At the least, it deters people from spinning up Astra Pro requests via the ChatGPT interface willy-nilly.

Those were 2 flagship projects that will now make ChatGPT a lot less useful to people who pay.

We're in trouble. OAI wanted to raise $30B and raise $11B they are cutting costs.

I don't use Claude chat, but this is a great way to ensure my usage doesn't disappear on me.

I had it for a couple weeks before it disappeared. It probably decreased upgrades and monetisation too much.

The friction is intentional. Esp with the broader ‘AI trade’ showing jitters; they need good financial numbers.

I wonder what impact this will have when a sizeable chunk of developers are no longer having convos and providing feedback? Could make a noticeable difference.

I also recall reading the following: https://newsletter.semianalysis.com/p/anthropic-subscription...

I will see once this transition actually happens but if another provider offers more for less and a cost of switching is inherently low as most of project lives on the disk and not in cloud, then why not move? For my workflow pro astra has been very helpful, likely will be less so if it becomes far too expensive.

If you want to plan out features then use agents to implement them in small chunks, deepseek has been comparable to Sonnet 5.5.

But this is just early stages for me, I was fully using Claude models for a while and decided to switch after getting annoyed buying extra credits for personal projects.. I was like “eh I’ll just throw a few bucks in openrouter and see how it is this time” and it’s been so much cheaper than buying Claude api credits when my subscription is out

It uses tools in my custom harness about as well as Claude models and makes edits to files doing specific things about as well as I’d want.

For small evaluation ds is impressive.

I'm pleasantly surprised by grok models.

We use cursor at work and i'm at the $20/month plan.

I used to mostly use Gpt, composer and occasionally opus. But this month i maxed these models quota and I've been working with Grok 4.7 essentially non stop for the last two weeks.

Grok is fairly verbose and somewhat slower but in the end it does get the job done. I've done a surprising amount of work (planning, implementing, tool usage, analysis) on the base plan.

Well, if so I'll say something too – I would feel better giving money to a company led by a man who has stated repeatedly why he wants "maximum truth-seeking AI".

Ironically enough I'm currently on a ChatGPT Pro 200 package because OpenAI has been my platform for all my serious (non-coder-guided) coding work so far.

But with this new change I might actually finally jump ship. And maybe it would be to SpaceXAI and their Grok Ultra package. That's still $200/month for me, but I would get 50x more usage compared to plus, compared to "just" 20x with ChatGPT Pro 200. It looks tempting.

Musk's main concept summarised by Gemini: "Truth as Safety: Musk argues that the most important aspect of AI safety is ensuring the system is maximally curious and honest, even if that truth is unpopular.

Raising vs. Programming: He compares developing superintelligence to raising a child rather than strictly coding it, warning that forcing an AI to hold contradictory or untruthful principles will break its logic and make it 'insane'."

Elon Musk has likely done irreparable harm to my country. Harm that undoes generations of progress and sacrifice by people I've loved dearly.

I do wish that more people cared and it's supremely disappointing to me that they don't.

Everyone claims to be interested in the truth. Ideologically driven political agitators are not particularly credible on that point though.

Right, fair enough since Musk and the sort of people he supports and associates with are widely know for not being the biggest liars and hypocrites in generations.

Grok had specific baked in instructions to avoid negative statements about Musk and Trump, how is that for truth seeking? No other LLM company has done that from what I know.

https://www.rollingstone.com/culture/culture-features/elon-m...

DeepSeek’s Winnie-the-Pooh Problem (Feb 2025) - https://medium.com/humanai/deepseeks-winnie-the-pooh-problem...

I know Musk tried to subvert it before, but it doesn't seem to have worked much, or at least not much these days.

Yes, it's also easy to use real facts and get Gemini to criticise Google, GPT models to criticise OpenAI, and Claude models to criticise Anthropic.

But the point is that people think Grok is somehow super loyal to Musk, and it's simply not. Maybe it's because it's "maximum truth-seeking", I have no idea.

https://www.theregister.com/offbeat/2022/11/14/twitter-engin...

If you simply go by what he is saying, then you should also simply go by what anyone else is saying during the time Musk isn't talking. When he says he's Jesus, he's Jesus until someone else says he's not. That's as much as you can be trusted with.

OP is not even demanding others to adopt it, so I don't see why it rubs people like this.

