GPT-5.6 Sol, Terra, and Luna explained: Which model should you use?
 |
| Before building anything, let's first figure out which model to use. Credit: OpenAI/YouTube |
These days, new versions of AI chatbots don't just launch; they're unshackled and released to the public following government scrutiny. OpenAI's new GPT-5.6 models were – like Anthropic's Claude Mythos and Fable – apparently too powerful to just launch; but now, after some tinkering, they're available to you, dear customer.
In practice, it simply means that the new GPT-5.6 models are very powerful and smarter than before. In its introductory post, OpenAI shared a bunch of graphs showing just how much better GPT-5.6 is than the competition, whilst using fewer tokens and generally costing less.
OK, great. But GPT-5.6 is not just one model; it comes in three distinct flavors: Sol, Terra, and Luna. So what do different kinds of users get, what should they pay for, and which models should they (mostly) use? Let's dive in.
Free users get (almost) nothing
Sorry; if you're not a paying customer, you'll have to make do with OpenAI's previous flagship model, GPT-5.5. Any sort of access to GPT-5.6 models requires a subscription of some sort. Fortunately, GPT-5.5 is still quite capable at most tasks, but if you want the best of the best, you'll have to cough up the dough.
There's an exception to this: Free and Go users can access GPT-5.6 through ChatGPT Work. More on that below.
If you're a Plus or Business user, you can only get Sol (the most powerful model) at medium and higher effort settings. There's another, higher level of performance called Sol Pro, but that's only available for Pro and Enterprise users.
In terms of availability per one million tokens, the prices are: $5 input and $30 output for Sol. $2.5 input and $15 output for Terra, and $1 input and $6 output for Luna.
SEE ALSO:Visa is connecting with ChatGPT to let AI agents automatically make purchases
Sol, Terra, or Luna?
Why are there three models in the first place? Well, OpenAI always had a multi-tier model; for example, previously users were able to choose a "mini" version of the main model to get results done cheaper. Now, the model has been split into three tiers.
If you're a paying customer, you're free to use all three. But you know how it is in the world of LLMs: If you pick the smartest one, your usage limits will get hit faster (yes, there are always usage limits, even if you throw a ton of money at OpenAI).
In the simplest sense, GPT-5.6 Sol is the smartest model, Terra is in between (with roughly GPT-5.5 level of performance), and Luna is the cheapest, fastest, but also least capable of the bunch.
The breakdown is as follows: Terra is a "balanced" model for everyday work. That's the one you should be asking most of your questions. Don't underestimate it, though, as OpenAI claims it performs better than Anthropic's Fable 5 in some cases.
Luna is cost-efficient, and should be used for easy, non crucial everyday tasks; think recipes and movie recommendations. Again, OpenAI says it outperforms Anthropic's Opus 4.8 in some cases, so it's not a slouch, either.
Sol should be reserved for coding, deep research, planning, and cybersecurity: The most demanding tasks. Of course it comes at a (literal) cost: While OpenAI claims it spends less tokens than Anthropic's Fable 5, Sol will still hit usage limits a lot faster than the other variants.
Fun fact: If you ask GPT-5.5 about any of this, right now, it'll give you completely wrong answers. Hopefully OpenAI will fix this soon.
Wait, what's this ChatGPT Work thing, then?
Oh yeah, OpenAI also launched ChatGPT Work, which is a new agent in ChatGPT that can access and take actions on your apps and files, and work in the background until a task is finished. It's powered by Codex (OpenAI's software engineering agent) and GPT-5.6. Think about it as your buddy that will go through your emails and files, browse the web, fetch the relevant data, and create that presentation your boss wants before the day is done.
ChatGPT Work is rolling out to Pro, Enterprise and Edu users first on web and mobile; this will be expanded to Plus and Business users "over the next few days."
On the desktop, Work is available for everyone, including Free users.
Oh, and one more thing: The fact that ChatGPT Work has a built-in browser also means that OpenAI is sunsetting its standalone web browser, Atlas. Sorry.
How about GPT Live?
GPT Live is a new version of ChatGPT Voice and it will show up when you start talking to ChatGPT.
