Well that was one of the reasons why i build my own agent/harness - to have actual full observability of what it is doing and why it is doing what it is doing.
Don't really understand why other agents to give those insights but well found my solution.
1) What kind of guardrails are you using? Do you do models like prompt injection checkers? Or what guardrails are useful that do not know the environment the agent lives in?
2) I heard Bifrost also works for this. Did you try that as well, pros/cons?
I'm not the person you asked but yes pretty much. I've found the prompt injection protection models pretty limited, but get good results from gpt-oss-safeguard with a policy describing various prompt injection techniques. More expensive but more flexible and reliable and I only run it on web fetch and web search results. Costs me like 5-20 cents a day depending on what I'm working on.
Still working on deploying it and setting it up with cloudflare sandboxes so I don't have deep expertise and need to experiment first before I can give advice
OpenRouter does all this for you. It's pretty nice. Although for apps I use another server to lock down the requests even more and prevent the client from seeing my API key.
LLMs work over a JSON API. You can read the JSON to see what's going on.
Using other people's harnesses I always got frustrated by not being able to see what the model sees. What are we actually sending it?
Context is pretty important when it comes to LLMs. Arguably, context is everything!
Usually the most straightforward way to see it, was to intercept the HTTP requests and dump them to a file.
This is usually trivial, just change the OPENAI_API_URL (or equivalent) to http://localhost:8080/ or whatever, which can be running a 20 line Python script, that just forwards the HTTP requests, and dumps them to a file.
It's pretty straightforward, although it really shouldn't be necessary in the first place.
Yeah you should run an LLM proxy, but imagine a world where LiteLLM and all the other LLM shovels weren't buggy vibecoded messes. And with the velocity of updates I feel there has to be vulnerabilities there. Or am I just holding it wrong? Anyone got a good alternative?
This general phenomenon is not new, but it has been so obvious during 2026. Everyone want to consume everyone else hard work, to create their own slop. Just pure egoism.
Hi, I'm sorry you feel this way about LiteLLM. We're actively working on improving code quality / reducing bugs. Were there any specific pain points you faced with LiteLLM?
If you're open to a convo, I'd love to hear what specifically about the product was not great and improve it for you :)
I'd love to do something like this, or use API keys, but I still don't understand how people justify the cost of using API keys vs. subscription.
I tested it once, I'm in on the 20x claude max ($200/month) and I either max my weekly limit every week or come really close. But it only took 2-3 days to hit $200 spend with api keys.
The current crop of flash models are more comparable to sonnet than opus but they are very affordable.
The reality is that the subscriptions are heavily subsidized. Pretty much every sub that's not run by one of the money-burning frontier labs has had to drop it or dramatically increase prices and lower limits in the last few months. Cheap tokens are an acquisition tool, not a real product. It sounds dramatic but it really is the drug dealer sales funnel.
At some point they're going to drop the real prices in your lab and you're going to have to figure out how to pay that much or pay someone else a lot less. Better to get a handle on it now.
>At some point they're going to drop the real prices in your lab and you're going to have to figure out how to pay that much or pay someone else a lot less. Better to get a handle on it now.
So just wait until that happens? What's the point of scrimping and saving now? Not to mention you're paying for a premium in the meantime.
It's not obvious the subsidization will end. Loss leaders are a thing. Most enterprise at this point are paying API pricing
To better predict what they'll do with it we'd need to know the proportion of token usage via API vs subscription. Keeping an affordable, slightly unprofitable subscription to get people in isn't a bad idea. They also get marketing benefits from the word of mouth
Also, some subset of developers on subscriptions are going to develop products that use the API, and some of those products are going to be successful.
In the optimistic case you describe, what smart customers want to watch out for is being lured into vibing your projects into such byzantine slop-complexity that subscription plans and trailing models are no longer sufficient to make progress on them.
Vibe-coders talk about surrendering understanding to the models, trusting that there will always be a model able to overcome and work through any technical debt they happen to accumulate, but the danger that still lurks there (at least for the naive) is that you'll need increasingly large contexts, increasingly sophisticated models, and increasingly large token spends just to keep that ball rolling.
Between profitable enterprise accounts being ineligible for subscription, eager but busy people paying for underutilized subscriptions, and naive vibe-coders falling into their pit trap, you can indeed imagine subsidized subscription accounts being available for a long time and remaining an effective value hack for people careful enough not to fall into their pit tral. But only insofar as they are that careful.
Also they can always find another way to make money. Make subscription users product and sell some analytics like "show most trending project being built this week".
You cannot predict the future, but you can identify risks and attribute a likelihood. A risk is that the subsidized inference will either end or be reduced significantly in the next 1-2y. If you think the likelihood is high enough it makes sense to ensure your systems are prepared for it.
> It's not obvious the subsidization will end. Loss leaders are a thing.
Loss leaders are a thing so long as they actually do lead somewhere. If your primary value proposition is conveyed by the loss leader itself, where do you go from there?
(meta commentary)
really gnarly word choice: "drop" in "they're going to drop the real prices in your lab"?
I took your meaning as "they're gonna drop [a bad surprise on you by raising] the real prices"... so at least it's not quite as tortured as the ambiguous "support for version N dropped today" (where you often can't even infer the direction from context).
this isn't just a tangent, I'm having trouble understanding the point you're trying to make, which hinges entirely on deciding if you meant "drop" literally
They meant "they're going to drop a bad surprise by charging you API prices", so yes, a better wording would be "they're going to give you a bad surprise in the form of forced API pricing".
The term "drop" becoming a synonym for "release", "put out" or "put into effect" for things outside of music/albums is one of the most confusing linguistic innovations of the last decade.