I’d naively expect openrouter to be a thing that ships with its own agent/harness and lets you pick the model that requests are sent to. But the website implies that’s not what’s happening.

How does the merge impact you then?

OpenAI is committed to helping us find the proper valuation of our workflows and we’re committed to paying far less than the perceived value these tools bring. Reminds me of the old Oracle pricing strategy of “how much you got?”

I say this working in an LLM startup where we serve literally unlimited usage (with the only caveat being that usage is serialized) via subscription and, unlike OpenAI, are actually profitable. I guess when you're hundreds of billions in the hole you're desperate to try anything to save the sinking ship, even if it clearly won't work.

We use Chat precisely because of that!

It's not unreasonable for them to switch but you really should not be so dismissive, this will likely have a massive effect on many.

This is a way of misusing the system and taking away GPU from people who are paying for metered work and the quotas became unsustainable.

I'm using it exactly the way it was designed to be used; that's why they allowed access via Chat. They've just decided merging is more economical.

Others building harnesses around chat or simply falling back to manual copy/paste mode once hitting limits. Ironically AI makes it easy to automate the tiniest loophole to make it a problem as soon as word gets out.

It's a non-issue for everyone not trying to tokenmaxx and sticking to the spirit of subscriptions, but theoretically these folks SHOULD be on a higher tier that reflects their actual usage.

Edit: I feel like I'm coming off as a corpo sycophant defending OpenAI here... not my intention. Both big AI companies encourage this behaviour by creating complex and inconsistent rules around their plans such that it is impossible to know what a subscription is buying, a large chunk of the value coming from discretionary but quite predictably regular resets. As much as they have every right to change their plans... they got what they asked for with all the folks exploiting the loopholes. I just wonder if we'll ever get any sort of consumer law around these "mystery box" subscriptions with meaningless tiers so everyone can stop wasting energy on finding, exploiting and then fixing pricing loopholes.

As for the 'mystery nature' of subscriptions; you can always pay API rates if you would like. If their API rates are too expensive, very cheap models like DeepSeek V4.1 Flash, etc, are available.

Just because API rates are ridiculously expensive doesn't mean we can't complain about how convoluted subscriptions are.

Offering an alternative, overpriced product in order to anchor the other one doesn't make OpenAI immune to criticism. As it stands it is impossible to make an informed choice about which subscription from which provider to buy or whether API price vs subscription is for you because there is no indication at all what amount of usage you get in a subscription, and it's highly likely to be different for different people on the same plans.

Imagine if Netflix offered subscriptions to watch a certain number of hours of shows. But they didn't tell you how many hours you would get, and each show deducted your hours at a different rate (which also was not disclosed and constantly changing). People would be mad.

So happy to have the open models!

> The reduction announced in our September 29 email is canceled. If you already had the higher Work and Codex allowance, you keep it. If you had the lower allowance, it has doubled to match. Your subscription price stays at $200/month.

Retail cost of $2500, would love to know their actual cost for these credits.

* 85% inference margins = $375

* 50% abandonment (since it does expire; and general prepaid credit abandonment is pretty high; especially given how much credits they dished out; so probably on the low end) = $188

Plus, it's unclear how many users opt out of training, so more usage and RLHF data on their best models from long-term power users, might work out to be a net benefit to them; since these kind of "non-abusable" (i.e. not announced beforehand) announcements don't suppress monetization as much.

https://x.com/JesusIsGod/status/2116228727059411248 am I misinterpreting? Or was it always 10x? Everything is so vague, no wonder they call it "vibe-coding" :)

edit: A lot of people on twitter say it has always been 10x for weekly so what do I know

I've been using Chat for all my Codex/Work tasks precisely because of the fact it didn't affect usage. Slightly less convenient and smooth but still directly against GitHub repos and other connectors so fantastic value in my Plus plan.

Another reason after the recent Pro 200/500 shenanigans to reevaluate my sub.

All good things and token-burning workarounds come to an end...

And then giving a bunch of bullshit reasons that make no sense.

I think everyone should start thinking about how these AI companies that have absorbed huge amounts of money are becoming more and more evil, and develop the ability to say "We have choices."

I think it's time to make these AI companies return the resources they have occupied, and seriously reflect on themselves.

I think we should make these companies' stock prices crash, demand a pause on unreasonable AI construction, and cancel subscriptions to those AI services that are expensive as hell and control your entire system.