We've covered this in more depth here, but the bottom line is that GPT Live can listen and speak at the same time, allowing it to keep up a more realistic conversation.
Wrapping it all up
The new GPT-5.6 model is smart. It comes in three flavors: Luna, Terra, and Sol, with Sol being the most capable variant, Luna the most affordable one, and Terra somewhere in the middle. You can currently only get them on paid tiers, unless you're using ChatGPT Work on desktop. And ChatGPT Voice has also gotten smarter with GPT-Live underneath, a model that can listen and speak at the same time.
Topics Artificial Intelligence OpenAI
Luna Medium for explorer
ReplyDeleteLuna high for mechanical tasks
Terra High for everyday tasks
Sol medium for more complex impl
Sol high-xhigh for planning and review
Why Terra despite benchmarks? Luna seems to be distilled smaller model which tends to cut corners and lie. Even though it passes benchmarks little bit better for same cost, I'm not risking using it for serious stuff.
Sol medium vs terra high is questionable though. I would like to test if it worth to just use sol medium instead of terra.
Sol low doesn't make economical sense to use vs other models according to benchmarks.
Here's 5.6-sol pro's recommendation:
ReplyDelete| Task | Model | Effort |
| --------------------------------------------------------------------------- | ---------------------- | ------------- |
| Mechanical, isolated, explicitly specified patch | `gpt-5.4-mini` | low or medium |
| Normal implementation from a completed plan | `gpt-5.6-luna` | medium |
| Implementation requiring substantial debugging or cross-file reasoning | `gpt-5.6-luna` | high |
| Ambiguous architecture change, difficult migration, repeated worker failure | `gpt-5.6-terra` or Sol | medium/high |
คู่มือของคุณวัดจาก โควต้า ของคุณอะนะ นี้คือ คู่มือ หรือแค่มาเล่าให้พวกเราฟัง
ReplyDeleteHow often do you have to redo a task in a higher model/effort though? Redoing a task again means you burned more tokens than if you just started with the higher model/effort to begin with. Also, there are the invisible unknown unknowns that creep in when a model can't handle the task but does things with confidence. I just opt for the best availalble, and have it spawn subagents for subtasks it deems appropriate. It generally seems to know better than I would.
ReplyDeleteI've quickly discovered that Sol Ultra is insane. And I don't mean "insanely good," I mean "deeply problematic and probably unusable for most tasks."
ReplyDeleteI have a few projects that are reasonably complex, so I applied Sol Ultra to audit them and identify problems. My experience is that, given an open-ended prompt, Sol Ultra consistently demonstrates two problems:
Problem #1: Massive scope creep. Given a simple instruction, Sol Ultra will blow it up into authorization to do all kinds of things that aren't required or implied by that instruction.
- Example #1: I had an idea for changing a feature. I asked Sol Ultra to review it, analyze it, and provide any recommendations. Sol Ultra ran off and thought for over an hour, and when I stopped it to ask what it was doing, it explained that it was finishing implementing the feature that I had specified across the entire codebase. I didn't ask for an implementation, I asked for an analysis.
- Example #2: I asked Sol Ultra to audit a codebase and to write up some recommendations to ./temp/PLAN.md. Sol Ultra did as asked. Then, I provided my responses for the first three and asked Sol Ultra to record my answers in ./temp/PLAN.md. Again, Sol Ultra ran off and silently processed my request for 20+ minutes. When I stopped it ask wtf it was doing, it said: "I was re-audit the entire codebase and documentation to ensure consistency with your instructions." I didn't ask for a re-audit, I asked for it to record my answers for future use.
- Example #3: I asked Sol Ultra to perform one phase of an audit that was described in ./temp/PLAN.md. It suggested adding a matrix of features to requirements to ./temp/PLAN.md, which I approved. Sol Ultra went off and worked for 2+ hours, and when I stopped it, I found that it had generated:
./temp/PHASE1_CONTRACT.md
./temp/PHASE1_REQUIREMENTS.md
./temp/PHASE1_LIFECYCLE_MATRIX.md
./temp/PHASE1_PROTOCOL_MATRIX.md
./temp/PHASE1_JOURNAL_AND_CONVERSION_MATRIX.md
./temp/PHASE1_FEATURE_ACCEPTANCE_MATRIX.md
...several of these as 200kb+ files with over 2,000 lines of text. Just a metric shit-ton of documentation, enormous and bloated and ridiculous. And any change to the architecture, no matter how trivial, would require reviewing and updating all of these documents as well as the actual codebase.
