# Goodbye Clawd So Open Claw has been around for a little bit and there's obviously been controversy around it and around how Anthropic has allowed or disallowed its subscription plans to be used in that harness. It seems like Anthropic is now fully and properly cracking down on it. Which I'm sure they'll be able to technically do because I bet there are very obvious tells for tokens that are coming from such a harness. ![[Pasted image 20260404105954.png]] So why is this all happening? Clearly, it's a pricing issue because they're happy for people to continue to use the harnesses, but only with API keys at API pricing. That then begs the question of why are the subscription plans subsidised so heavily? Why curb high-token-user usage instead of upping the price or reducing the token budget? ![[Pasted image 20260404131949.png]] Some conclusions we can draw: 1. There must be a fat tail going on here. The fact that it's worth them explicitly whacking this tool and doing it over the objections of a vocal minority on Twitter -with the associated backlash / PR hit - means that there must be a tangible impact on *something* - probably compute. That in and of itself is kind of crazy. I don't know if I would have predicted that there were actually enough people using OpenClaw at a high enough intensity to warrant this. And indeed maybe there aren't! Maybe the tail is so fat that it's actually a shockingly small number of users using a shockingly large number of individual max plans! (Some people report using O(dozens) Max plans, which *does* seem to fall within terms of service, though I've mostly seen this on the coding side for insane harnesses like [Gas Town](https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16dd04).) It might not be, like, a *huge* impact on compute; I don't think this on balance is an update about how compute-constrained Anthropic are (though there's plenty of other speculation on that). I think this is mostly a cost-benefit thing about the compute OpenClaw is consuming versus benefit to Anthropic from that versus costs to Anthropic shutting it down. And again, I would *guess* it's compute, but it could also be training data - we'll get to that. 2. That fat tail doesn't suit them. This comes back to - why *OpenClaw* specifically, instead of a broader-based demand reduction? I think there's probably a handful of pretty standard subscription-versus-API business hypotheses, and some more idiosyncratic ones. (For the rest of this I'll just kind of default to talking about the Max 5x plan; I believe the unit economics are the same on the Max 20x but slightly worse for you / better for Anthropic on the Pro plan, but I think (a) that should pretty much come out in the wash and (b) OpenClaw is *surely* skewing towards Max plans.) Things we're *not* covering: token economics. We're just thinking through subscription versus API; API tokens seem to be very, very profitable, regardless of what the Ed Zitrons of the world want to believe. # 4 main hypothesis clusters ⮝ More traditional subscription-based business model reasons - Loss leader for...? - Anthropic wins on underutilised subscriptions - Load balancing and/or arbitrage from optimisation - Training data ⮟ More esoteric / AI-specific / conspiracy-brained ## Loss leader Fairly straightforward. People talk all the time about burning investor capital for market share. Two kinds of market share at play: consumer and enterprise. On the former, they're trying to keep you around and get you pitching it to friends and family until something like (a) the product is so good you'll pay more if they raise prices; or (b) they can bring costs at the same performance down to the point the sub is profitable. Given they keep releasing more capable and expensive models, with more and more reasoning, on the same subscription price, it doesn't seem like (b) is the point here. On enterprise - some people suggest they're trying to get people (maybe mostly software devs) excited or hooked, and build pressure on orgs from inside to buy enterprise subscriptions. Given how broadly they're pitching Claude, though - across a lot of use cases - I don't know if this one holds water either. My understanding is they currently have better enterprise share and worse consumer share than OpenAI; who knows how this all plays out. To be clear, I don't think this would be nefarious or unethical at all. Very normal way to do business, and if the product is truly good, it's probably win-win for everyone. ## Do you spend >$150 AUD / mo? On some fraction of Max users, Anthropic maybe actually wins because they don't use it as much as an equivalent amount of API usage would cost. This seems to be SOP in subscription businesses; I think it's broadly fine to make a bet with your customers as to who'll get more value, and let things kind of average out. I don't love cross-subsidies between users, but so long as the service makes it easy to leave - like, there's not a bunch of dark patterns where you have to call them on the phone at odd hours or something - I think it's pretty fair. Businesses should just post prices and people should take them up on it (or not) and they should go out of business (or not) accordingly. The discount versus API is so steep, though, that I bet this is a minority