> At some point they're going to drop the real prices in your lab
The real price for tokens at a specific level of intelligence is dropping rapidly. Between models improving and new hardware getting faster, I think prices will be going down, not up.
> Better to get a handle on it now.
I experiment with a lot of providers and models. I strongly disagree that anyone needs to start migrating now in preparation for prices maybe going up in the future.
It doesn’t make sense. Deal with a price increase when it happens, if it does.
Switching models and providers isn’t some huge lift. I can jump to a new model or even hardness and get a feel for it in a day or two. It’s not some huge undertaking that requires preparation and planning.
If prices go up, try something at that point in time. Don’t worry about it until then.
When Claude Code subscriptions really started to catch on, there was a popular idea on HN that Anthropic's end game was that basically every dev adjacent role would need to pay for a $200/mo coding plan.
Then, both Anthropic and OpenAI in lockstep developed a enterprise pricing structure that openly coerced large businesses to pay API prices where it was feasible to spend $200/day rather than the comparatively cheap $200/mo.
Even if token pricing relative to intelligence does trend down over time, that doesn't mean the pricing structure will actually translate to lower costs, particularly in the cash cow B2B market.
Companies still believe they have IP to protect. Even when all they do are basic CRUD systems or on the verge of collapsing if they don’t renew some contracts.
From experience, if you work somewhere with very disconnected higher ups a sales person shows what your token usage is vs spend on the monthly subs for whatever set of users you have and then show an API price with a sample bill for even “more users” at that same number of tokens that is 30% lower. The crux of course being that your existing user base is not going to use fewer toke a just because you signed more people up and have a discount for the status quo.
Then when you consolidate all users to a single provider you find out you have a 2x to 3x larger bill because you… logically have more users at the same average token spend than the quote.
At my previous job, all while managing a full team, I did a lot of porting of legacy apps to modern standards to maintain various obligations. It was AI gen work in between meetings and paperwork. It was seen as such a success that I was asked to move to an IC role to focus on it within the “Elite new products team” that was supposed to be all AI and the latest tech. I showed up to find that they were pushing out low quality manually written code in 12 hour days six days per week to appear like LLMs were in use to save face because they lacked a token budget or the skills to properly manage one under the APIs. They had dismissed someone who had spent multiple thousands per month on Agent swarms. It was the perfect situation to make their 3x budget issue a 10-30x budget spend issue.
For many many reasons I left that job and got a new one but that was the cherry on top.
The API/subscription split as always there. I don’t know why you’re talking about it as a bait-and-switch development.
That doesn’t change anything about what I said, though. The relative price per task or price per unit intelligence is also going down in the APIs too.
> Even if token pricing relative to intelligence does trend down over time, that doesn't mean the pricing structure will actually translate to lower costs, particularly in the cash cow B2B market.
What are you talking about? The APIs serve the same models as the subscriptions. Unless you think the price of APIs has been going up (it has not) this argument doesn’t make sense
> At some point they're going to drop the real prices in your lab and you're going to have to figure out how to pay that much or pay someone else a lot less.
This only applies under the assumption that everyone tokenmaxxes their subscription.
I have visibility across a few different Claude team subscriptions. I'd say roughly 10% of people max their subscriptions.
I think I've only seen one organization where the under-used subs don't more than pay for the tokenmaxxed ones.
And that is ignoring the 90% margin that API inference runs at.
At some point companies will need to have the “scary” conversation about the value proposition of adding another human to the team, or adding a comparable amount of token spend to the team.
$1k/person sounds like a lot, but a ten person team at $1k token burn / head / month is spending $120k. How much does a new hire cost, tco? Does velocity from one more person compare with velocity from one more person’s worth of tokens?
Sure, learning to be efficient matters. So does taking the time to consider cost vs alternative. You might be surprised by what you find, in either direction.
Counter point - models get better, ways to run them gets optimized and hardware will eventually come down in price.
The share of tasks where "cheaper" models are enough also increases for same reason. For majority of use cases it will quickly become race to the bottom. And that means part of the pie for the "premium" models of frontier AI labs will quickly become smaller and smaller.
The drug dealer gives you a free hit because it is profitable for him in the long term for you to be addicted to his product, not because he is your friend.
...yes? It doesn't literally kill you if use more tokens than usual?
And we're talking about subscription plans from OpenAI/Anthropic here. Switching to a different model and/or harness if the price changes isn't a "the world has run out of opioids and I'm going to need a detox center" level crisis.
The original question was "are you sure an OpenAI subscription isn't heroin?" Different impacts of overuse (death vs slop code) feels like a valid point of difference.
This is a weirdly melodramatic analogy.
If you want to flatten everything to "all things habit forming are definitionally equivalent", you'll get some pretty funky results.
You're equivocating, they were talking about addiction, you're talking about overuse/overdose which is where the analogy breaks down for pretty obvious reasons.
If you're genuinely curious, I'd recommend searching around about the history of this persistent urban legend. It's really interesting.
But the short version is that there are lots of people who want to buy any given drug and drug dealers sell them. I don't know exactly what you mean by "WAIT for people to demand the more profitable drugs", but there typically isn't a whole lot of waiting. There's a lot of demand out there. Giving away a bunch of free fentanyl to random people in the hopes that one of them instantly gets addicted would be a pretty wasteful and high risk sales tactic, especially given the number of people that want to buy it already.
Personally at home I use the free github copilot and then openrouter when that runs out. Last month I used $0.90 of openrouter. Zero so far this month. When my bill starts hitting double digits I'll consider options.
