Author: Digital Life Kazik
I'm numb. Weren't people just a few days ago all shouting that models should slow down? How come these models are being tossed out like they're free? Yesterday Grok 4.7 PK MiMo v2.6, was that just foreplay for today's two big guys? Early this morning, OpenAI's GPT-6 Sol and GPT-6 Luna, as well as Anthropic's Claude Opus 5.5, were all officially released.
I really want to ask, every day everyone says slow down, where exactly is the slowdown? All scumbags. And this plot, I always feel it's inexplicably familiar, then I flipped through historical articles.
Last time, GPT-5.3 Codex went head-to-head with Claude Opus 4.6, today history directly repeated itself... Except, Claude Opus 4.6 became Claude Opus 5.5.
GPT-5.3 Codex became GPT-6 Sol.
The two companies are still these two companies, still that same pair of rivals who won't back down.
Let's go one by one.
I. Claude Opus 5.5
First, Claude.
Frankly speaking, no matter how much I dislike Anthropic, no matter how much I think this company is an idiot, no matter how many times they ban my account, objectively speaking, Claude is still the best and most comprehensive tier of model in my mind.
I still remember to this day the shock Claude Fable 5 brought me.
And this time Claude Opus 5.5 is the first model in Anthropic's brand-new Claude 5.5 series, and from what I remember, it's rare for Claude to jump so many version numbers at once.
And from my own temporary experience (don't ask me why I can still use it after being banned, I just buy disposable accounts, the kind that get banned after 2 hours of use...), Opus 5.5 in the entire communication gave me a feeling of returning to Opus 4.6, with a very strong sense of being alive.
Although their positioning for this thing is actually the cost-effective model in the Claude series, it really does have a way of reminding me of my old white moonlight.
I summarized the basic information, roughly like this:
Model name
claude-opus-5-5
Context
1M Token
Max output
128K Token
Reliable knowledge cutoff
June 2026
Thinking
Adaptive Thinking, always on
Default effort level
medium
Input price
$4 / million Token
Output price
$20 / million Token
Cache read
$0.2 / million Token
5-minute cache write
$5 / million Token
1-hour cache write
$8 / million Token
Compared with the previous generation Opus 5, the overall work cost can directly drop by 40%.
At the same time, output speed is more than 30% faster than Opus 5.
Currently on some non-large-scale tasks, Opus 5.5 can already achieve Fable 5.1-level performance, while large, high-difficulty tasks still require Fable-level models to step in.
On Coding tasks, it's basically the current SOTA.
For example, Terminal-Bench 4.0, which reflects the ability to complete complex, multi-step tasks in a real terminal environment, Opus 5.5 set a new high score of 66.4%.
Anything involving Coding is basically SOTA, very much like when Opus 5 first came out and comprehensively surpassed Fable 5 in Coding execution.
But two things are different. Terminal-Bench-Science, this task is more research-oriented, Opus 5.5 is 58.7%, while Astra is actually higher here, 64.6%.
AutomationBench is the same. This task is actually cross-software work, Opus 5.5 is 40.0%, GPT-6 Astra is 41.4%.
And that Computer Use, although GPT-6 Astra has no score, everyone knows this thing can't possibly be bad, it's the absolute SOTA.
From these places you can also see some of the differentiated development directions in which Anthropic and OpenAI are competing. Anthropic leans more toward Coding ability, and also has extremely strong aesthetics. They also believe Coding is the cornerstone of the path to AGI, but OpenAI is more focused on comprehensive Agent capabilities, including reasoning, software operation, scientific research, and so on.
But frankly speaking, simply looking at an evaluation set in one dimension, I think it's completely impossible to assess a model's true quality.
Like you know, even when developing products, there's still a very core thing, which is your upfront planning and architecture design ability. This thing, Fable is a god at it. It depends on your large parameter count, on your world knowledge, on your emergent intelligence.
You say Opus 5.5 really can surpass Fable in certain specific executions, but you say overall Opus 5.5 is stronger than Fable? Then I think you might as well believe I'm Qin Shi Huang.
Overall, Claude Opus 5.5 is basically Fable 5.1's own official distilled model.
In many vertical scenarios it may be stronger, and has a smaller parameter count, is faster, and cheaper, but at the same time it also brings Token waste. This is actually a trap. Once on high-difficulty tasks, some non-flagship models will instead continuously think crazily, try crazily, but still can't get out and can't solve the problem, and instead burn more money than flagship models.
Just like Opus 5.5 is indeed very strong, but its output Token count per task directly exploded, so actually a lot of its intelligence is bought by using super-long reasoning. The downside is that once it falls into a predicament, it explodes on the spot.