DeleteTakeaway message: Sol needs boundaries. If Sol is working on a request much longer than you'd expected, stop it and ask what it's doing.
Problem #2: Preference for complexity without value. Sol Ultra can overdesign anything to include features that are totally disproporationate to the context.
- Example #1: I asked Sol Ultra to review an app framework that contains instructions for sessions to build macOS apps. The build pipeline is straightforward: generate the .app bundle, copy it over the existing .app bundle in /Applications, and restart the running process. Sol Ultra proposed an intensely overcomplicated "swap/register/verify" mechanism that includes keeping every old .app version for rollback. Bonkers level of overdesign for a simple build process.
- Example #2: The same app framework requires apps to perform a health check on app start and to report any problems in the Preferences dialog. Sol Ultra raised the alarm about "blocking health errors that prevent the Preferences dialog from being shown" and proposed a "recovery dialog" to deal with these issues. This is an app framework that applies to every single project, from Calculator on up, and Sol Ultra thinks that they all need a mandatory "recovery dialog" like Windows has. Ludicrous.
- Example #3: Some of my apps talk to each other via IPC. Sol Ultra raised a cybersecurity risk of apps eavesdropping on one another, recording responses, and engaging in impersonation and injection attacks by replaying responses. Sol Ultra proposed a cybersecurity mechanism where every response includes a cryptographic signature generated over the body, protocol, request, response, and timestamp. This is for a set of apps that I designed for strictly personal use and that use IPC to talk to each other in my trusted environment. Absurd.
In general, Sol Ultra feels like it was trained on high-security, ultra-robust projects - stuff like financial services or cryptocurrency, where audit trails and security are essential - and that it wants to apply those design principles to every project regardless of scope or context. It could blow up "Hello, World" into a 10MM-line codebase with its own quantum-secure git repo.
Other problems of Sol Ultra:
- A tendency to stop providing status updates and engage in long chains of tool calls and responses without Thinking output, reducing visibility in what it's doing, which probably amplifies the "went off the rails" scope-creep behavior.
- An amplified Amelia-Bedelia-like tendency for literal interpretation in isolation, where it cannot understand statements in context.
- The same fundamental deficiency as 5.4 and 5.5: It does not verify facts before reporting them, resulting in false factual statements.
My overall conclusion is that Sol Ultra is... not good. It is like a severely autistic individual, subject to extremely literal instructions, extreme tangents, and obliviousness to context and proportionality.
I intend to use Sol Medium or Sol High for most tasks, and might step that back to Terra if those behaviors can't be reined in.
"I only escalate when the previous model actually struggles"
ReplyDeleteBut then you're wasting tokens on struggling instead?
Luna high all day today for chunking through discrete tasks, planned out beforehand with sol medium. Has been nice. Also realised I was chewing through token usage because my speed setting was on 1.5. I don’t even remember enabling that so worth checking if you don’t need it
ReplyDeleteHow do you now the current model is struggling?
ReplyDeleteI'm currently creating new features in my project and I'm using Medium SOL to plan and execute the plan. Why am I using this strategy? Medium SOL is very good at planning and executing, it basically doesn't come with errors and it creates automatic tests. It's much better to pay a little more and have a good product. Cheaper models leave errors that you then have to correct, wasting time and spending more money.
ReplyDeleteI’m on 20x plan, am I The only one thinking is like having infinite usage?
ReplyDeleteWhen I first enabled 5.6 terra medium, I explained to it the frustrations I had with 5.5 (gawd the slop, lazy, sub par performance drove me to the brink of canceling my openAI sub).. the contrast of using claude cli and codex cli makes codex cli like something that is deliberately working against you, ignoring custom rules, not using skills, and shipping sub par experience) that said, so far terra medium has deliberately delayed, inject dozens of arbitrary pauses, claims to be doing things "off-screen" when it isn't not... sits at the prompt waiting. So I just now enabled Sol high hoping for a better experience, I am at the end of the line for open AI codex cli... interesting enough, I get WAY better results by using codex web, and even codex desktop app. Something has to give.