Different matter at work but even there we are paying for API costs and have a monthly allowance (starts at $200/month but increases fairly easy to get)
You might not have a choice. For example, if you want to get a guarantee from Anthropic that they don't train bases on your prompts, you will need an enterprise deal, which is pay-by-token.
I'm with you there. I'm on expensive plans with everyone because I run out of tokens and rotate/use the models that are best in their niche...
CLI Model Pct Remaining Reset Date
claude all models 15% 3d
claude Fable 50% 3d
grok all models 0% 4d
agy Gemini 36% 5d
agy Claude and GPT models 0% 5d
codex codex 0% 6d
And I also just used two of my three get-another-week-free tokens for codex.
If you've got an agentic swarm, or are working on multiple big projects at once, or have agents parsing data... you burn through tokens fast. The last four days in API would have cost me well over $3k.
I'm always curious, what kind of software exactly are people building that requires multiple parallel agents and thousands of dollars per month in AI tokens per developer!
I think in many cases the token bloat comes from using a fleet of autonomous agents that "collaborate" over a problem. Using one coherent agent loop with a human involved might not feel as futuristic but it's more effective and efficient (right now).
We have two wildly different cohorts of people. Those who see these tools as a technological curiosity and those who see them as a means to a specific end. I know that if I finish a task with a certain set of constraints that I will get paid a certain amount of money. I'd much rather be HITL on one very good agent loop if there is real money on the table. If you are mostly having fun, being curious, or exploring (which is fine!) then the cost of subscriptions might feel safer. Running an IRL version of factorio can get very expensive.
I also think the API tokens perform better but I have zero evidence other than observing the quality of my work product relative to that of my peers. I have almost a religious fervor about using prepaid tokens now. I am still using GPT5.5 in my copilot with custom model configuration for most heavy lifting in visual studio. It is more than sufficient. My average monthly token cost is <$100. Paying for a subscription would be more expensive because I have a predictable and controllable amount of demand.
This is analogous to variable vs fixed rate electricity. Paying the real time wholesale price is a massive economic advantage if you can figure out how to control the demand side of the house.
> This is analogous to variable vs fixed rate electricity. Paying the real time wholesale price is a massive economic advantage if you can figure out how to control the demand side of the house.
rrr_oh_man is claiming that wholesale fixed rate electricity is generally cheaper than varible rate domestic electricity. Or that the wholesale-bit is higher order than the variable rate-bit. Not that it's a big issue, bob1029's point doesn't actually care about one situation being wholesale. It's just a somewhat narrow, (in this case) inconsequential, but also valid, nitpick.
I often read opinions where people say their particular method of more hands-on working with AI is the most efficient and reasonable way.
From 500 hours working on my mobile app project outside of work, I can confidently say that using anywhere between 2 to 10 conversations to parallelize work is much more effective in my project.
The models being served in the API are the same as via the subscription, there are no shenanigans here, particularly not concerning quality.
And GPT 5.5 was great at the time, I made heavy use of it for my project, but it's long outdated and already being retired.
You also mentioned using it in copilot, which is a third-rate harness that is frustratingly buggy and inefficient.
There's no comparison to Opus 5.5 in Claude Code.
Use whatever works for you, but frankly, I think you should try something new.
Just not finding this to hold true. The difference in work quality for peers isn’t a matter of how many threads they run, but how much attention they give each. Which is a skill.
> I know that if I finish a task with a certain set of constraints that I will get paid a certain amount of money.
So not salary… ok. That explains some of it. Getting paid by the task might flavor one’s approach. But also one could reason, more tasks, more pay… somewhat confusing.
> I also think the API tokens perform better but I have zero evidence other than observing the quality of my work product relative to that of my peers. I have almost a religious fervor about using prepaid tokens now
And this appears to be the other half. It looks no less futuristic than the more wireheaded approach, just different in form.
I think businesses are footing the majority of that bill right now due to regulatory requirements, but for individuals that price increase is def coming. Maybe I'm too pessimistic, but enjoy it while it lasts :')
For a lot of large scale day to day work, deepseek / zai models (the latter comes with a subscription) can be very cost effective and very competitive actually.
Another thing of note is that when coding at 'scale' to my surprise over time with memory aggregation across users/sessions these models can be better than the evals would have you believe because they seem to be differently/less (?) RL'ed (aka actually listen to instructions), which results in them being more fastidious about writing memories and benefiting from them, especially over time. These are of course mostly subjective opinions but they were surprising, and not what we were expecting.
Relevant: the proxy we wrote and use:
https://yapp.latchlabs.dev is partially open sourced: still working on releasing more of our memories stuff.
People using API keys heavily, are either making enough money from whatever they're doing to afford Anthropic's high prices, or are using a mix of locally hosted and cloud-based models. E.g. using a cloud model as a planner/orchestrator, with a local model as a subagent that follows the orchestrator's designs/instructions.
I had a subscription for a few months. I found that at the end of every week I would succumb to the sunk cost fallacy and try to squeeze all the usage out of my subscription. I'd rather pay a little extra in LLM cost to not have that issue.
> I still don't understand how people justify the cost of using API keys vs. subscription
Like a big white truck or a 600mm lens, expensive gear should be purchased by an employer. I think most folks talking about API access are getting it through work/self-employment.
Anthropic needs to retune their personal plans. Opus is the Amelia Bedelia of models. Maddening! I shelled out $100 for Fable to drive Opus and finish a personal project, but that's probably my last time loading up on usage credits. I put a second $100 towards API credits so I could dink around with API access and understand conversations like these.
There's another big element that is being completely overlooked.
People who use API keys on their dime are usually running DeepSeek 4.x or GLM 5.x (Flash or otherwise) and similar models.
