Yeah, that makes sense. It’s still questionable what better means in that context. It will for sure be more intelligent in a way but I just think the newer models don’t behave as nicely as 5.5 did in a agentic setting. What I mean is that 5.5 just found the best middle ground between doing things independently while still not doing more than I asked for.
I have the exact same feeling. For me the last big leap in models was when the Codex-Family came out (GPT-5.1-Codex to GPT-5.3-Codex) and then the GPT-5.4/GPT-5.5 that brought the "Codex" capabilities back to the main stream. The GPT-5.6 family feels very different, might be a little bit more intelligent but behaving very differently, and GPT-6-Astra is along those lines. That's why my main driver still is GPT-5.5.
Freelance full-stack developer and AI engineer with 15+ years building web applications. My main lane is Laravel/Vue business software: SaaS products, internal tools, API-heavy apps, legacy PHP modernization, and projects that need an experienced person to stabilize them and move them forward.
I also build AI features into production software: OpenAI/Anthropic/Google API integrations, agent-style workflows, scraping/data pipelines, queues, and content/publishing systems.
Founder/operator background from SaaS, ecommerce, SEO/content, and marketplace projects, so I can help with product scope and business tradeoffs as well as implementation. Best fit: remote freelance or contract work, especially 20+ hours/week or focused rescue/build projects.
It definitely feels like subscription limits have dropped since GPT-5.6 came out. But I couldn't really find any evidence for it, going through my history. All I could come up with is that 5.6-sol uses about twice the tokens compared to 5.5 (both on xhigh) [0].
One thing I just did was to stop four long-running Codex sessions that ran on 5.6-sol and switched to 5.5 and asked it "Please check if you're really going against the actual goal or have you drifted away from that?" and all four replied something like "Yes: I had started to drift"
I always set thinking to high for the main session or subagents doing audits, medium for everything else.
I think xhigh or max has a place if I'm doing something genuinely complex. But in general use it's slow and I have the impression the highest often gives worse results - over-thought, over-engineered. High or medium might give better results for less complex tasks.
I have seen that kind of drift in smaller models (e.g. DeepSeek V4 Flash) when setting the thinking too high. So less thinking would lead to better results. But that's not something I'd expect from a SOTA model. Higher thinking effort should lead to same or better results.
Freelance full-stack developer and AI engineer with 15+ years building web applications. My main lane is Laravel/Vue business software: SaaS products, internal tools, API-heavy apps, legacy PHP modernization, and projects that need an experienced person to stabilize them and move them forward.
I also build AI features into production software: OpenAI/Anthropic/Google API integrations, agent-style workflows, scraping/data pipelines, queues, and content/publishing systems.
Founder/operator background from SaaS, ecommerce, SEO/content, and marketplace projects, so I can help with product scope and business tradeoffs as well as implementation. Best fit: remote freelance or contract work, especially 20+ hours/week or focused rescue/build projects.
Freelance full-stack developer and AI engineer with 15+ years building web applications. My main lane is Laravel/Vue business software: SaaS products, internal tools, API-heavy apps, legacy PHP modernization, and projects that need an experienced person to stabilize them and move them forward.
I also build AI features into production software: OpenAI/Anthropic/Google API integrations, agent-style workflows, scraping/data pipelines, queues, and content/publishing systems.
Founder/operator background from SaaS, ecommerce, SEO/content, and marketplace projects, so I can help with product scope and business tradeoffs as well as implementation. Best fit: remote freelance or contract work, especially 20+ hours/week or focused rescue/build projects.
Anthropic, OpenAI and SpaceX all want to IPO within this year. There's just not enough money in the market to buy all those shares. So people might sell their shares in other companies to buy in at the IPO, then when the next one goes public they might sell the shares they just bought to jump onto the next one and so on. I don't think that there was a situation like this ever before.
The amount of new shares for sale could be very large in those three IPOs:
SpaceX: up to $75B [1]
OpenAI: at least $60B [2]
Anthropic: more than $60B [3]
Together, that would be about $195B+ of IPO shares to buy.
For comparison, all U.S. IPOs together raised $44.0B in 2025 [4].
All IPOs in the world together raised $171.8B in 2025 [5].
So where should the money come from? Either from selling shares in other companies or from loaning money which would only make sense if the Fed brings back ZIRP.
They might be conflating the valuation of these three companies with one of the recent analyses that the presumed valuation of the current “AI Economy” is an order of magnitude over the valuation of the total wealth of the world.
Another scenario would be that they release very little shares but because of their market cap they make up a significant portion of an index so index funds are forced to buy a lot of their shares which would drive prices to insane heights.
Well it will always be good in a way, but probably won't become better in the near future. Opus 4.7 was a downgrade in a way that gives Anthropic more control and better margins. And they keep Mythos away from the normies, giving access only to large corporations who pay millions for it.
I'd say it's a mixed bag. Yes, price increases are/were expected. But blocking 3rd party harnesses from their subscription and also moving SDK/claude-p access out of the subscription is blocking innovation and therefore future use of Claude models. What I mean is, while Claude models are SOTA, Claude Code is not. It's full of bugs and shortcomings and new innovative harnesses might make better use of the model, but this won't happen now as they are blocked. Same for all the people building their own scripts/workflows around claude -p or the SDK, they will now stop inventing new stuff on top of Claude.
I think it also doesn’t help that they’re all different flavors of the same tool and it’s very easy to jump between them. None of them actually made a particularly distinct product, and they want us to use their tool to make/justify the nebulous billion dollar product.
As for subscription/token costs, even with increases they’re not even remotely covering costs. If people actually paid what it cost for these companies to even break even, nobody would be using these tools. They simply aren’t that consistently useful despite all the grand claims. They can be useful and in some areas they are very useful, but nobody is going to spend thousands of dollars a month to have something rewrite their emails regularly. And it’s not like these companies are trying to target one industry. They want to target everyone.
I wonder if there's a bit of a standoff situation here where every company feels like they have to burn enormous amounts of cash training/iterating on their models because if they stop, they'll get leapfrogged too quickly. So unless they all agree to just settle where they are more or less and focus on building out functionality and tooling, they all have to keep spending at an unsustainable rate.
reply