OpenAI Launches New Coding-Focused GPT-4.1 Models

[MacRumors]

OpenAI today announced the launch of three new GPT models that are available through the OpenAI API. Called GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, the models are not available for ChatGPT at the current time.

GPT-4.1 includes major improvements in coding, instruction following, and long context, according to OpenAI. The models outperform GPT-4o and GPT-4o mini in all tasks, especially coding tasks. The new models support up to one million tokens of context, and can use that for better long-context comprehension.

On SWE-bench Verified, GPT-4.1 showed a 21.4 percent improvement over GPT-4o for coding, and a 26.6 percent improvement over GPT-4.5. The slimmed down GPT-4.1 mini matches or exceeds GPT-4o with reduced latency and an 83 percent lower cost. GPT-4.1 nano is OpenAI's fastest and cheapest model, ideal for classification, autocompletion, and similar tasks.

OpenAI says that many of the improvements in GPT-4.1 have already been incorporated into the ChatGPT version of GPT-4o, with more functionality to be added in the future. All GPT-4.1 models have a cutoff date of June 2024, which means they are up to date on events that took place before then.

With the introduction of GPT-4.1, OpenAI is deprecating GPT-4.5 in the API, as GPT-4.1 offers much of the same functionality at a lower cost.

GPT-4.1 costs $2 per million input tokens and $8 per million output tokens, while GPT-4.1 mini costs $0.40 per million input tokens and $1.60 per million output tokens. GPT-4.1 nano is just $0.10 per million input tokens and $0.40 per million output tokens. Customized models are available at a higher price.


Article Link: OpenAI Launches New Coding-Focused GPT-4.1 Models

 

[Populus]

Wait wait wait… OpenAI is deprecating (at least on the API models) the recently announced GPT-4.5 because they just released GPT-4.1 which is superior? Hmmm why not call it GPT-4.6 if it’s superior?

 

[neotint]

4.1? ClosedAI continues with the weirdness... It'd be better if they named it 4o-Newer, and then after that 4o-Newer-Newest.

 

[Ad47uk]

Seeing how reliable these AI things are at answering even a simple question, I will leave well alone. What needs to be done is to decrease the amount of energy they use.

I don't use this AI stuff, even turned it off on the Mac, so no chance of me using any other.

 

[karnac]

Maybe they could use AI to come up with a logical naming scheme.

 

[Adam089]

And then there is Siri

 

[sw1tcher]

GPT-4.1 includes major improvements in coding, instruction following

This will be helpful in the school system with getting students to follow instructions. We need to making sure that first graders, or even pre-Ks, have A.1. teaching in every year. This will be a wonderful thing.

 

[mr_user1]

Wait wait wait… OpenAI is deprecating (at least on the API models) the recently announced GPT-4.5 because they just released GPT-4.1 which is superior? Hmmm why not call it GPT-4.6 if it’s superior?

No. These models are specifically targeted at doing coding and scientific tasks. Not general purpose models like GPT-4.5 and are not meant to replace GPT-4.5 for general purpose tasks.

Their naming conventions suck. Perhaps they could ask ChatGPT to help them name their models better.

 

[HiRez]

4.1? ClosedAI continues with the weirdness... It'd be better if they named it 4o-Newer, and then after that 4o-Newer-Newest.

Yeah, their naming scheme is just awful. Even for technical people it can be difficult to understand/remember what is what. For non-technical people like those selecting a model in the ChatGPT interface, it's absolutely incomprehensible. And even worse, counter-intuitive at times. I understand it's a rapidly-changing technology with a need for multiple models, but it's nuts that they think this is fine.

 

[LV426]

Skynet type situation, anyone?

 

[HiRez]

"ChatGPT, come up with a clear and understandable naming scheme for all the ChatGPT models."

Sure! Here’s a simple and clear naming scheme for ChatGPT models that makes it easy to understand their generation, capabilities, and special features at a glance:

Core Format: GPT-<Version>-<Variant>

• <Version> = Model generation (e.g., 3.5, 4, 4.5, 5.0)
• <Variant> = Optimization type (e.g., Base, Turbo, Omni, Mini, etc.)

Standard Variants Explained

• Base: Raw, unoptimized model with standard performance
• Turbo: Optimized for speed and efficiency (used in Plus plans)
• Omni: Multimodal (can handle text, image, and possibly audio)
• Mini: Lightweight, cheaper and faster variant for basic tasks
• Pro: Premium variant with extended memory/tools/accuracy

Example Names (Using This System)

• GPT-3.5-Base: Original 3.5 model, standard speed & capabilities
• GPT-4-Turbo: GPT-4 optimized for fast and efficient inference
• GPT-4.5-Omni: GPT-4.5 with full multimodal support
• GPT-4-Mini: Lightweight GPT-4 model for cheaper, quick use cases
• GPT-5.0-Pro: Next-gen premium model with top-tier performance

 

[picpicmac]

Just in time for the new COBOL programs...