RE: Subsidies:
What if subscription costs are still high enough to be profitable (I believe they are for OpenAI).
Then maybe we should be saying that API users are paying an order of magnitude more than the cost, at least at the very top end (frontier models). Likely because big cos are willing to pay top dollar for what is the best and unmatched elsewhere.
I am on the 25 dollar plan and I never even come close to hitting the weekly limit. Still manage to ship quite a bit. Not sure how folks hit their Claude limits. I have handle on the codebase though.
We (some friends and I) built it because we wanted to understand how these things worked and also had max accounts. It has since grown into a ton more, but is still in essence a very full featured local proxy.
I’m not sure what people are doing with these subscriptions to feel the need to monitor what’s going on this closely. A $100 OpenAI subscription feels effectively infinite for my usage at home
I've never touched my ISP-provided router. But maybe it's just how it works in my country (Malaysia), we pretty much only have 1 ISP with monopoly, so maybe that helps.
I let them install the free ISP router a year back when they installed fiber in the neighborhood. I assumed I would swap it out the next day, but I haven't had any issues like back in the day. I love shopping for a nice router as much as the next tech nerd, but decent drop-in home networking seems to be pretty commodity now days.
In similar spirit, the biggest motivation for me to set up DeepSeek Flash over Opus 5.5 would be seeing the reasoning traces, which is a mere curiosity.
> We need answer explain LLM reasoning traces. Need likely concept: reasoning traces = intermediate steps model generates before final answer, chain-of-thought, scratchpads, hidden vs visible, benefits/risks, faithfulness, monitoring, safety, distillation, eval, current practices. Need explain clearly. Desired oververbosity 5.
Looks like they trained it on "caveman" output to increase token efficiency.
Cache TTL will get you anyway. For Claude subscriptions, subagents still have a 5m TTL and before I discovered that, their constant waits for ~10 min tests or similar tasks would destroy my 5h usage. And for my API subscription at work, Claude has a 5m TTL for all stuff, and that's nowhere near long enough. You have to be monitoring this unless you want to absolutely torch your company's budget.
Yeah my problem is that I need to toggle it between "really active constantly writing" sessions and "running a bunch of long tests and periodically pinging me about it" ones. The former are best with 5m, the latter with 1h.
Ultimately, Codex doesn't suffer from ANY of this so I'm starting to think about just using it for work subscription use as well.
$100/mo. is basically the cost of a cable TV package. Often $120-170, actually.
If you're the kind of person who gets enjoyment being able to code all sorts of things you were never able to before, or 20x faster than you were able to before, then $100/mo. is extremely reasonable in comparison.
I'd argue it's money much better spent than on cable.
Also, if a hobby of yours runs you $100/mo., that's actually very much on the cheap side for a lot of hobbies.
Also, $20/mo. is for hobbyists using it for 2 hours at a time a few times a week. $100/mo. is for people working on a side project every evening and every full weekend.
Because the details we care about right now already travelling over the network and the industry is adopting OTEL, where you can export and collate kernel things if you need to
With just a 12gb VRAM GPU + 64gb of CPU RAM, and you can actually own your own infrastructure. I can't wait for hardware prices to fall so that becomes even more accessible.
Tailscale makes Aperture available for free, and while it's not something you own, it is in your tailnet, so at least you have a certain amount of control over how things access it.
It gives a nice view of what agents are doing, with lots of configurability and extensions, and the team behind it has been updating pretty steadily.
I happen to be working on an extension to Aperture, so I have my own reason to promote it. But I do think it's a nice easy solution, and I promoted it to people before I was extending it.
I'm building a "bring your own data" layer for agentic work of all sorts, using the webhook method for now.
In the context of Aperture, I watch for agents looking for certain types/sources of data, and instead provide them with an in-tailnet API for accessing that data.
In a personal context, think email, calendar events, music playlists, fitness app records, transactions. I've also found it useful in a professional context, but I think I need to explore that a fair bit more.
It keeps credentials out of agents hands, and provides some relatively standard ways for controlling what agents can see and change.
I need a fair bit of hardening work before I try to get random folk to use it, but I might post something more here in the next week. :)
It is a not proxy. It can clearly show whats going on across all agents. The proxy's are generally inconvenient and yet another point of failure. I think the author got to the resort option quickly
I am boggled how the first step of an llm after you make an orchestrator isn't some kind of proxy that records your personal internet surfing to be used by your personal AI later. Every college should be running in house AI with their own in house search engines and in house any service. Don't go to any college that wants you to use chatgpt or other corporate nonsense. If they can't make local LLMs work, they are not a real college.
I don't do this but I do have a system that imports all my conversations with AI (claude code, chatgpt, claude app, etc.) to a storage system I call Chatacombs: a unified storage of most conversations I've had with AI, as detailed as I can easily get.
For subscriptions, it is easier to collect the JSONL files from ~/.claude/projects/ and similar directories. They are better organized (e.g., subagents) than what the proxy sees.
However, the best value comes from collecting agent sessions across the whole organization. You'll need to scrub credentials and handle security.
What’s the best way to pull (ie archive) and search these across a few harnesses? I remember seeing a project or two trying to standardize them, but can’t find anything that seems to be a good sensible solution
Ask claude to do it end to end, then write a solution that fits your needs from it's experience with you driving it.
I did exactly this - probably the single most valuable piece of software I've "written". It's the backbone of every piece of code coming out of any agent under my control (via MCP toolset).
Mlflow does it pretty nice including session gathering and optional data enrichment (eg what files were touched, what model was used, git remote).
We use these session IDs for commenting on issues and stories touched by AI for context tracing.
