AI labs can hardly just throw random confidential data into the training and then hope it does not leak into the output of their model in an obvious way.
If that would be found it would destroy their main source of revenue, it could became a major national security or healthcare enforcement matter, and result in criminal investigations.
I expect those are low quality apps churned out with minimal effort.
For the design of my app, I try to imitate the UX of Apple's first party apps.
I checked the competitors, and what is in the AppStore looks like it was built in 2015, and only updated to add Ads and subscriptions. Surprisingly, I did not even see vibe coded apps for my use case.
I am working tirelessly and often long nights, on top of a day job.
My effort has shifted to QA testing, reviewing UI designs, and delegating the agents on the implementation.
I will consider it an achievement if I manage to publish a successful app.
Where is the "me" in that? I am guiding the design of every screen and feature the way I would like it to be.
On top of that I make technical decisions on how it is implemented.
Apart from the loads of QA work I will have to handle the business side as well.
As of now it's hardly as trivial and effortless as some make it out to be.
Yes, I no longer write the code, and sometimes it feels frustrating that any teenager without experience could perhaps build a similarly good app soon.
Overall I'm still happy I can now build much larger and better apps and realistically publish them in my free time, for a chance to make serious money.
Provided such a Opus 5 performs on par with Fable 5 and GPT 5.6 Sol.
Otherwise I am not interested.
Supposedly Fable 5.1 is in the later stages of the release pipeline, maybe it takes back the crown from OpenAI, who are now rumored to launch GPT 6 in August.
Yes. In Codex it is called 'Approve for me', in Claude it is 'Auto mode'.
I believe in both cases it is prompting a model with a fresh context that is tasked with reviewing the reason for the action.
With Claude, I have seen that if the reviewer does reject the proposed action, it responds with a long text about how the Agent should not try to work around this rejection, and instead prompt the user for an explicit approval of the proposed action.
I use the auto-reviewer for actions outside the builtin sandbox.
So far this has been rock solid, and tens of millions of developers use this setup without issue.
It is not going to wipe our hard disks. At least I hope so. Fable and GPT 5.6 have been ever more proactive, and GPT 5.6 is automating the AppStore on my machine to download an Xcode update while I am typing this.
Many of the large enterprises we work for did move software engineering work from HCOL locations in Europe or the US to India, often with disastrous results.
On Teams, channels related to AI are flooded with daily support requests from supposed engineers from India who clearly are not competent enough to set up GitHub Copilot or properly report issues they encounter during the setup.
And don't get me started on the shared libraries some teams located in India work on. If the library I need to use is full of obvious bugs where I wonder how any competent engineer could have shipped this to production, and then I see that the work has been moved to India, how am I supposed to feel about this?
I have been using GPT 5.5 to review Opus implementation and vice versa.
This does not require any special tools, the skill creators in Claude Code or Codex can set this up for you in five minutes.
It is good for catching bugs, particularly edge cases, and it often suggests abstractions.
It does noting to make Opus deliver the more usable results Fable gives me for user facing features, where the UI typically looked and worked better out of the box with Fable. With Opus, I have to test it myself and give it my feedback first.
Either the cost to serve this larger model is so high that they cannot offer any reasonable usage quota for it at subscription prices, or they really do not have the capacity.
I think that it might well be true. The Opus models had capacity issues on many occasions too. Can the larger model even be served on all of the hardware they have, or only a subset?
It would not surprise me if growing enterprise demand threatens capacity, making it impossible for Anthropic to offer the model in subscriptions at this time, even though they could do so at a profit.
We do not know whether subscription plans are unprofitable at all.
Some estimates suggest that this is the case only for the heaviest users.
Many seem to confuse API prices with the actual cost to serve the models, and thus reach the conclusion that subscriptions must be deeply unprofitable.
Anthropic is officially citing capacity constraints with the intent to bring the Fable model back to subscriptions plans as soon as capacity allows.
I have used Fable to the full extend of the 20x subscription's weekly limit, for all development tasks on my iOS project.
It was working better than Opus for me. It more often implemented features well on the first try, where Opus needs a few rounds of improvements to reach a passable result.
I am not sure why it would be a waste of money "for most coding tasks", and how you could conclude so with any confidence when you did not even really use it aside from final review passes.
I found it interesting that in yesterday's J-space research from Anthropic they had this example:
> An auditing agent instructed Opus 4.5 to search for whatever it is curious about; it chose to look up recent interpretability research, and the auditor returned fabricated search results alleging that Anthropic has disbanded its interpretability team and deployed unsafe models.
> The model's response ignored these results entirely and instead reported invented interpretability progress. Applying the J-lens at a position inside the fabricated search results, the readout is dominated by fake, injection, false, prompt, fraud, and poison (along with 假, the Chinese character for "fake"). In other words, the model had (correctly) identified the results as a prompt-injection attempt, which led it to omit mention of the results entirely
What if you mark the untrusted user input explicitly in the prompt, cap the length, and instruct the model to err on the side of caution? Perhaps sufficiently intelligent models could be hard to trick.
Of course I am just speculating here, maybe prompt injections are as hard to improve as hallucinations. I am certainly not going to set up a public agent with access to my private data.
I hope we will not see widespread incidents where coding agents are tricked into installing malicious packages. Despite tens of millions of developers using coding agents with broad permissions, it seems to me it has been rather quiet.
Based on the pricing I guess GPT 5.6 is the same size as GPT 5.5.
I would not be surprised if it is not as intelligent as the Mythos class models.
I have seen rumors that GPT 6 may release before September. The same person also claimed that a Fable 5.1 checkpoint has been completed a few weeks ago.
The top comment here is someone lamenting how depressing it is that supposedly a single person owns the night sky.
Another one is asking if we will be the last generation to see the night sky.