ReplyDeleteReset the Codex saved memory. That's probably your issue.
DeleteI concur. My own testing concludes Luna high is comparable to 5.5 very high, but with slightly lower token usage, and slightly better output (“better” here meaning Luna will go a little further than 5.5 in terms of executing the task given against the existing scope and structure)
ReplyDeleteSol should only be used for complex refactors or persistent issues the other models struggle with.
Terra seems somewhat pointless at the moment, every test it seemed to take longer than other models, burn more tokens and then eventually timeout. Could be a capacity issue, but my experience with it has been poor.
I am on Pro 20x plan and had no issues with running Sol Ultra with fast mode the entire time on a big cleanup and refactor with Superpowers Skill set. Only problem is that during implementation phase where it frequently reviews its work with sub agents it ends up making almost no real progress because an agent figured out that we need to harden the code against quantum computing and a potential alien invasion despite the task being relatively straightforward.
ReplyDeleteI gave it a redesign plan with a lot of designs from claude, its been 2 fucking days, 2 whole days of it working, I'm going insane like wtf.
DeleteToo true have tried dialing back to Sol medium might just try Luna on xhigh see what progress we make.
Delete😂😂😂😂 I genuinely burst out laughing in public! it really does love to overthink security, it's really good till it's not
DeleteWhy don’t use 5.5 xhigh for me that does the trick and I don’t feed like I miss anything. Is faster and burns much less tokens
ReplyDeleteSol xhigh in hermes orchestrating Luna xhigh in codex cli. 5 hour quota on 20€ (plus) subscription gives me ~1 hour of heavy project work currently. For me the GPT 5.6 quota limits are surprisingly generous on that small sub
ReplyDeleteWith my resets left I’m struggling to max out my usage before they expire. Running 4 projects on sol high currently in parallel. 2.5 hours left and I’m at 25% of weekly still remaining.
ReplyDeleteSol ultra and max only.
ReplyDeleteI’m driving with sol daily, because my code is complex, Terra and Luna eat up way more usage than sol, just trying to figure out what they’re looking at
ReplyDeleteTerra medium has been fine for me token wise.
ReplyDeleteIf you are to believe the Artificial Analysis benchmarks on capability vs cost, There doesn't seem to be a place for Terra. Luna xhigh or Sol on the lower thinking basically close the gap. It was posted in one of the AI Reddits but I'm too lazy to go fetch it.
ReplyDeleteI dont use it heavily because I cant 2-4 prompts and I have to wait 5 hours
ReplyDeleteEstou usando no direito para redigir peças processuais.
ReplyDeleteO luna no máximo consegue analisar bem processos e organizar documentos para protocolo, o sol ultra consegue redigir petições complexas de temas extremamente avançados sem precisar de qualquer correção.
https://imgur.com/a/Pwy9XnI
ReplyDeleteThis is my config now run with sol as orchestrator. basically skipping all of the terra but i don't dare to go luna xhigh for now, usually theyt hink too much on xhigh
for implementer, did you consider 5.4-mini? it's cheaper and i've had good results for month, so i'm hesitant to move to -luna. Lots of things describe it as if Luna is closer to spark, but it's so early i don't really trust hot takes like that yet
Deletexhigh or max?
DeleteWhich one ? Luna or sol? Sol alternate between low-high
DeleteLuna mostly high, usually on xhigh model overthink too much
Luna has the most advanced level, the max, but it needs to be activated in the settings; it was disabled by default.
DeleteYes i dont dare to touch the higher level of reasoning for now (as i only have plus plan too).Have you seen much benefit increasing it to max for you?
DeleteDid you make this sub-agent yourself? If it's a published/publicly available one, I'd like to know where to get it!
DeleteYou can do the same thing out of the box in Pi agent with PI subagent addon.
Deleteyea i did make this myself i tell GPT to do it based on my usage patterns.