I just wrapped around the Mlflow setup to make this simple CLI helper to set it up. Ofcourse the name suggests only Claude but it’s Codex, CC, GHCP and Agy. https://pypi.org/project/claudetracing/
GDPR doesn't cover unrecorded human memory, so you're fine to overhear whatever in a day-to-day setting. If you then do anything that looks like digital processing or cataloguing that information though, then you're making it a GDPR issue.
Unless you are a robot, you'll be fine. GDPR is about digitally recorded information. If you are a robot, it's probably best to delete that information.
If you have dyslexia and I know it, that brings it down to 20% of people. Combined with some other data, this is a pretty good factor to narrow it down to "who you are", right?
Isn't any personal data a PII? You just need a couple of personal o enable t identify individual. And if this is in an org then amount of individuals is quite small, so sometimes it needs just one personal detail.
Is a medical condition not PII? Together with the fact that it's linked to a specific person, their tasks, responsibilities, mannerisms. Even an IP address is PII, not sure how a full chat log isn't, even if it's not text-book PII.
That's not what GDPR is. I'm assuming you don't have the time or energy to research it, so the quick summary for you is that it does actually protect your privacy and gives you easy and fast avenues to acquire the data a company has about you (in a portable format), request deletion of your personal data, and requires that you explicitly consent to any processing of your data (in detail).
If a company doesn't follow these laws, the fines are proportional to the revenue(?) of the company, so the fines can get quite existential if a company fucks up.
In practice, this is not "shoved down" my throat at all. In practice, this means that I can request companies delete all my data, and they respect this, in practice. I can also request a .zip file with all data they have of me, and I will get it within a couple of days max.
It also means that e.g. your family doctor has to have you sign a document if they want to enter your data into a system, or if that system might be accessed by a 3rd party (such as IT support).
This isn't theoretical; this works. I've requested data deletion from various companies and requested data exports from various companies (and it worked without any issues every time).
It also means that, at any company I work at that handles PII, everyone is VERY careful touching anything with PII. The fact that you have to implement data retention periods, limit the amount of data you collect and retain, etc. forces companies to respect your privacy even if they don't want to or it's inconvenient.
If you want "progress at any cost", particularly at the cost of human lives, you're right. We're behind in the ever so important AI race, and we're missing that privacy of people, health of people, happiness, and the future of humanity is entirely irrelevant unless it makes you money, today.
Well, it's not just the AI race, it's just about everything. Who the hell would want to be an entrepreneur and start a small tech business of any sort in the EU when you are going to be taxed and regulated down the throat from every direction. It's great that you're able to obtain a zip file of what information a company has about the things you've searched for on Amazon, all it costs you is all the hundreds of thousands of individuals who never start a business because they don't want to deal with the headache of touching PII with a thousand foot pole and risk getting sued into Oblivion, so in exchange you just get the large soulless corpos that are the only that can afford to hire people to deal with PII and the enthrenchment of Google/Microsoft and Facebook is entrenched even further.
> Who the hell would want to be an entrepreneur and start a small tech business of any sort in the EU
Probably only people who are interested in living in such a place, and those who care about that sort of stuff are maybe less likely to choose total annihilation and exploitation for profit, so you see less multi-billion dollar companies that have to promise they're not evil, I guess?
It's a different culture, that doesn't necessarily make it bad.
If a company can't deal with PII safely, they SHOULDN'T be a company in the first place. If you can't have a zoo without starving the animals and running them over every now and then, maybe it's fine that you can't start a zoo, even if countries exist where you could do that instead and be fine (as long as you hide the bodies).
Technically you can still run Claude-Code through proxies though, even with the subscription.
Right now I run my Claude-Code sessions through a very simple mitmproxy, logging everything that goes in & out. It's very low level, sure, but it works.
I really want to know if someone can point to where it says this is ok in an unambiguous way. Reading the Antropic docs ir is not clear to me that they differentiate personal mitm proxies from proxies used to share the sub. And if they attempt to detect the latter and ban people, how can they properly distinguish the two even?
I've shared a sub with my husband for a long time, and we never ran into trouble.
Then I shared a sub with 3 people in total and we were detected within a few days.
I don't know what metrics they use, because I use my sub on a handful of servers and different IPs. But I'm guessing there's some kind of fingerprinting going on, where they can actually determine that prompts are being sent by different people
238 comments
[ 5.1 ms ] story [ 34.0 ms ] threadDon't really understand why other agents to give those insights but well found my solution.
I can't orchestrate agents in sandboxes without it.
1. It can issue virtual keys, so provider API keys never go near the sandbox
2. I can see everything
3. Guardrails
And much more
Looking forward to it.
2) I heard Bifrost also works for this. Did you try that as well, pros/cons?
I need to use the desktop site checkbox on chrome in android otherwise the UI is mushed together.
Let me know if any issues, happy to help
For clarity, I do like blog posts like these.
Using other people's harnesses I always got frustrated by not being able to see what the model sees. What are we actually sending it?
Context is pretty important when it comes to LLMs. Arguably, context is everything!
Usually the most straightforward way to see it, was to intercept the HTTP requests and dump them to a file.
This is usually trivial, just change the OPENAI_API_URL (or equivalent) to http://localhost:8080/ or whatever, which can be running a 20 line Python script, that just forwards the HTTP requests, and dumps them to a file.
It's pretty straightforward, although it really shouldn't be necessary in the first place.
homelab running liteLLM
If you're open to a convo, I'd love to hear what specifically about the product was not great and improve it for you :)
My [LinkedIn](https://www.linkedin.com/in/reffajnaahsi/) if you're open to a convo.
