Token Town 001: What Is Anthropic Doing?!  🔗 **Source:** [Token Town 001: What Is Anthropic Doing?!](https://share.snipd.com/episode/9605ebc3-0391-4a5e-86fb-7b5d7ec939f7) ## Highlights - **Anthropic Shifted Programmatic Usage To Token Billing** - Anthropic reclassified non-interactive uses like claude -p and the Agent SDK to be billed at API/token rates instead of subscription coverage. - That move forces heavy GUI/third-party tool users to pay for token-heavy programmatic calls that previously rode on subsidized subscriptions. Transcript: Ian Landsman So we're going to go right with the big topic, which is Claude clarified and changed their pricing. So that now you can't... Yeah, clarified. Clarified. Yeah. Clarified it so that on your subscription plans that is for interactive usage in their clients and then any other usage that's programmatic whether you're using claude dash p to call It on the command line or using their agent sdk which i don't know a ton about but i assume it's for embedding it in other things, then that's going to be billed by the token, except you get A credit for the equivalent amount of your subscription. So if you have a $200 subscription, you get $200 in programmatic credits and that you will use any programmatic access. And then when that runs out, you can choose to either obviously have it stopped or to, you know, have overages and pay by the token. Aaron Francis This is a furtherance of the spiritual open claw restrictions. So you'll maybe remember that they have basically applied a one-off patch to restrict open claw usage in the past. That's old news. What's new news, which I guess is just news, is that they have further clarified or tried to clarify that all non-interactive clawed usage is now billed at the API rates. So that feels like a huge punch in the face, but then they also gave us a little bandaid of, oh, but we're giving you a few extra tokens to fix the punch in the face we just gave you. So yes, I think you set it up correctly that clodp, which is the headless kind of like one-shot execution mode on the cli clod p and the agent sdk are no longer covered under your subscription Plan which is a huge blow to tools like conductor polyscope t3 code anybody that puts a pretty UI or GUI on top of a Cloud session was probably using the agent SDK, maybe in an outside case Using Cloud P, and those are now hosed. ([Time 0:01:51](https://share.snipd.com/snip/c5eeec78-d997-4a79-84cd-2787d7da16cd)) - **Change Is Capitalism Driving Whales Off Subsidized Plans** - Aaron frames the change as a business decision: Anthropic is reclaiming costs from heavy 'whale' users who were subsidized by low subscription prices. - He reasons Anthropic wants to push GUI/desktop clients toward paid enterprise/team plans to capture telemetry and revenue from high token consumption. Transcript: Aaron Francis My initial reaction is welcome to capitalism. That's my initial reaction. It's like I'm the company. I do whatever the hell I want. And in that regard, I respect it fully. They are, they are, this is not a public good. This is not free speech. This is not gas, water, power, air, lights, whatever. This is a private company making their own decisions. Now, I think the more interesting question is, is this a wise decision? We don't have enough information to answer that. But my read on it is there are going to be, and are, a lot of people on Twitter that are mad, which as I'm sure you will tell me, Twitter does not a market make. Twitter is the worst. Yeah, that's your bit. That's your bit. So whether that's representative of the broader Claude user base, customer base, I can't speak to, but I think, I still think it's somewhat unclear. They've tried to draw the line between interactive and non-interactive. I think if they don't, I don't know. I think think they're trying to push people to use cloud desktop that that's my that's my renewed take on it is they let a thousand desktop clients bloom conductor polyscope etc and on Subsidized usage which is their right and now they're pulling that back and saying oh if you want to use their desktop clients, you're going to have to pay through the nose. But if you want to use ours, it's closer to free than not. And I'll take a little bit, I'll push back a little bit on the $400 of usage. I think if you're on the $200 max plan, you're getting like $4,000 of usage. Oh, I agree. And then that's the problem, right? You've been using... But you're only paying $200. I agree. I agree. But that's why people feel like this is a red pool, because you've been in conductor burning $4,000 of value while paying $200. And now if you