All I want to know is how old is the model and how much does it cost. I can figure out which one I want to use based on that, assuming that newer models are always better.
Trying to convince us there is a difference between GPT-6-Sol and GPT-5.6-Terra or whatnot is ludicrous to the point of being insulting, especially when new models come out every week.
What's unclear to me is who OpenAI thinks they're marketing to with this form of branding. These different models don't really mean all that much to the vast majority of people using their products who aren't developers, and developers aren't helped at all by the way they've been naming said models. Are they merely scared that they'll become irrelevant because Anthropic decided to give their models quirky names like "Opus" and "Fable"?
If OpenAI really wants to give their models names, they should name the generation of model and then have the different sub-models named by purpose or capability level. After all, I wouldn't use Mini for a job that Nano could easily do, and I wouldn't use Nano for a job that the full version of GPT-* necessitates. Similarly, I've had to discover exactly how Luna, Terra, and Sol are appropriate for different complexities and task types. OpenAI could help me skip a lot of those steps and just tell me what each model distillation is good for without causing me to look through their pricing page and make educated guesses. After all, shouldn't they not want me to pay attention to how much they're charging me?
All of this makes the days of frontend framework churn seem quaint and actually preferable.
The issue that OpenAI had when they had mini and nano models is that ambiguous the differences between those and everyone just used the base model anyway. I have no idea what type of job mini can do that nano couldn't or vis-à-vis.
I do wonder if it would just be better if they were named 6-small, 6, and 6-big?
In my experience, Nano won't reliably handle complex open-ended tasks and is mostly suited for very explicit instruction that it can't screw up. It's no different from how there are some chores you can give to kids and there are other tasks you need at least a teenager for. If the decision tree of the task is very clear and conventional, Nano can be cheaper than giving the task to a relatively overpowered model, especially if it's something where the output is rigidly structured. This makes it well suited for skills that essentially run CLI commands and generate output, especially because it is usually faster. Mini is more like a discount version of the base model, and Nano is the dollar store version. Mini is more of a generalist and a fairly good deal if you have a moderately complex task that is conventional, but can be less conventional that what Nano can handle. I mostly used gpt-5.4-mini this year for my side projects because it's a pretty good generalist while significantly saving on costs. It is, however, somewhat dumber than the base model and more prone to ignore or forget rules you give it. I'd have just used a base model, but the low cost of Mini and Nano made them appealing to me. Maybe I'm a cheapskate, but I have hundreds or possibly thousands more in my pocket than many other users because of that.
This workflow I settled into with Mini and Nano didn't map cleanly on to the current generation of model tiers. With the price of Luna, you'd think it would be a replacement for Nano. In a sense it is, yet I didn't find that Terra became the new Mini. Terra is more powerful, better at explaining its own decisions, yet I've also found it to be relatively stupid while charging me more to use it. On the other hand, Luna with its reasoning set to "high" is what I consider to fill the role of Mini, and is good enough such that I no longer use Mini. Sol and Astra are great, but they're pricey. It could be my own brain and its bad perception, but so far I don't get the point of Terra. Luna succeeded at reverse engineering some abandonware with a very complicated licensing and virtualization scheme, and did so over SSH into a Windows VM with only PowerShell on the other end. Terra did such idiotic crap to my flashcards app that I stopped using it for anything after that.
This is why I find OpenAI's naming unhelpful and kind of pointless. I don't really care about the benchmarks that all these models are commonly run against. They're not that useful, IMO. OpenAI could easily give early access to these models, get a ton of feedback, and provide better insight to customers on how these things behave. Even calling Terra "gpt-5.6-overpriced-cheating-dumbass" would be better than wasting my time and money figuring it out myself. But that wouldn't make OpenAI as much money.
> In the Coding Agent Index, Sol improves but Luna regresses: In OpenAI's Codex harness, GPT-6 Sol (max) scores 57 in the Artificial Analysis Coding Agent Index, up 2 points from GPT-5.6 Sol (max), with gains in Terminal-Bench 4.0 (43% vs 37%) and SWE-Atlas-QnA (58% vs 54%). At $2.99 per task it costs ~50% less than GPT-5.6 Sol (max) and sits on the Pareto frontier of Coding Agent Index vs Cost per Task. GPT-6 Luna (max) scores 41, down 2 points from GPT-5.6 Luna (max), with lower scores in SWE-Atlas-QnA (44% vs 49%) and DeepSWE v1.1 (64% vs 66%), at ~60% lower cost per task.
https://x.com/ArtificialAnlys/status/2102462962758033624
Given that they had to discontinue sales of the 20x Pro plan after the Astra release due to compute constraints, I wonder if 6 Sol & Luna are smaller vs their 5.6 counterparts?
At least it sounds good on paper, the the graphed results do give me pause as it seems the lower cost might come from a slightly nerfed base model combined with more thinking, going by the more erratic scoring curves and the lower no thinking baseline score. I’ll have to try it out but I really hope they haven’t nerfed Luna/Sol to make this price point possible!
Rarely have I seen such hogwash. It seems to be a mix of virtue signalling and trying to push the perspective that being "90th percentile" (on what exactly?) requires extensive AI use. You are telling me you expect each researcher to generate USD 7000/d or USD 140k/m in AI cost? Or is that a way to abuse tax laws in some way so they can claim their own payments for tokens as expenditure on the other side of the ledger?
From context, I took this to mean 90th percentile in token usage. So yes, being a top token user does require extensive AI use.
Overall, I expect for most people think the winner of today was Anthropic. I personally am preferring Opus 5.5 at medium over GPT-6 Sol Max, in very very early tests. Similar price range, more capability.
But competiton is great, these are solid releases by OpenAI today.
Model update Input Output Reduction
------------------------- ------------- ------------- ---------
GPT-5.6 Sol → GPT-6 Sol $4 → $2 $20 → $10 50%
GPT-5.6 Luna → GPT-6 Luna $0.20 → $0.10 $1.20 → $0.50 50%