I tested it once, I'm in on the 20x claude max ($200/month) and I either max my weekly limit every week or come really close. But it only took 2-3 days to hit $200 spend with api keys.
The reality is that the subscriptions are heavily subsidized. Pretty much every sub that's not run by one of the money-burning frontier labs has had to drop it or dramatically increase prices and lower limits in the last few months. Cheap tokens are an acquisition tool, not a real product. It sounds dramatic but it really is the drug dealer sales funnel.
At some point they're going to drop the real prices in your lab and you're going to have to figure out how to pay that much or pay someone else a lot less. Better to get a handle on it now.
So just wait until that happens? What's the point of scrimping and saving now? Not to mention you're paying for a premium in the meantime.
It does make me sharper I think. If you only ever use claude or whatever you don’t even really know exactly how tied to it you are.
Funny enough the drug dealer thing still holds. A deal’s a deal if you’re going to buy it anyway.
To better predict what they'll do with it we'd need to know the proportion of token usage via API vs subscription. Keeping an affordable, slightly unprofitable subscription to get people in isn't a bad idea. They also get marketing benefits from the word of mouth
Vibe-coders talk about surrendering understanding to the models, trusting that there will always be a model able to overcome and work through any technical debt they happen to accumulate, but the danger that still lurks there (at least for the naive) is that you'll need increasingly large contexts, increasingly sophisticated models, and increasingly large token spends just to keep that ball rolling.
Between profitable enterprise accounts being ineligible for subscription, eager but busy people paying for underutilized subscriptions, and naive vibe-coders falling into their pit trap, you can indeed imagine subsidized subscription accounts being available for a long time and remaining an effective value hack for people careful enough not to fall into their pit tral. But only insofar as they are that careful.
That’s risk management 101
Loss leaders are a thing so long as they actually do lead somewhere. If your primary value proposition is conveyed by the loss leader itself, where do you go from there?
I took your meaning as "they're gonna drop [a bad surprise on you by raising] the real prices"... so at least it's not quite as tortured as the ambiguous "support for version N dropped today" (where you often can't even infer the direction from context).
this isn't just a tangent, I'm having trouble understanding the point you're trying to make, which hinges entirely on deciding if you meant "drop" literally
“Drop [an unwelcome surprise] in your lap” is a common English idiom.
The real price for tokens at a specific level of intelligence is dropping rapidly. Between models improving and new hardware getting faster, I think prices will be going down, not up.
> Better to get a handle on it now.
I experiment with a lot of providers and models. I strongly disagree that anyone needs to start migrating now in preparation for prices maybe going up in the future.
It doesn’t make sense. Deal with a price increase when it happens, if it does.
Switching models and providers isn’t some huge lift. I can jump to a new model or even hardness and get a feel for it in a day or two. It’s not some huge undertaking that requires preparation and planning.
If prices go up, try something at that point in time. Don’t worry about it until then.
Then, both Anthropic and OpenAI in lockstep developed a enterprise pricing structure that openly coerced large businesses to pay API prices where it was feasible to spend $200/day rather than the comparatively cheap $200/mo.
Even if token pricing relative to intelligence does trend down over time, that doesn't mean the pricing structure will actually translate to lower costs, particularly in the cash cow B2B market.
Then when you consolidate all users to a single provider you find out you have a 2x to 3x larger bill because you… logically have more users at the same average token spend than the quote.
At my previous job, all while managing a full team, I did a lot of porting of legacy apps to modern standards to maintain various obligations. It was AI gen work in between meetings and paperwork. It was seen as such a success that I was asked to move to an IC role to focus on it within the “Elite new products team” that was supposed to be all AI and the latest tech. I showed up to find that they were pushing out low quality manually written code in 12 hour days six days per week to appear like LLMs were in use to save face because they lacked a token budget or the skills to properly manage one under the APIs. They had dismissed someone who had spent multiple thousands per month on Agent swarms. It was the perfect situation to make their 3x budget issue a 10-30x budget spend issue.
For many many reasons I left that job and got a new one but that was the cherry on top.
That doesn’t change anything about what I said, though. The relative price per task or price per unit intelligence is also going down in the APIs too.
> Even if token pricing relative to intelligence does trend down over time, that doesn't mean the pricing structure will actually translate to lower costs, particularly in the cash cow B2B market.
What are you talking about? The APIs serve the same models as the subscriptions. Unless you think the price of APIs has been going up (it has not) this argument doesn’t make sense
This only applies under the assumption that everyone tokenmaxxes their subscription.
I have visibility across a few different Claude team subscriptions. I'd say roughly 10% of people max their subscriptions.
I think I've only seen one organization where the under-used subs don't more than pay for the tokenmaxxed ones.
And that is ignoring the 90% margin that API inference runs at.
$1k/person sounds like a lot, but a ten person team at $1k token burn / head / month is spending $120k. How much does a new hire cost, tco? Does velocity from one more person compare with velocity from one more person’s worth of tokens?
Sure, learning to be efficient matters. So does taking the time to consider cost vs alternative. You might be surprised by what you find, in either direction.
The share of tasks where "cheaper" models are enough also increases for same reason. For majority of use cases it will quickly become race to the bottom. And that means part of the pie for the "premium" models of frontier AI labs will quickly become smaller and smaller.
1. In most cases, like this one, the product isn't heroin.
2. Drug dealers don't actually do that.
And we're talking about subscription plans from OpenAI/Anthropic here. Switching to a different model and/or harness if the price changes isn't a "the world has run out of opioids and I'm going to need a detox center" level crisis.
This is a weirdly melodramatic analogy.
If you want to flatten everything to "all things habit forming are definitionally equivalent", you'll get some pretty funky results.