want to burn $4,000 of value, you're going to have to pay $4,000. ([Time 0:05:14](https://share.snipd.com/snip/a4b2ddb1-f113-445b-90f0-ae183d816147)) - **Enterprise Adoption Gives Anthropic Pricing Power** - Ian notes Anthropic's business momentum gives them leverage: adoption by enterprises has surged and revenue multiplied, enabling stricter policy shifts. - He argues the company can now prioritize migrating fragmented individual $200 plans into negotiated team/enterprise contracts. Transcript: Ian Landsman What's given them the power in some ways to do this is like there was just a, one of these companies that tracks this stuff, just how to graph out, showing how like Anthropic use has now Passed OpenAI in business for the first time. And we know like Anthropic's like 80X their revenue in the past eight months or whatever. And so, you know, that gives them a lot of power to be like, hey, people are loving this. People are using it. That is not all Twitter devs correct using two hundred dollar month plans right this is big companies spending hundreds of thousands of dollars a month in tokens and so i think that part Of it is that in the sense of like we want to get all these companies that are right now there's still a lot of companies that are just like hey i gave everybody use your business card, buy $200 a month plans, individual plans, right? Let's get those people, let's get them into team plans, into enterprise plans. Let's, enough of this, like everybody's just out there, like being individual buyers, because they're not. They're working on a team with 30 other people. They're all buying cloud individually. Let's get them in teams, maybe for big companies doing this so we're going to give you some discounts it's a now negotiating thing right if you're really using a ton of tokens and and the Businesses aren't really going to care like the the businesses are like yeah if you're more productive with this i don't care if this costs me more tokens now so uh you know more costs Than those tokens i mean they care a little bit but they don't care that much if you're actually being more productive with it. ([Time 0:07:42](https://share.snipd.com/snip/00a5694f-0a1b-42a9-8b3e-38cb254181b1)) - **Opt Into Anthropic's Temporary Credits Now** - Anthropic temporarily increased weekly limits 50% through July 13th as a PR cushion after the billing clarification. - Check and opt into the $200 programmatic credit offering in your dashboard because Anthropic did not auto-apply that credit. Transcript: Ian Landsman Well, this, I mean, this was weird because it happened right after the other thing and it was like, is this, I guess this is a make nice type of thing, but it's basically, you know, the quad Code announced that weekly limits are 50% higher through July 13th. You know, this is this whole game both of them seem to be playing a lot of, right? Like, I feel like every week, both OpenAI and Anthropica are like, hey, free tokens, more tokens, limited time tokens. You're it's oh, it's a long weekend, more tokens like it's token town, baby. Yeah, token town, that product manager guy whose name I can't remember for OpenAI is always like, reset the tokens. Yeah. Yeah, I reset the tokens for everybody. Go have at it. And so they seem to have found this carrot, right, as a way of like when things are a little rocky or when there's an opportunity to pull from the other to, hey, we're throwing some free Tokens out there. But I mean, I don't know. These things are so hard to know. I assume it's effective for them because they seem to be doing it all the time but i don't know i don't know what what my take on is beyond that yeah i think this was i think this was pure damage Aaron Francis Control frankly and it was obviously planned ahead of time so they put out the tweet that's like hey the deal has fundamentally changed um everybody went ham on that one and then they Were like hey by the way weekly limits are 50 higher um and i think it was just trying to like split split the media coverage on yesterday's announcement frankly i will say um the the 50 Weekly limits that goes through july 13th so that's two months the uh upcoming subscription interactive non-interactive goes in to effect one month from now um and then this 200 sorry I punched you in the face thing that has to be specifically opted into that part yeah we didn't talk about that part that's an important so it's like hey you know we're cutting you off but We're giving you 200 if you go opt into it it's not automatic and that's like guys like come on what could it possibly cost you to just save that pr disaster and opt everybody