So Anthropic also gives this warning.
When your results are important, you must use the highest-level model.
Then you can look at the cases of Claude Opus 5.5 run by big shots on X, still quite stunning, especially in aesthetics and details, with huge enhancements.
From @notjazii's test comparing GPT-6 Astra and Claude Opus 5.5, in some scenarios it's even a bit more refined than GPT-6 Astra:
But the cost is that compared with GPT-6 Astra, it's slower and more expensive.
However, for Opus 5.5, comparing it with GPT-6 Astra is itself a kind of praise.
Opus 5.5 also strengthened communication this time.
Anthropic said one of the most common pieces of feedback they received for Opus 5 was that sometimes when writing things it was more roundabout, had many terms, and wasn't direct enough in expression.
5.5 puts the most important information first, reducing strange wording and unnecessary explanation.
At the same time, in creation, it also has the feel of Claude 4.6.
So, in summary, Claude Opus 5.5 is a model I think is very great.
Anthropic has pushed down a lot of work that only flagship models could do in the past to a price point that can finally be used regularly. I guess it's also because the pressure from OpenAI in the past was too great.
Their own cost guide directly suggests:
For daily Feature, Debug, Code Review that you keep an eye on, use Opus 5.5.
For cases where the actual result matters more than the Token price, long periods of unsupervised operation, or Opus 5.5 failing continuously, then switch to Fable 5.1.
Fable is starting to look more and more like an expert-level consultant.
Opus 5.5 is more like a main employee.
This is Claude's new model today.
II. GPT-6 Sol and Luna
Finally, something I can use normally.
GPT-6 Sol and Luna.
Because no matter how good Claude is, I'm just a disposable account, not something I can use year-round, so it's useless.
So, there's no way around it, the main force is still GPT-6.
What's especially coincidental is that a few days ago when I wrote that article about GPT-6 Pro + MCP, I even specifically wrote a line:
"GPT-6 Sol will probably be released very soon. After it's out, for execution, I'll probably switch to GPT-6 Sol without thinking, and only switch to GPT-6 Astra for high-difficulty tasks."
Then, it finally came.
Currently, it's already live in Codex and directly usable.The biggest criticism of GPT-6 Astra in the past was that it was too expensive, and the quota was completely insufficient.
So this time, GPT-6 Sol and Luna are aimed directly at reducing costs.
Astra is still OpenAI's strongest model.
For the hardest, most important work where you don't want to make any compromises, keep using Astra.
But real-world work has different scales, rhythms, and budgets.
So the significance of GPT-6 Sol and Luna is the same as the relationship between Fable 5.1 and Claude Opus 5.5:
Bringing the capabilities from Astra's generation of training methods down to faster, cheaper models.
Almost on the same day as Claude, talking about the same thing.
Who can buy more intelligence for less money.
This is the Pareto frontier.
In terms of pricing, the new generation is priced as follows.
Sol and Luna are 50% cheaper than the equivalent models of GPT-5.6.
The most outrageous thing is Luna. If we look only at input and output prices, it's actually already cheaper than Deepseek.
Of course, we all know that the real bulk is actually caching. Hmm, you can see that this cache price is still about 3 times higher than Deepseek.
However, the most suitable scenario for Luna, I think, is actually the automation tasks behind various applications, just like the dozen or so information processing tasks requiring large models behind my AIHOT, with tens of thousands of requests every day.
In this kind of scenario, the cache hit rate sometimes isn't particularly high. Overall, it's already good if it can reach 30%. In my view, Luna's advantage becomes relatively obvious.
Other basic information is as follows.
These knowledge cutoff dates are actually all different.
Luna is the newest, reaching May 18.
Then there's the GPT-6 Sol that everyone cares about most. Frankly speaking, in terms of capability, it's somewhat worse than I expected, but in terms of cost reduction, it's stronger than I expected.
You can see that on AA, it's only a little stronger than GPT-5.6 Sol, but Opus 5.5 directly took first place by a cliff-like margin.
Although AA's benchmarks are sometimes criticized by many people now, and aren't that accurate in many details, the overall direction won't be too far off.
So GPT-6 Sol is more about, based on the new model and new architecture, trying to stay a bit better than GPT-5.6 Sol while drastically reducing costs.
For example, this AutomationBench is the benchmark we mentioned above when talking about Claude, where it was stronger than Claude's side. This benchmark tests Agents across 47 tools to execute real business processes in sales, marketing, operations, customer service, finance, HR, and so on.
