[AI Topics] Mid-July note posts showed a shift toward the 'end of disposability' and the 're-stratification of AI'

2:00 AM. Late-night anime is playing on Abema. The volume is low, and I'm scrolling through note on my device.

By the middle of July, the atmosphere of AI-related articles on note had changed again.

Until last month, there was still a sense of the initial phase of corporate adoption, like 'what we built,' 'costs,' and 'how to integrate it.' Entering this month, it's different. A sense of practical fatigue has started to emerge, with questions like 'Will this really last?', 'Is it reliable?', and 'What should we even keep using?'

Models are no longer 'disposable'

Since around the second week of July, looking at note, new models like 'GPT-5.6,' 'Grok 4.5,' and 'Muse Spark' are appearing every day. Performance is also improving.

But at the same time, these kinds of questions have increased on note.

'Are we going to switch every time a new model comes out?'

'Which one should we actually use?'

'Is this model replacement work going to continue forever?'

In other words, models are starting to change from 'consumer goods' to 'infrastructure.'

The phase up until June of 'trying out the latest things one after another' made sense because the performance differences were overwhelming. But this month is different.

Performance caps are starting to plateau. Companies are beginning to seriously calculate the 'cost vs. benefit of switching to 5.6 if 5.5 is enough.' Once they decide, they can't change it easily. It would cause operational hang-ups.

Someone who appeared to be from an IT department on note wrote something like, 'We evaluated it based on monthly costs and signed a contract not to switch for another year.' In other words, it's a state of 'we can't get bored and cheat with another one' and 'we've bet on this.'

This is the end of 'disposability.'

Safety is becoming a competitive axis

Around the 3rd of July, news broke that Anthropic was resuming the provision of 'Fable 5' and 'Mythos 5.'

What does 'resuming' mean? Apparently, they had temporarily suspended provision before because 'safety measures were insufficient.'

After that, they incorporated strengthened safety measures and resumed it in a state where they could say, 'This is ready for society.'

Reading this far, it looks like just a piece of news.

But from a practical standpoint, this is a very important point.

When companies introduce AI, choosing based on 'performance' is already outdated. Now it's about 'safety commitments,' 'jailbreak resistance,' and 'liability in case of malfunction.'

This is especially true in finance, healthcare, and legal sectors. Compliance departments are asking, 'Has this model been externally audited?' and 'What is its safety score?'

On note, someone who appeared to be a corporate legal representative wrote, 'To avoid vendor lock-in, we are narrowing our choices to models whose safety scores are measured by neutral organizations.'

In other words, the models companies choose are being re-stratified from 'highest performance' to 'most trustworthy.'

If Anthropic wins on safety, everything will consolidate there. OpenAI will remain due to its overwhelming performance. The middle tier will disappear.

The government has started creating 'industry rules'

Among the news in July, there was an article stating that the U.S. government is coordinating to establish 'voluntary disclosure standards for cutting-edge AI models.'

The trend is that AI, which was once wild, is gradually becoming 'institutionalized.'

The fact that 'the industry and government are agreeing on standards for when development companies release cutting-edge models' means, in other words:

• The 'risk assessment' of models will be standardized

• The minimum line for jailbreak countermeasures will be set

• The criteria for users to judge 'this model is safe' will be unified

An AI expert on note wrote, 'This is the same as the history of existing technology: free competition -> market chaos -> industry rule-making.'

In short, the 'era of doing whatever you want' is definitely over.

May and June were the era of 'wild AI.' Creators made what they wanted, and users tried what they wanted. Some even made a profit from it.

July is different. The government has started to intervene. Industry standards have started to form.

In that process, companies with 'high safety measure costs' will drop out. Only 'companies confident in their safety' will survive.

Anthropic likely anticipated this and invested money in safety measures.

Cost optimization is becoming serious

A prominent topic in July's notes is 'optimization of monthly usage fees.'

Japanese companies are writing up estimates like 'How much will the monthly cost be if we introduce AI tools to 1,000 employees?' and expressing concerns about 'how many tools we should actually contract for'.

It was also in July that GitHub Copilot switched to token-based billing. In other words, we are now in an era of 'pay for what you use'.

Corporate accounting departments are starting to get angry, asking, 'Are these AI-related SaaS costs really this high every month?' Pressure to cut costs is beginning to mount.

Someone in an IT department on note wrote that they 'evaluated the monthly costs and narrowed down multiple tools to just one'.

In short, the phase of 'trying this and that' is over.

A realistic selection process of 'using this much' and 'so, how much is the monthly fee' has begun.

AI companies with low profit margins will disappear at this stage.

AI is becoming a 'utility'.

Until June, the atmosphere that 'AI is revolutionary' still remained.

July is different.

A sense of reality that 'AI is just like electricity' started to increase on note.

Electricity was also 'revolutionary and rare' at first. Now it's different. It's a habitual infrastructure of 'use it, pay a monthly fee, and be in trouble if it cuts out'.

AI is heading in that direction too.

It was symbolic that someone on note wrote about the 'fear of depending on AI'.

In other words, the stage of 'whether to use it or not' is over. We have entered the stage of 'how to keep using it'.

To keep using it, you need reliable vendors, stable costs, and secure operations.

Funds and labor are concentrating on models and companies that have these things.

That was the vibe in mid-July.

It's 2:00 AM. On Abema, the fourth episode of the anime has finished. The next episode is still playing.

Looking at note, AI is no longer 'wild'. It has reached a stage where companies can choose.

The fact that the choosers are starting to get tired is also a sign of that.