Embrace open source, LLMs that you can privately deploy yourself, and keep your data in your own hands.

I think China is really doing well right now, open sourcing useful models, not letting hardware prices rise crazily, and even pushing down Token prices.

As for security issues, since they're already open source, we can do Fine Tune afterward to solve these problems.

And 4, someone has done experiments showing that 4 small models can reach the level of claude fable.

I really hate those arrogant AI giant companies. They act so calm and unbothered, wanting to own everything and control everything. It's disgusting.

I'm really looking forward to China developing a cheap, non-CUDA AI machine, keeping the price around 1000 USD, a machine that can run deepseek flash level models, becoming a dedicated AI machine rather than a training machine.

Then there will no longer be Token anxiety, no need to react to all kinds of decisions made by those AI companies, and no problem of giving your data to others.

It's time, brothers.

Other people also have that choice and many choose differently from you, but no one’s taken your choice away.

IMHO, they will be seeing a slow down in growth (not total numbers) suggesting the AI bubble is about to pop, against the backdrop of their investors now oversized cash budget that is still growing.

They're souring the milk, adding the anti features now they have the captive audience.

With this change even non-coders will jump ship. Not sure why OpenAI is doing this now. They should improve their models now to match the competition.

The models may switch occasionally in terms of who the class leader is in certain disciplines, but at the top level either provider is enough for professional coders.

Saying that either is only suitable for non-coders sounds gatekeepery and very out of touch.

Might affect go and Plus users.

Their 5 hour limit works better with a "constant background trickle" usage pattern. I don't think the $20 account is a good fit for hobbyists-who-have-day-jobs (which is where I think it should be targeted). They should have an offering that meets "binge" needs.

Once this change comes into effect and I'm not able to use the chat side of things for more casual inquiries and ideation it'll be the motivation I need to try out Claude or Grok again. I'd rather have a second subscription with a different provider than double up on one for more of the same.

At OpenAI/Anthropic scale, they're can't just scaling up-and-down on demand on rented autoscale GPUs (to be honest; that's not really practical anymore for even medium size volume); the peak hours are basically the same for everyone.

Just look at this announcement: "your usage will be measured in tokens, which we do not define, nor give you a way to measure them. The amount of tokens you consume will depend on things you can't control". What is left unsaid is that your subscription will contain an unknown number of those tokens you can't measure anyway, and that they will change the number whenever they feel like it.

You pay for… what, exactly?

I have no idea how this kind of sale is legal, especially in jurisdictions with consumer rights protection (EU).

- https://developers.openai.com/api/docs/pricing

- https://platform.openai.com/tokenizer

Edit: It's some info, but not much around caching, or what context does vs just what you type in the chat, how reasoning or tool use affects cost/token burn, what control you might have on output tokens, etc.

My point is that in normal world, when you buy something, it is verifiable. You buy vegetables by weight and you can weight them yourself. You buy gas at the pump by the liter, and you could measure it yourself. You buy electricity by kWh and you can measure it.

With AI subscriptions, you get "some amount" of "something", neither of which is defined, and neither of which you can measure.

Also, you can't always measure those tokens independently, and the definition of a token changes over time (see the rage with one of the Opus updates).

What's your guaranteed down and upload speed? The ISP's usually guarantee far lower speeds than they advertise.

Token subscriptions are not really a completely novel concept.

Hello,

In the coming weeks, we’ll start testing a combined experience for Chat and Work tabs in ChatGPT with a small group of users. We’ll use their feedback to improve the experience before bringing it to more people.

This will enable you to move from exploring an idea to having ChatGPT carry it through, all in one place. You can bring in your files and plugins and ask ChatGPT to build a website, create a presentation, or do things for you on the web, while reviewing and guiding its work. Codex will remain a separate experience in the ChatGPT desktop app.

*A shared allowance for ChatGPT and Codex*

As part of this unified experience, your use of ChatGPT and Codex will draw from the same allowance in the coming weeks. We’re preparing carefully for this transition to help prevent disruption to your workflows. We expect most users won’t need to wait for their usage to reset more often than they do today. Usage will depend on model and task, with more advanced models and complex tasks using more of your allowance. Unlimited Instant is included in your plan and will not draw from your usage.

You will be able to see your combined remaining usage and reset times in Settings → Usage. If you use your full allowance, you can wait for it to reset, upgrade your plan, or buy credits to continue using eligible features.