GPT-6 Sol xhigh:
33.2%.
Average cost per task:
$0.27.
GPT-6 Astra low is 30.3%, but the task cost is 3.9 times that of Sol.
Fable 5.1 plus Opus 5 fallback is 31.4%, and the cost is at least 8.9 times that of Sol.
Note that the comparison here is still with Claude Opus 5, because Opus 5.5 was released almost at the same time as GPT-6 Sol, so OpenAI's chart simply didn't have time to include it. If it had been included, it actually wouldn't look very good either, and wouldn't be very favorable for OpenAI.
But this chart already explains Sol's positioning very clearly.
However, what's powerful about GPT is that its token efficiency is just absurdly good. If it says it saves money, it really can save money.
And this time, compared with before, there's an update point that I think is great, which is still the factual error rate. I've always said that GPT is almost becoming my fact-checker, because it really has almost no hallucinations. It's still just too strong on this point. In the past, I always thought 5.6 was already pretty good, but this time they made another major optimization, and it's basically already at the level of GPT-6 Astra. Here you can also see the reasoning level. If you turn on light or medium, it will often make some factual errors. This is also why I always recommend that in daily use you turn it on high, and even if you're using Luna, you have to turn it on the highest.
Then I also tested it myself, and there is a clear downgrade in aesthetics and some details.
For example, controlling Blender to build a motorcycle, you can see that in terms of detail completion, it's far worse.
This is GPT-6 Astra's Temple of Heaven.
And this is GPT-6 Sol's Temple of Heaven. Many details have problems, and the door directly has a bug.
But in terms of cost, it dropped by about 70%, so overall it's still acceptable.
When I was writing the draft, GPT-6 Sol and GPT-6 Luna had already started rolling out in ChatGPT Work and Codex.
But the awkward thing is that OpenAI actually said these models are not yet available on Chat, which means that in chat mode, what's still provided to everyone is GPT-5.6 Sol.
This part is something I don't quite understand. Your costs have already dropped drastically, so why don't you also switch chat to lower-cost models for everyone?
Final thoughts
I know that by the time everyone reads this, you'll definitely still ask one question in the end.
So how exactly should I choose?
There are really, really, really, really too many models now. New models come out every day. It's really exhausting.
I can only say that I'll describe my choices to everyone as much as possible.
1. If you can subscribe to Claude without getting banned.
Logically and emotionally, I really hate Anthropic, that XX company. But from a user experience perspective, if you can subscribe to Claude and won't get banned, and can use it long term, I still recommend subscribing to Claude.
Use Claude Fable 5.1 for complex planning and plans, and use Opus 5.5 for task execution. This may indeed currently be the best combination.
2. If you've been banned by Claude but can subscribe to overseas models.
Then just subscribe to ChatGPT without thinking. Frankly speaking, at the model level, GPT is still a small notch below Claude, whether in creation, cognitive insight, or coding.
But it has almost the best consumer-facing experience, an unlimited chat mode, the best client Codex, and GPT-6 is still T0-level strong in the world.
Use GPT-6 Astra or the unorthodox GPT-6 Pro for complex planning and plans, use GPT-6 Astra high or GPT-6 Sol xhigh for execution, and use GPT-6 Luna to run large-scale batch automation tasks. It will be very delightful.
3. If you can only subscribe to domestic models.
Frankly speaking, Qwen, Kimi, GLM, MiMo, and DeepSeek, none of these have that kind of generational difference. Just use whichever one you're used to. The only troublesome thing is that the model lineup isn't as strong as those two companies. Kimi K3 and Qwen 3.8 Max are more suitable for planning and proposals, while GLM-5.3, Mimo v2.6 Pro, and DeepSeek V4.1 Flash are more suitable for execution.
Just remember, always subscribe for only 1 month, don't subscribe for too long, at most buy a quarterly card, and absolutely do not buy an annual card.
OpenAI and Anthropic are still too mature. The current environment is also somewhat different from before.
In the past, these two companies were responsible for launching frontier intelligence, in other words, doing high-end.
Then domestic models did the mid-range and cost-effectiveness.
But now, their supply chain seems to have become more mature, just like Apple, and they've also started to take all domains. Your mid-range, and even your ultra-low-end, I want all of it.
However, for all users, this is also a good thing.
Because now intelligence finally seems to be getting cheaper and cheaper.
In the past, for equivalent AI performance, its cost almost dropped by 47% every quarter. Intelligence, in another few months to half a year, may really become an asset that everyone can afford, just like coal, water, and electricity.
At that time, it may truly be the era of great prosperity.