I'm not sure about that. Someone using AI for productivity every day will experience massive issues if you suddenly take it away.
It's even the whole business model of MS Copilot: They don't work with companies to have their data structured in the past or going forward.
They don't help you to structure your business "life", they offer a tool you consume to make sense of it while you use it.
The moment you stop using it, you'll struggle to even make sense of your own past actions.
> "2. Drug dealers don't actually do that."
It would be a weird business if they only WAIT for people to demand the more profitable drugs...
But the short version is that there are lots of people who want to buy any given drug and drug dealers sell them. I don't know exactly what you mean by "WAIT for people to demand the more profitable drugs", but there typically isn't a whole lot of waiting. There's a lot of demand out there. Giving away a bunch of free fentanyl to random people in the hopes that one of them instantly gets addicted would be a pretty wasteful and high risk sales tactic, especially given the number of people that want to buy it already.
Someone who pays $20 a month may only use 10% of their max budget.
Different matter at work but even there we are paying for API costs and have a monthly allowance (starts at $200/month but increases fairly easy to get)
CLI Model Pct Remaining Reset Date claude all models 15% 3d claude Fable 50% 3d grok all models 0% 4d agy Gemini 36% 5d agy Claude and GPT models 0% 5d codex codex 0% 6d
And I also just used two of my three get-another-week-free tokens for codex.
If you've got an agentic swarm, or are working on multiple big projects at once, or have agents parsing data... you burn through tokens fast. The last four days in API would have cost me well over $3k.
We have two wildly different cohorts of people. Those who see these tools as a technological curiosity and those who see them as a means to a specific end. I know that if I finish a task with a certain set of constraints that I will get paid a certain amount of money. I'd much rather be HITL on one very good agent loop if there is real money on the table. If you are mostly having fun, being curious, or exploring (which is fine!) then the cost of subscriptions might feel safer. Running an IRL version of factorio can get very expensive.
I also think the API tokens perform better but I have zero evidence other than observing the quality of my work product relative to that of my peers. I have almost a religious fervor about using prepaid tokens now. I am still using GPT5.5 in my copilot with custom model configuration for most heavy lifting in visual studio. It is more than sufficient. My average monthly token cost is <$100. Paying for a subscription would be more expensive because I have a predictable and controllable amount of demand.
This is analogous to variable vs fixed rate electricity. Paying the real time wholesale price is a massive economic advantage if you can figure out how to control the demand side of the house.
Wholesale electricity is almost always cheaper.
Sorry, I have to be that guy... https://xkcd.com/386/
The analogy is false.
Domestic electricity is priced per kWh, wholesale electricity is priced per kWh, the latter is, as a rule, cheaper. It is not fixed vs variable rate.
> wholesale electricity is priced per kWh
> the latter is, as a rule, cheaper.
I think you’re missing some detail. Why do two things that are “priced per kilowatt hour” have to have different prices?
YES. You got it, thanks! That's why I was referencing #386 :)
From 500 hours working on my mobile app project outside of work, I can confidently say that using anywhere between 2 to 10 conversations to parallelize work is much more effective in my project.
The models being served in the API are the same as via the subscription, there are no shenanigans here, particularly not concerning quality.
And GPT 5.5 was great at the time, I made heavy use of it for my project, but it's long outdated and already being retired.
You also mentioned using it in copilot, which is a third-rate harness that is frustratingly buggy and inefficient.
There's no comparison to Opus 5.5 in Claude Code.
Use whatever works for you, but frankly, I think you should try something new.
> I know that if I finish a task with a certain set of constraints that I will get paid a certain amount of money.
So not salary… ok. That explains some of it. Getting paid by the task might flavor one’s approach. But also one could reason, more tasks, more pay… somewhat confusing.
> I also think the API tokens perform better but I have zero evidence other than observing the quality of my work product relative to that of my peers. I have almost a religious fervor about using prepaid tokens now
And this appears to be the other half. It looks no less futuristic than the more wireheaded approach, just different in form.
Another thing of note is that when coding at 'scale' to my surprise over time with memory aggregation across users/sessions these models can be better than the evals would have you believe because they seem to be differently/less (?) RL'ed (aka actually listen to instructions), which results in them being more fastidious about writing memories and benefiting from them, especially over time. These are of course mostly subjective opinions but they were surprising, and not what we were expecting.
Relevant: the proxy we wrote and use: https://yapp.latchlabs.dev is partially open sourced: still working on releasing more of our memories stuff.
Oct 16 I switched machine
Excludes work done on local and Claude Chrome plugin.
You spent $720 in four days on a $100 per month subscription? So that table isn't correct?
Do you mean the table shows what it'd have cost you if you weren't on the $100/m subscription?
Yep, that. There's tools out there that will add up your token usage and then estimate what it would be at API pricing. That's likely what they used.
Like a big white truck or a 600mm lens, expensive gear should be purchased by an employer. I think most folks talking about API access are getting it through work/self-employment.
Anthropic needs to retune their personal plans. Opus is the Amelia Bedelia of models. Maddening! I shelled out $100 for Fable to drive Opus and finish a personal project, but that's probably my last time loading up on usage credits. I put a second $100 towards API credits so I could dink around with API access and understand conversations like these.
People who use API keys on their dime are usually running DeepSeek 4.x or GLM 5.x (Flash or otherwise) and similar models.
RE: Subsidies:
What if subscription costs are still high enough to be profitable (I believe they are for OpenAI).
Then maybe we should be saying that API users are paying an order of magnitude more than the cost, at least at the very top end (frontier models). Likely because big cos are willing to pay top dollar for what is the best and unmatched elsewhere.