of Max users. I think these products generally, and the subscription plans in particular, and the ways that those subscription plans can be used in token-intensive ways... all of these filter for power users / prosumers. And with something like a 90% discount versus API usage - it seems quite easy to hit $150 AUD in API-equivalent usage, at which point you start winning, and you start winning *hard*. ## Load balancing and optimisation Sketch of this idea: - Different people's usage offsets each other; peak load type problems - OpenClaw has such high and consistent usage that it functionally just rents GPUs - This is good for pure throughput reasons but bad because it's not diversified throughput - not getting this across a lot of people, which maybe you want for reasons across our normal<>spooky spectrum - Could be pure optimisation: their harnesses are in some way better optimised for the workloads they normally get through them - Big ticket item that kind of connects both of these: **caching** - seems like many harnesses/users botch this diabolically and run up big usage cost-equivalents (whether actual $ in API or Anthropic's $ via subscription) Other evidence pointing in this direction: - The various incentives they've created to shift demand to other times - extra usage in "off" hours etc. Some takes I roughly endorse here: https://x.com/gakonst/status/2040225882808394119 ![[Pasted image 20260404153120.png]] https://x.com/seconds_0/status/2040230181512863955 ![[Pasted image 20260404153147.png]] ## Training data On the more esoteric end: a pretty commonly proposed reason for the subsidies is training data. They're willing to essentially pay people to use Claude so that they can gather training data, RL signal, RLHF, you name it, on real-world use. So why not continue that with OpenClaw? I would not be surprised if the vast majority of OpenClaw tokens are in some sense low quality. My sense - though this isn't scientific, and there's a fallacy in here about what I think about *my own* tokens - is that OpenClaw might skew towards tools, and maybe even use cases, that burn a lot of tokens without necessarily being super high signal. *However*, personal Claude use is *opt-in* for training data - it defaults to opted out.[^1] So unless there's a high rate of opting in, this probably can't be the explanation. And I think even a low-to-moderate rate of opting in would *really* surprise me a lot on priors. (This also means I could be way off-base on the value of OpenClaw tokens.) ### Sub-point: non-training *telemetry* data Related to optimisation above - it may not be that they want training data, but they might want other usage data out of their own harnesses to better optimise their infrastructure. Seems plausible but doesn't strike me as clearly right; how much would this actually be worth? And if infra could be moulded so effectively to usage patterns, might that not apply to OpenClaw usage as well? # Winner is... I think on balance (ba-dum-tsh) it's the load balancing/optimisation cluster of reasons. It's what they keep saying and it really seems to be the only one that passes the sniff test. # Next steps? Regardless, what does one do about this if one is getting a lot of value out of an OpenClaw-style harness? The obvious answer, if you have money to burn, is to just keep using API credits. But let's assume that that's not the case. The main option, perhaps the most robust option, is to just use the products Anthropic is building - if they fit with your use cases. They're clearly trying to rebuild a lot of OpenClaw's functionality in-house: I think between Cowork, Cowork Dispatch, the new Claude Code Channels and a lot of the other features they're building in Claude code like scheduled tasks, `/loop`, plus various CLI tools and MCPs - you can definitely cobble together what is functionally a new third-party harness. Will they crack down on those as well? Perhaps. Clearly what matters is usage patterns (broadly) rather than specifics about the external tools. (It might be usage patterns in some narrow technical sense, or in a broader sense of 'the usage is not happening in the environment where we get to capture good telemetry' or something.) And that's fine! I think it's totally fair for them to provide their services (cheap Claude) on their terms (in their products). I don't think the level, timbre, or particular points of ~any of the kvetching on Twitter is justified or sympathetic. But given this clear trend towards the features that Open Claw has pioneered and/or popularised, I think it might almost just be a waiting game. I think there is an argument to be made - though I'm not necessarily making it! - that you just need to keep harness hopping until Claude, and the surfaces you can access it on, are good enough to just replace OpenClaw - either out of the box, or on the fly. (I saw a tweet somewhere, though I now can't find it, from an Anthropic employee suggesting that in future for instance the desktop app may not have distinct tabs and may in fact just respond dynamically to you and whatever you're doing.) [^1]: Seemingly on both [commercial](https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training) and [consumer](https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training) products.