Deletehttps://github.com/hindraxxx/subagents_configs i do have installable file so i can share it to my friend but it hasn't been tested yet.
do notes that i introduced an additional SUBAGENT_ROUTING.md that's imported inside the AGENTS.md so that by default my agent always uses this smaller subagent to reduce my token usage. Feel free to take reference/ use the installer above
I have a similar set up including an agent dispatch md. I haven't really tinkered much yet, but is there a way to know if it actually spawned a subagent with the pre-determined model? It's told me that it only inherits the session default/model. I asked it to do a test, and then it tells me it can't select a model or the feature isn't available.
DeleteYou can add a specific command in agents.md so the orchestrator / subagents to explain itself (what model and reasoning did they uses) when a task delegated
DeleteThanks! I'll give it a try.
Deletewhat plan you have and how long does 5h limit last with this setup when actively working?
DeleteIts plus, and considering how token hungry 5.6 is.. i made like 6-7 changes (small-medium changes with some back and forth minor adjustment / steering) with my 5h limit which i guess are fair.
DeleteUsually i exhausted it 2-3hour into the session
Sol Ultra when you have 1% usage left, you know what to prompt 😉
ReplyDeleteWhat’s the prompt?
DeleteIt will still eat up your weekly limit. Tho, it bypass the 5-hour.
DeleteI already knew this but wasn't sure can you explain ?
DeleteHaha. It works only for 5h limits
Deletewhat prompt do you mean?
DeleteI had a similar finding and shared my workflow here https://github.com/Jogan/soluna-workflow. I'm using Sol Low as my driver.
ReplyDelete> Unclear task that requires exploring several parts of the repo
ReplyDeleteI ran a single (n=1) benchmark and this is what stood out to me the most. Terra used a lot more tokens than the other two but it also found a lot more issues. Sol was the winner on being both a strong model and token efficient. Due to that I use Sol low for the most part. It more expensive but it uses less tokens so its not more expensive.
So for me its
- Sol low as a smart default
- Terra for exploration
- Luna for edge cases and small tasks
I don't see Terra as less expensive. Its just different. Half the cost but using 2 times the tokens isn't cheaper. Having a different model is great for parallel agents because they're going to behave differently and go down different paths. Deploying all three for deep search is great.
I’m aware benchmarks aren’t that reliable but they showed that for every level of effort of Terra there’s another model that’s as intelligent as a lower cost.
DeleteInstead of Terra low use Luna high, Terra medium -> Luna xhigh, Terra high -> Sol low, Terra xhigh -> sol medium
They are still good for adversarial review just as you pointed out they are different models and may come with different things. But I just use Fable 5 or Sonnet 5 for that too.
I’m starting to think “difficulty” is the wrong axis. I’d rather escalate based on blast radius: a hard but isolated refactor can stay on Luna, while a simple auth or config change might deserve Sol.
DeleteThat’s a good point. Cost per token and cost per completed task are definitely not the same thing. My experience with Terra was similar — it explored much more, but also burned far more of the limit. I haven’t tested Sol Low enough as a default yet, so I’ll try it on a few normal tasks and compare it with Luna XHigh. What kind of task did you use for the benchmark, and which reasoning levels did you compare?
DeleteIt was a code review on a PR. Checked for correctness and other stuff that Sol recommended checking.
DeleteAll low effort. They all had one issue they found in common and they all found issues the others didn't. I think running all three at the same time would produce really good results for code reviews.
Speed Luna > Sol >>> Terra
Token efficiency Sol > Luna >>> Terra
Focused review Sol
Broad tracing Terra
Distinct edges Luna
Compared with Sol:
- Luna finished about 14% faster.
- Luna used about 66% more tokens.
- Terra took about 2.6 times longer.
- Terra used about 2.9 times as many tokens.
i jus use sol xhigh on 5x and then lwkey go on twitter to complain.
ReplyDelete😂
DeleteI already have a “task-sizing” skill on my main project that determines many things based on the task complexity. I just tried adding model evaluation to it. Unfortunately Codex doesn’t know what it is. :( just gets GPT-5 (leading me to believe they route our requests as they see fit, and selector is mostly cosmetics).
Idea was that if you launch a task that is not well suited to model capabilities, it will stop and recommend an appropriate model.