We (some friends and I) built it because we wanted to understand how these things worked and also had max accounts. It has since grown into a ton more, but is still in essence a very full featured local proxy.
I don't remember the last time this was true. These things don't have a 'just works' mode.
> We need answer explain LLM reasoning traces. Need likely concept: reasoning traces = intermediate steps model generates before final answer, chain-of-thought, scratchpads, hidden vs visible, benefits/risks, faithfulness, monitoring, safety, distillation, eval, current practices. Need explain clearly. Desired oververbosity 5.
Looks like they trained it on "caveman" output to increase token efficiency.
Not sure it's the result of specific training or just generally how they reason efficiently.
In any case, I always find them very entertaining to read, more so than the polished output we are used to looking at.
It was also interesting in the HF incident write-up or other research like "Hacker Opus", which said this gem:
It's something, at least.
Ultimately, Codex doesn't suffer from ANY of this so I'm starting to think about just using it for work subscription use as well.
$100 a month on an LLM subscription for hobby usage is an absurd amount of money IMO.
If you're the kind of person who gets enjoyment being able to code all sorts of things you were never able to before, or 20x faster than you were able to before, then $100/mo. is extremely reasonable in comparison.
I'd argue it's money much better spent than on cable.
Also, if a hobby of yours runs you $100/mo., that's actually very much on the cheap side for a lot of hobbies.
Also, $20/mo. is for hobbyists using it for 2 hours at a time a few times a week. $100/mo. is for people working on a side project every evening and every full weekend.
...proceeds to use codex.
With just a 12gb VRAM GPU + 64gb of CPU RAM, and you can actually own your own infrastructure. I can't wait for hardware prices to fall so that becomes even more accessible.
I happen to be working on an extension to Aperture, so I have my own reason to promote it. But I do think it's a nice easy solution, and I promoted it to people before I was extending it.
I'm building a "bring your own data" layer for agentic work of all sorts, using the webhook method for now.
In the context of Aperture, I watch for agents looking for certain types/sources of data, and instead provide them with an in-tailnet API for accessing that data.
In a personal context, think email, calendar events, music playlists, fitness app records, transactions. I've also found it useful in a professional context, but I think I need to explore that a fair bit more.
It keeps credentials out of agents hands, and provides some relatively standard ways for controlling what agents can see and change.
I need a fair bit of hardening work before I try to get random folk to use it, but I might post something more here in the next week. :)
It is a not proxy. It can clearly show whats going on across all agents. The proxy's are generally inconvenient and yet another point of failure. I think the author got to the resort option quickly
my own ai gateway/proxy
For subscriptions, it is easier to collect the JSONL files from ~/.claude/projects/ and similar directories. They are better organized (e.g., subagents) than what the proxy sees.
However, the best value comes from collecting agent sessions across the whole organization. You'll need to scrub credentials and handle security.
I did exactly this - probably the single most valuable piece of software I've "written". It's the backbone of every piece of code coming out of any agent under my control (via MCP toolset).
I just wrapped around the Mlflow setup to make this simple CLI helper to set it up. Ofcourse the name suggests only Claude but it’s Codex, CC, GHCP and Agy. https://pypi.org/project/claudetracing/
https://github.com/CauchyIO/claudetracing
https://grady.dev/projects/pharos/
So id be very careful with chats and treat them like chats, not like code.
Is my brain now subject to GDPR regulations?
Not an attack on you, but my word am I thankful I don’t participate in the European regulatory regime.
not an attack on you, but I am thankful to have guardrails for greedy scrupulous corporations
If a company doesn't follow these laws, the fines are proportional to the revenue(?) of the company, so the fines can get quite existential if a company fucks up.
In practice, this is not "shoved down" my throat at all. In practice, this means that I can request companies delete all my data, and they respect this, in practice. I can also request a .zip file with all data they have of me, and I will get it within a couple of days max.
It also means that e.g. your family doctor has to have you sign a document if they want to enter your data into a system, or if that system might be accessed by a 3rd party (such as IT support).
This isn't theoretical; this works. I've requested data deletion from various companies and requested data exports from various companies (and it worked without any issues every time).
It also means that, at any company I work at that handles PII, everyone is VERY careful touching anything with PII. The fact that you have to implement data retention periods, limit the amount of data you collect and retain, etc. forces companies to respect your privacy even if they don't want to or it's inconvenient.
If you want "progress at any cost", particularly at the cost of human lives, you're right. We're behind in the ever so important AI race, and we're missing that privacy of people, health of people, happiness, and the future of humanity is entirely irrelevant unless it makes you money, today.
Probably only people who are interested in living in such a place, and those who care about that sort of stuff are maybe less likely to choose total annihilation and exploitation for profit, so you see less multi-billion dollar companies that have to promise they're not evil, I guess?
It's a different culture, that doesn't necessarily make it bad.
If a company can't deal with PII safely, they SHOULDN'T be a company in the first place. If you can't have a zoo without starving the animals and running them over every now and then, maybe it's fine that you can't start a zoo, even if countries exist where you could do that instead and be fine (as long as you hide the bodies).
Right now I run my Claude-Code sessions through a very simple mitmproxy, logging everything that goes in & out. It's very low level, sure, but it works.
Then I shared a sub with 3 people in total and we were detected within a few days.
I don't know what metrics they use, because I use my sub on a handful of servers and different IPs. But I'm guessing there's some kind of fingerprinting going on, where they can actually determine that prompts are being sent by different people
The actual question is: do many programmers work in businesses that have < 150 people?
Primaily designed as a local LLM ops project, but will definitely do proxying for claude-code, opencode, codex, etc
https://llama-dash.dev