into that 200 Ian Landsman Automatically i had to assume it's because there's like they just think there's so many claws out there running that people aren't even using anymore and things like that. And they're like, well, we can opt them out of it and they'll turn off after they hit the limits, I guess. I don't know. It's very weird. They've done this like the second or third time they've done this. They did it with some other thing too, where they were giving out the free credits or something. They're like, but you have to opt in. And it's like, go to your dashboard. Yeah, you sent me an email about it. Just like put the credits in my account. ([Time 0:10:28](https://share.snipd.com/snip/1e1b314a-4466-42c0-86e3-170a2fb351fb)) - **Solo Term Orchestrates Claude And Codex Workflows** - Ian uses Solo Term as an orchestrator: he writes in Claude, then spawns Codex for review and bug-finding, combining Claude's conversational flow with Codex's thoroughness. - He calls Codex his 'German reviewer' and uses Solo to switch models per task for specs, code, and reviews. Transcript: Ian Landsman I am rocking Solo Term by Aaron Francis is my orchestrator of choice or my harness of harnesses, if you will. Meta harness. I'm going to own that phrase. Meta harness. Oh, you should actually. I am. You should actually trademark that. Oh, that's a good idea. Yeah, you got to trademark it. I got to get you business minded. That's for the other show. Okay, that's for the other show. You got to trademark that. That's for the other show. Yes, the meta harness is solo term. And I'm still rocking the Claude, man. I'm a Claude guy. Even Josh Pickford also said he misses the Claude already after 10 hours being away from it. He misses the Claude. I feel like the actual software of Claude Code itself, as well as the way the model reacts to things, I find superior in Claude. I do acknowledge that Codex is very good. And technically, I like a lot of the stuff it can do. I don't like using it. And solo term solves that problem for me because I do stuff in Claude. I write specs. I build software. I write code. And then if it's anything of significance, I say, hey, spin up a Codex client and take everything we just did, feed it to Codex, get its feedback, and integrate its feedback, blah, blah, Blah. And then it does that. It works great. It finds bugs. It has different ideas. We get them working together. And so I get the good side of Codex without having to actually use Codex. And it's great. Aaron Francis So I have a small Codex plan and a big Cloud plan. You use Cloud CLI because you like the vibes of the actual software. And then if you need more of the like rigid German engineering, you have Claude call off to Codex via Solo. Is that right? Ian Landsman He's my German reviewer guy. Right, exactly. Like, I mean, I'm fine with Claude, I think does a good job with specs and the code. I'm actually pretty fine with everything with Claude, but I do think there are advantages to just the different models, how they see things, what they choose to check. And yeah, and Codex is at times right now, I feel like can come off a little more thorough. One of my issues with that is sometimes it sounds very thorough. Sometimes it's too thorough. Yeah, in fact, it's just gone crazy. So that's one of the things I don't like about codex, but it doesn't seem to do that too much in these scenarios. If you say, hey, look at these commits, review the code changes, blah, blah, blah, it seems to stay on task pretty well with that. And at least so far, which I've been doing this for a while now, it holds up well. So that's pretty much my AI stack in terms of what I'm actually coding. I'm not doing anything really in anything else. I'm in Solo. I got Claude going. I'll have a Codex if I need it. But mostly I use it through Solo's MCP of Codex. And that's kind of the stack. ([Time 0:15:47](https://share.snipd.com/snip/d19b80fb-9174-4736-a6c9-d8c13c539d50)) - **Prepare For A Long WhackAMole On Desktop Clients** - Expect an ongoing 'whack-a-mole' between desktop clients and provider restrictions; Anthropic may deploy telemetry or heuristics to detect programmatic terminal input. - Prepare for friction: desktop-based interactive workflows may require workarounds or formal enterprise licensing. Transcript: Ian Landsman I do think there's always this level too. It's like in software, in my own product, HelpSpot, like we have licenses and you can put a certain amount of effort into preventing people from hacking license. But at some point, like some people just do it and that's fine. So they're going to have the same thing. Like if