I think this should be built in to Codex. Having 40 different variants is silly.
Comment deleted by user
DeleteWhat makes you think that?
DeleteLow-key the method 😉
Delete"tibo wen reset"
DeleteYap that seems like a lot of people work flow, sol ultra then post on Reddit or X, Tibo when reset. Wait 1 hour, get nothing just bitch posting about how OpenAI is scamming you.
Delete100% I can’t help but laugh when people complain about ultra eating their usage like??? Are we so fr 💀
DeleteUltra is like a top fuel drag car
Medium - high is ur GT3 Nurburg winner
Ya. I saw the way they show the kind of prompt they use. If you going to put build me an app to make money and out ultra, I am not surprise max 2000 account is also insufficient for you.
DeleteEver since all those vibe coder refugee come to codex, that is the kind of post we see everyday.
This graph seems pretty useful
ReplyDeletehttps://x.com/ArtificialAnlys/status/2075739292052463646
That's really good to know, thx!
DeleteWould be better if they added a time per task label to each point
DeleteI don’t think this release was well thought out. Too many overlapping options, unclear guidance from codex, contradicting reports of which model and raining level works well for what tasks.
ReplyDeleteMany are as confused as you are.
It is genuinely quite confusing. For the second day in a row, I have been clicking through Reddit, reading everything, and experimenting with the different models. Gradually, you begin to develop a sense of how they work.
ReplyDeleteIt is also important to understand that all Ultra tiers across the models use sub-agents, although these are only necessary for more important tasks.
Just use Luna Max and plan with any higher model you can afford. Simple as that. Don't overthink it
DeleteThey're likely gaining valuable insight into how people use these different strata for particular tasks. I'm sure they are very aware of what they're doing UX-wise.
ReplyDeleteSomeone explain to me how 5.6 Terra on XHigh eats through my entire 5 hour window in 1 prompt whereas 5.5 xhigh is supposedly more expensive and less token efficient and yet lasted me significantly longer. I think something is buggy with the newest model release / ChatGPT codex combined app.
ReplyDeletesame. 5.6 seems very expensive. If the price is "lower" then they might have changed the tokenizer. Anthropic did the same with Sonnet 5. Cheaper on paper per token, but the tokens are smaller now so you consume more tokens and thus more credits/money.
Deleteit makes no sense to have multiple intelligence levels for each model. The people here might do some research and ask to find out best use cases but for the vast majority this is decision paralysis. If the highest intelligence is higher than the lowest two or three of the next tier model then they shouldn't exist because the lower tier model will probably be more cost-effective or faster.
ReplyDeleteAnd if they want to stick to the current breakdown then each tier model should be very clearly specialized for a certain task.
I would go further as to provide only the 3 tier models with a default intelligence router depending on the task. eg planning, execution, research, etc. But of course, they should have done a comprehensive internal eval for the router which should be under continuous refinement to allow teams more time in this fast paced environment. There should be a dedicated team studying and optimizing this under the hood to offload the cognitive tax of the end user.
But of course, we are not forgetting the powerusers building specialised agentic workflows who might want to do their own internal testing to find the most optimal tier models for each workflow. For them, we would have the option to enable all intelligence levels.
This is me my take as a former UX designer.
Your take makes sense from a UX perspective, but I think we're just not ready for this yet.
Delete"I would go further as to provide only the 3 tier models with a default intelligence router depending on the task. eg planning, execution, research, etc."
Nobody has figured out the answer to this yet, and there is a lot of variance across use cases, which is why your proposed UX cannot be implemented.
"But of course, they should have done a comprehensive internal eval for the router which should be under continuous refinement to allow teams more time in this fast paced environment. There should be a dedicated team studying and optimizing this under the hood to offload the cognitive tax of the end user."
There does not exist a canonical set of representative use cases for anybody to test this meaningfully. Millions of users are discovering millions of new use cases every day.
it took decades after websites were created for the term UX designer to be coined and formalized with actual roles within companies. Some of the ideas I talked about aren't that easy behind the scenes but if even these trillion valuation companies with all that internal AI compute can't figure it out then 1. what does it say of the AGI they keep selling that is around the corner?, 2. how do they expect people to understand which model tier and effort mode to choose from? specially as you mentioned, the variance, not just after a new model release but what's truly going on behind the scenes? the quantizations, system prompt injections, model checkpoint switches, internal harness, etc etc.