The OpenAI team

With the implied, "...but some will"

Note that they also started tracking tokens for normal chats in one's ChatGPT profile in the last couple weeks. Scroll to "Token activity" on: https://chatgpt.com/profile

likewise, suppose you buy a 1gb plan from a mobile isp, then you find a youtube tutorial that you're interested in. the nugget of information you need is 23 minutes in, but you have no way of knowing that upfront, it may not even be there. you can exert some control over how many bytes you send in the headers of the get requests to youtube, but you have little control over how youtube compresses the video you requested. also you download 2 30s ads and a sponsored segment before finding what you were looking for. unfortunately you're now out of data on your plan. do you blame your isp for their shady sales practices? is it youtube's fault? what did you pay for?

You're billed by the hour / mile, you can choose an economy vehicle type, but the Toyota driver kept getting lost, swapping out vehicles, changing the price, when you finally reach your destination, you need another top up.

For example, on Android 17, there is Data Saver mode. Your per-app settings can control mobile data usage, in the background, or unlimited. Your WiFi networks can be individually labeled as metered, non-metered, or automatic. In the YouTube app, there are settings for your video quality, and how to adjust that quality based on your connection, if it is mobile data, or metered WiFi. YouTube Music has its own settings, and since it handles podcasts, you can opt out of video entirely.

Furthermore, there is monitoring of data usage. You can set the time period: for example, my mobile data metering resets on the 9th of the month. You can see usage for each app during that time period, split out between mobile and non-mobile data. I can set a warning threshold and a hard limit for mobile data usage.

I only use free LLMs, and not very much. I've only reached a limit once. But I do see that they have indicators or meters for usage. So perhaps check again, look harder, maybe ask your LLM how you can monitor and restrict token usage.

It's not how you are using them. They're setting up complex agent->subagent webs that exponentially consume tokens, and "vibe" coding things via 5-10-100 agents just scaffolding whatever. There's no amount of human information that would stop that.

The problem with tokens isn't that they can't track it, it's that the labs can seemingly change it's definition anywhere in the stack regardless of what it technically means to the LLM you are using.

It should be megatokens.

They have an open source tokenizer on GitHub, they indicate which tokenizer they use per model, and they have a simple to use website that tokenize text and visualize it.

Most consumer protection laws can be easily satisfied by offering a full refund.

I have no idea if that's the same definition of "token" that OpenAI's billing department uses.

But, some people do yearly subs. In that case you're handing over money without even knowing what you're getting.

Hoping not

A Dot consuming credits passively with Astra just to deliver some low value “email says a package will be delivered” message.

Scheduled tasks have more control and focus.

Personally, I didn't have any reason to look elsewhere because a couple of Pro accounts ensured I always had access without worrying about limits (the reason I never kept an Anthropic account active for more than a month).

I'm sure we'll quickly start seeing how the frontier models aren't all that necessary.

Combine this with the fact that you can't trust OpenAI with your data, I think they are blowing up the tiny moat they might have had.

It's also worth mentioning that their Dot product got nerfed in under two weeks. I went from rethinking my workflows to dropping the product completely overnight between the data loss and network firewall "issues".

It's too bad because I've been fine with using just ChatGPT since the beginning, simply because they were most generous with limits.

All good things come to an end I guess.

The truly good news is they aren't reducing the $200 plan usage. I was going to downgrade my plan after this month, now I won't.

I consider myself a Heavy cgpt user since 2022. To the extent where I’ve made over 30k ish in solo dev side-projects since 2024 that wouldn’t have been possible without hiring a team.

I have not once gotten a over limit notice from openai.

Alternatively, I used to max out my claude tokens very quickly even on fairly small projects.

The only difference I can tell from how I see other people utilizing these tools is that I never once “fire and forget.”

My project flow is: 1000 rounds of text back-and-forth on architecture, data structures, competitive analysis, tool use and calling, etc… and then the last 10% is code writing test development and then usually about five rounds of CI/CD iteration until I get to a MVP or something shippable.

Are the majority of people attempting to one shot and then when they fail they just keep trying to one shot?

In my experience, you can get 90% of the way to the technical solution before generating any code whatsoever.

That’s been true before LLM’s continues to be true now.

Why do you need to get more done at once?

It sounds like you’re describing a one shot but with more steps

There are plenty of coding agents, but I thought ChatGPT is the only one frontier model for any conversations..