it ends up being a low number of people who jump through whatever hoops are necessary to actually like make interactive clawed work for claw or whatever. And if it's a pain in the butt to do it, like I think they'll just leave it and it'll be fine. But if it becomes really easy to do and everybody does it and it's costing them a ton of money, then yeah, they'll probably, then there's the next layer of how do we prove we're humans at The terminal typing these strokes and all that stuff. So yeah, this is what's bleeding edge here. You know, we're on the bleeding edge. ([Time 0:23:52](https://share.snipd.com/snip/0b788f92-253d-4a2f-8938-c06ea81193c0)) - **Use AI To Hack Yourself Regularly** - Turn AI tools loose on your own codebase to find vulnerabilities because models like Mythos and GPT 5.5 Cyber dramatically accelerate attack planning. - Run automated AI-driven pentests to catch common low-hanging exploits before adversaries run bots at scale. Transcript: Ian Landsman Man, it's scary out there. I think if you're not spending a lot of time thinking about your security posture, you are making a tremendous mistake, and pretty soon you're going to pay very badly for that. So you've got to be locked in on the security stuff we've been doing all security stuff for the past month or two um yeah and we have these i mean the current models are already great like If you really send a model thoroughly through your code base it's going to come up with like five six huge issues and it's going to come up with like 200 other secondary but not inconsequential Type issues, most likely of like things that if, if thing A goes bad, this thing B will then be bad for you. So yeah. And we just had the report out how, uh, you know, the, if we ever see mythos, but the people have access to mythos anyway, uh, on the latest previews, like it's, you know, beat some evaluations The first time that in the past only humans could do this one specifically there's a 32-step like corporate network attack that they estimated takes 20 hours for a human to complete All the steps and mythos was able to do it in six out of ten attempts at doing it um and then also the chat 5.5 cyber version also has, you know, advanced the capability well above what 5.5 On its own does. Not quite apparently to Mythos level, according to this data, but still very good. Still far exceeding everybody listening to this, your capability. And even Opus 4.7 and 5.5, you should have them in there. Yeah, I'm a little irked that they have these tools and they're not letting us poor little guys, like only Cisco systems get to use it and not poor userscape and try hard. But hopefully soon we will have access to help defend ourselves. Aaron Francis So this is a report from a Twitter account that I am not vouching for, but it sounds very official. It says the AI Security Institute. And And it's like, oh, it's an institute. That's amazing. And it does have Mythos Preview as the top cyber hacker followed somewhat pretty closely behind GPT Cyber, which is GPT 5.5 Cyber, which is different than GPT 5.5. Ian do you have access to GPT 5.5 cyber is that where you had to like scan your eyeball and pledge your firstborn to Sam Altman? Ian Landsman Well Sam Altman took all my data but he did not actually give me access to 5.5 cyber so I do not have access to it. Yes. They gave me access to a security tool they have which is different than these other security tools they're all working on. But yes, not yet. Aaron Francis So what's the recommendation here? Because Next.js, may it rest in peace, has had a new RCE, a new vulnerability like maybe once or twice a week. Tamstack got pwned by no fault of their own from a GitHub cache poisoning thing. We've seen targeted attacks on maintainers to get worms inside that. So what are we supposed to do? Are we just supposed to turn our tokens to like internal hacking? What's the plan here? Give me 30 seconds on what the hell we're supposed to do. Ian Landsman Some of this stuff is like way beyond your control, right? Like, I mean, you can never pull in something from NPM again. Like, it's probably not possible for you to do that. Obviously, companies are trying to like build in between layers and also had delays and all that stuff. But there's workarounds for all that. I think we're just in for a rough year or two. Like, I think that's just reality. But yes, you should be turning the AI on yourself and having it hack you. Because to me, especially more in smaller situations where a smaller company, the biggest risk is more from AI bots just crawling around that people are going to send out there and trying To