DeleteImagine if one had to select the effort every time we say Hey Siri, use effort Terra Medium... that will never happen for so many obvious reasons so it means they will have to figure it out eventually and whoever figures out the optimal algorithm like google did with the single search bar (solving queries with the lowest actions and decisions) will come up ahead.
we basically advocate and design around human limitations.
Codex right now is a product for early adopters who are willing to tinker and put up with unpolished UX. It will take time to mature and get to a more polished UX. Ref: https://en.wikipedia.org/wiki/Technology_adoption_life_cycle
Delete"Imagine if one had to select the effort every time we say Hey Siri"
Apple's strategy is not to cater to early adopters. They wait a few years until a product category is mature before integrating it into iPhones.
Luna seems to be the best value right now, mind you its not just a "mini" model, its MoE.
ReplyDeletesomeone made a post about cost per intelligence already: https://www.reddit.com/r/codex/comments/1us6pxy/the_56terra_is_perhaps_the_least_efficient_model/
In our team’s benchmark Terra xhigh has been cheaper than 5.5 medium but the intelligence is way higher. Sol medium is already quite expensive but Sol low is quite cheap and intelligent. Luna has not been good enough for our use but it is very cheap.
ReplyDeleteIf you use 5.5 high, I would recommend to start using Terra xhigh. It should be better and you should get more use. Go even Terra max if you do not hit limits or mix sometimes plan with Sol high/xhigh where it matters.
Thanks man, this is the kinda inputs I needed
DeleteAlso have a look at DeepSWE
Deletehttps://deepswe.datacurve.ai/
Now, that's a gem. Thanks a ton!
DeleteUse Sol high and it ate the hell out of my usage on the 5x plan. Was surprised after how much success I had with 5.5 and almost never hitting my 5 hour limits.
ReplyDelete5.6 Sol Light is stronger than 5.5 Extra-high and consumes less too
Deletesame problem, switch to terra xhigh and its a lot better in cost vs gpt5.5 xhigh, but im not sure how intelligent it's, what I can tell you is terra xhigh found bugs that fable medium didn't, and in my code comparisons before this release -> gpt5.5 xhigh write better code than fable medium (this was the only one that didn't eat the tokens and do an ok analysis of the code)
DeleteYes I'm having the same issue. I find that terra consumes limits similar to 5.5 but sol consumption is much higher.
DeleteI know right. Something must be wrong ... even Sol light used as much as 5.5 high/xhigh for me.
DeleteMy macOS chat just give a project to sol5.6 high without my consent and my 5 hour limit just gone… Now he’s just apologizing about why he should asked me first🤣
Deletei have this issue even with 5.6 luna...
DeleteHonestly I’m finding Luna to be a beast, xhigh is feeling comparable to 5.5 medium to me
ReplyDeleteFor narrow tasks it's fine. But tasks requiring world knowledge..I guess luna might lack there
DeleteLuna is really great if it is enough for the use cases. In our tests it tend to cut corners even on max effort and it has the distilled/benchmaxed feel. When the use cases are right, it is a beast but if use cases vary, then I would not recommend it, if the budget allows Terra.
DeleteOh nice , so terra will be one step better than both of these then right
Deleteyup it's like the mid-tier. i've been reluctant to try it since i was seeing some reports of it being worse value since luna is so much cheaper. lunas actually dealing with some bugs now for me that 5.5 was unable to sort. IMO from limited stuff i've known might be best to stick with sol for truly hard stuff and luna xhigh and maybe terra for everything else, but i have a feeling most the things terra can do luna can do as well, but cheaper. supposedly we'll get 2 resets today to try the models so, todays the day to fuck around and find out ;)
DeleteFor each level of terra there is a higher level of Luna that matches its intelligence (in coding at least) except terra max, but terra max is very inefficient. Basically, there is no use case of terra via api, the usage could be calculated different in actual codex.