 

[macmyworld]

No. These models are specifically targeted at doing coding and scientific tasks. Not general purpose models like GPT-4.5 and are not meant to replace GPT-4.5 for general purpose tasks.

Their naming conventions suck. Perhaps they could ask ChatGPT to help them name their models better.

Can they screw up code like they normal ChatGPT screws up answers to questions?

 

[skiguy45]

Core Format: GPT-<Version>-<Variant>

• <Version> = Model generation (e.g., 3.5, 4, 4.5, 5.0)
• <Variant> = Optimization type (e.g., Base, Turbo, Omni, Mini, etc.)

Standard Variants Explained

• Base: Raw, unoptimized model with standard performance
• Turbo: Optimized for speed and efficiency (used in Plus plans)
• Omni: Multimodal (can handle text, image, and possibly audio)
• Mini: Lightweight, cheaper and faster variant for basic tasks
• Pro: Premium variant with extended memory/tools/accuracy

Example Names (Using This System)

• GPT-3.5-Base: Original 3.5 model, standard speed & capabilities
• GPT-4-Turbo: GPT-4 optimized for fast and efficient inference
• GPT-4.5-Omni: GPT-4.5 with full multimodal support
• GPT-4-Mini: Lightweight GPT-4 model for cheaper, quick use cases
• GPT-5.0-Pro: Next-gen premium model with top-tier performance

clear as mud.

where does o1, o3-mini and o3-mini-high fit into this mess?

So for coding they now say 4.1 is the best, and 4o incorporates some features of 4.1, but using 4o for coding still actively suggests you try o3-mini-high.

This is what happens when engineers are allowed to name products.

 

[JosephAW]

Wait, another reason they’ll push off Siri to iOS 20 because all the coding they’ve done has gone down the drain.

 

[HiRez]

Can they screw up code like they normal ChatGPT screws up answers to questions?

Of course they can. We're a long way from AI being able to generate useful, stable general code. I have found AI sometimes useful for generating small algorithms that do one, specific, narrowly-defined thing. But entire apps? Not really. It will often get it like 70% right, but trying to figure out what it did and fix the other 30% takes more time and is messier than just doing it yourself. But over time that going to be 80%, 90%, and eventually maybe 100%. But I think we're still years away from that.

The sticky bit is increasingly, AI code generators are going to be trained on AI-generated code. Will it be able to adapt/innovate at that point? Will it matter anymore?

 

[HiRez]

clear as mud.

where does o1, o3-mini and o3-mini-high fit into this mess?

So for coding they now say 4.1 is the best, and 4o incorporates some features of 4.1, but using 4o for coding still actively suggests you try o3-mini-high.

This is what happens when engineers are allowed to name products.

Not saying it's perfect, it's not. The point is with one simple prompt to their own system, it came up with an initial response that's already better than what they're using. It's funny they can make a crazy-advanced AI system and at the same time not come up with a simple naming scheme for their products. But you're right, "letting engineers name things" is probably the root of the problem. Most companies have goofy, engineer-named projects, but they keep those internal and then let actual writers and marketing people rename the public-facing products.

 

[asiga]

Fully agree with previous posters: the naming scheme is nonsense. Even the release strategy is bad: why improve 4o without increasing the version number? If MacOS was developed by OpenAI, the successor to Sequoia would be developed on the Ventura tree and marketed as Ventura.

They are also missing proper naming for models for specific tasks: if you have models for coding, just call the successive versions gptcode-1, gptcode-2, gptcode-3, etc.

 

[skiguy45]

Why is this even a choice in the end-user facing product? The api, I get it, but in ChatGPT why can't the AI figure out the best model to use for any particular question?

 

[BornAgainMac]

Reminds me of USB naming standards.

 

[jsavvy]

Just in time for the new COBOL programs…

You joke but I still make the occasional update to COBOL programs. I’d love for AI to do the updates for me.

 

[KillerScoregasm]

I personally have zero coding knowledge and zero intent to learn coding and I’ve built 2 websites between Grok and OpenAI. So I for one, welcome any additional coding improvements that allows me to speak my needs and it builds what I want…

 

[edvj]

Maybe they could use AI to come up with a logical naming scheme.

 

[Macaholic868]

Today Tim Cook broke wind. Take that Sam Altman and OpenAI! Apple is leaving you in the dust. Siri improvement will be almost be here by iOS 20. So put that in your pipe and smoke it!!!!!!

 

[scottrichardson]

So basically just keep using 4o inside the ChatGPT App?