find obvious exploits. Basically the same thing as the humans used to do, but now you can do it 10,000 times faster and more effectively. So you got to make sure the really sort of low-h fruit stuff is taken care of. The rest of it, I don't have any good answers on. Like I don't have any good answers for when Next.js has a hack and like, but your whole app is Next.js. Like, I don't know, like, are you never supposed to update? You need to update to get the security patches. So what are you going to do? ([Time 0:29:13](https://share.snipd.com/snip/7719ab20-7d4a-4f76-93ca-7e801474a098)) - **Pin Dependencies And Enforce Minimum Package Age** - Pin dependencies and enforce a minimum package age to reduce risk from supply chain attacks and rapid malicious publishes. - Example: disallow packages published within the past 5–7 days to avoid immediate poisoned releases like the TAM stack incident. Transcript: Aaron Francis To keep So my narrow recommendation would be like this chatter just said, there are ways to pen versions in your dependency files and there are ways to enforce a minimum age. So you can't install something or rather you disallow your package manager from installing something that was published in the past seven days or whatever. Because like the TAN stack thing, you know, security researchers found it in like 12 hours or whatever. So keep your dependencies as pinned as specifically as possible, but if not, enforce a minimum age of five, seven days or something. ([Time 0:33:20](https://share.snipd.com/snip/9ecfd398-8f92-43f1-b3c5-14fc6b048a44)) - **AI Enables Massive Rewrites If Black Box Tests Exist** - Large-scale rewrites driven by AI are now feasible if you have strong external black-box tests, as shown by Bun's million-line Zig-to-Rust conversion. - Jared rebuilt internals in weeks and validated correctness because comprehensive outside-in tests kept the public surface stable. Transcript: Aaron Francis Bun uses Claude to convert to Rust from Zig in a one million line pull request. What are we thinking about this, Mr. Landsman? Big old pull request. A million additions. Ian Landsman Only 4,000 subtractions. So I'm not sure what's going on there. I assume they just left the other code mostly in I think it's the Canary release, so I think they rolled them both in. Yeah, so a million lines of code. I got the impression, I think he said, I couldn't find the tweet quickly, but I think he only worked on it for like a couple weeks or something like that, where it was like him and Claude. You know, Jared Sumner works at Claude now. Sorry, you know, he obviously has access to a lot of resources and as many tokens as he needs, presumably. But he got in there and rebuilt, converted the entire thing from ZIG to Rust. It sounds like mostly for the, like, compiler tooling for memory leaks and things. And that speed is, like, maybe a little bit faster, but it's not primarily about speed. But this is fascinating as somebody who runs an old app like the idea of like hey let's just convert this whole bad boy to something new in a week you've been running you've ever been running Aaron Francis A help spot for 20 years you could rewrite it all in a week yeah maybe i don't think i have the test coverage he has some amazing test coverage it sounds like where it's like perfect test Ian Landsman Coverage. Aaron Francis I think that is my, that's my big takeaway on this is he said all the tests still pass. And that immediately opens the question of, well, what are the tests written in? Which leads me to believe, and it has been confirmed by Jared later, that the tests are basically like testing from the outside in. And so he's able to rewrite all of the core and then run the test to make sure that it still passes. I think, yes, his stated goal, it got a little bit smaller, I think like 3 to 8 to 12 megabytes smaller in terms of distribution size. But his stated goal is to paper over or rather eliminate some of the frustrations he was having with zig things that are beyond my level of programming expertise but stuff like memory Safety borrow checkers all that kind of stuff so he switched it all over to rust um i i mean i think the value of black box testing and end to end testing is going through the roof right now, Because as long as and this is this is a big, big asterisk, as long as your app still works, it doesn't matter how often you rewrite it. Works. Like that's not trivial. But if in this case, Jared is saying my app, which is Bun, my app still works. I just happen to rewrite it all underneath. ([Time 0:34:25](https://share.snipd.com/snip/ec965da9-c962-4b50-a588-8cb1bdbe984c))