DeleteOhh cool then I’ll use luna xhigh for a day and see how it suits my workflow
DeleteBuddy it was said to roll out to all subscriptions over plus in the next 24 hours not the other way around 😂😂
DeleteDamn my bad man 😭
DeleteWhere did you see Sol being removed? That's news for me
DeleteModel Input Cached input Cache writes Output Input Cached input Cache writes Output
ReplyDeletegpt-5.6-sol $5.00 $0.50 $6.25 $30.00 $10.00 $1.00 $12.50 $45.00
gpt-5.6-terra $2.50 $0.25 $3.125 $15.00 $5.00 $0.50 $6.25 $22.50
gpt-5.6-luna $1.00 $0.10 $1.25 $6.00 $2.00 $0.20 $2.50 $9.00
gpt-5.5 $5.00 $0.50 - $30.00 $10.00 $1.00 - $45.00
gpt-5.5-pro $30.00 - - $180.00 $60.00 - - $270.00
gpt-5.4 $2.50 $0.25 - $15.00 $5.00 $0.50 - $22.50
gpt-5.4-mini $0.75 $0.075 - $4.50 - - - -
gpt-5.4-nano $0.20 $0.02 - $1.25 - - - -
gpt-5.4-pro $30.00 - - $180.00 $60.00 - - $270.00
The token pricing is cheaper and 5.6 takes less time to built stuff therefore it must be more efficient than 5.5, install codex-lb to monitor stuff
What is the cache duration on the 5.6 models??
Deletehttps://developers.openai.com/api/docs/guides/prompt-caching#prompt-cache-breakpoints
Interesting that these are now 30mins flat but have implicit breakpoints, previous models all had 24Hr
https://imgur.com/a/2hPoP9O
Deleteor the pooler https://github.com/icoretech/codex-pooler
Deletethanks for the table.
Deletewhy does it have redundant columns and which are the relevant?
I copied the entire oficial table from https://developers.openai.com/api/docs/pricing and I gotta admit i dont understand some columns.
Deletethanks. you forgot to add the labels above those columns
Deletelong context is >= 272k
https://imgur.com/a/0Xwv1Gi
I only have 5.6 in codex but can’t access api yet?
DeleteThank you bro so using 5.5 or 5.4 with an idea of saving tokens in mind is of no use then it’s better to switch to 5.6 thanks to you, I kept thinking using older models will use consume lesser limit or token
DeleteWell that was openai promisse with 5.6, suposebly "half the price of 5.5" but with greater capabilities, for what I have tested I experienced a greater consume comparing to 5.5 yesterday but it was a different task and different projects so I will trust it as an exeption , just purchased Pro x5, lets see how it goes
DeleteIt feels like they are funneling you towards 5x, because tasks which I did with 5.5 on xhigh now drain 10x as fast on 5.6, which I don’t understand how that is possible if 5.6 is supposed to be the same price wise (sol)
DeleteI discovered that 5.6 Sol light is better (for my application) than 5.5 Extra-high even without Fast mode.
DeleteYeah, but it seems fair to me honestly , they are dropping banger after banger, 5.6 , pets , cleaner UI design, etc , + if you plan to use codex to make money like I am currently doing you have to look at it like an investment, not a cost
DeletePets?
DeleteSwitch to 5.6, there is no reason to not do so. Currently the usage is a bit nebulous, sol uses more than 5.5 even though they are the same price, but you can drop to Luna xhigh/max and it holds up better than 5.5. The only way you can test this yourself is by using them, but overall even the weakest model in 5.6 is better than 5.5. Terra is to be avoided in API (Luna is just better value) but maybe it makes sense using it on the sub. It's all nebulous since openai doesn't really disclose how they charge usage.
ReplyDelete5.6 Sol Light is stronger than 5.5 Extra-high and consumes less credits.
DeleteVery relevant reply, for some reason you cannot just compare api pricing and assume it will hold for the subscription; very simple queries on sol drain like 5% usage per back and forth. And then I am talking about trivial tasks like interacting with programs on the CLI.
Delete(I know that this is not the purpose of sol, but it is just a test and 5.5 xhigh felt like unlimited usage in that regard)
Thanks will keep these in mind.
Delete