuned

Thank you

[lbrendle]

how can developers with a niche use case contact you?

[AlexisTL]

getting paid to say that ? underpriced lmao its 10x more expensive than any other api for a result you can almost achieve by instructing your ai properly.

[Diet]

Welp

[2034-SWE]

No no, not even close. You can’t replicate what this model is capable of (for my use-case and architecture) with any other model. Nothing comes even close.

Not getting paid — I’m just that blown away.

You might end up using one of my next 3 major disruptive global projects over the next 1-3 years believe it or not

[PaulBellow]

I’ve not tested it a lot myself, but I’m sure it’s “.5” better!

[2034-SWE]

I should add, that if $200 in API consumption with GPT-4.5 produces >$50k in profit… you’d understand why I think it’s underpriced. That’s all I’ll say for now

God bless

[Diet]

to be fair, it can redeem itself:

unlike 4o, the poor boy: (first assistant response hand written)

[AlexisTL]

200 is the price of pro its not the api pricing. your 200 will be far gone before making your 50k.

GPT-4.5

Largest GPT model designed for creative tasks and agentic planning, currently available in a research preview. | 128k context length

Price

Input:
$75.00 / 1M tokens

Cached input:
$37.50 / 1M tokens

Output:
$150.00 / 1M tokens

[2034-SWE]

This reminds me of Wes Roth (in YT) talking about how AI’s capabilities (in comparison to a human) is graphed as a jagged edge;

baseline horizontally = human intelligence,

and vertical points = above and below the human baseline, for various different domains/tasks.

For example, coding something may be top 1%, and counting r’s can be bottom 1%. Focusing on how to utilize the strengths is the best daily subconscious objective to sustain focused, curiosity-driven discovery and innovation built on AI

[2034-SWE]

I was referring to API pricing. Not the Pro subscription.

$1k in API consumption for >$250k profit.

[Bravepanda]

I use ChatGPT a lot. I’m definitely interested in 4.5 and im so sad to hear you might not continue. Please keep it. It has been so essential for me that I can’t think of my life without it