A couple of months ago I wrote Looking for tokens in all the wrong places, where I tried a pile of coding subscriptions looking for a good, affordable place to run open-weights models. I ended it with “ask me again in a month,” so here we are (a little late).

The question I was really trying to answer is whether a stack of cheap open-weights plans can stand in for a $100 a month Claude or ChatGPT subscription for real, daily use.

The short version is no. But we’re soooo close! The models are getting good and the plans are cheap. The hard part is what happens when you run out.

If you caught the (not so) recent Hacker News thread asking why nobody’s freaking out about DeepSeek 4.1 Flash, you might have come away thinking open models are expensive. The top comment says “I tried the cheapest provider on openrouter and burned through $50 in a few days,” which is a quarter of a Codex subscription. I get it, but that’s a pay-per-token bill up against a heavily subsidized flat-rate plan. Open-weights models have flat-rate plans now too, and those are what I’ve been trying.

Where to get them

OpenCode Go Plus

In part 1 I said that if OpenCode sold a $50 version of Go I’d probably buy it twice. On September 28 they shipped Go Plus for $40, which is close enough.

Plain Go ($10) is a bit of a dabbler’s plan, with about 490 Kimi K3 requests a month. Plus gives you 2x to 8x more depending on the model, and the biggest jumps landed on the ones I use most:

Model	Go ($10) requests/mo	Go Plus ($40) requests/mo	Bump
GLM 5.3	1,080	8,640	8x
Kimi K3	490	1,960	4x
Hy4 preview	6,770	27,080	4x
GLM 5.3 Flash	31,580	94,740	3x
DeepSeek 4.1 Flash	173,330	346,670	2x

Those are OpenCode’s own estimates for very cache-heavy requests, and the rows don’t stack, since all the paid models share one pool. OpenCode also lists Kimi, GLM, Hy4 and MiMo as zero retention and not used for training, so Go is the one I’d pick for anything private.

rout.my

rout.my is the first place I’ve found that gets you a usable amount of Kimi K3 without breaking the bank. You buy a daily token pool, each model eats into it at its own rate, and it refills every night at midnight UTC.

Use kimi-k3-fw, which is Kimi served through Fireworks, a good, reputable provider that for some reason is absurdly cheap here. It costs 4x against your pool, so the $15 Basic plan (10M tokens a day) gets you about 2.5M Kimi tokens a day, and the $25 Plus plan about 6M, which is enough to actually work with. GLM 5.3, DeepSeek 4.1 Flash, Hy4 and MiMo 2.6 are on Basic too.

One thing to know before you sign up: rout.my seems to be a Russian company, you pay through Boosty or Patreon, and it’s not clear how private they are. My experience has been good, but I’d keep sensitive code away from it and stick to the open-weights models there.

A few more

ClinePass, which I’d just bought when I wrapped up part 1, is fine, basically OpenCode Go with a few more quirks. Its models endpoint lists plenty of models the pass doesn’t cover, so stick to the ones marked as ClinePass models and call them by their full cline-pass/<model> name.

Command Code’s $10 GOAT plan gets you $70 of credits, mostly on open models.

Ollama Cloud is still a solid place to get GLM, but it’s swapped its weekly limits for a monthly allowance ($60 of credits on Pro), which makes running out a lot more painful.

Use a separate card for each one

I learned this one the hard way with Synthetic. Their cancel flow shows a little robot saying goodbye and a message that your account is cancelled, but back on the dashboard the subscription is still active. I’m pretty sure it’s a bug, but I couldn’t find a way to reach support, and my question in their Discord never got an answer.

Luckily I use privacy.com to make a separate card for every provider I test, so I just paused the card. If you’re going to try a few of these, do the same.

The gray market

There’s also a whole parallel market of Chinese token resellers. I’ve tried a few but won’t name them. They feel legally gray to me, partly because many seem to resell access from subscription plans, and a lot of them quietly swap models on you (you pay for Opus and get Sonnet). If you want to go that route anyway, Veridrop (in Chinese) ranks thousands of them from public test reports, and for Claude it checks cryptographic thinking signatures, which are a lot harder to fake.

What they’re like to use

Kimi K3 is my favorite. It’s nipping at the heels of Opus 5.5 on personality, it’s quite smart, it has a genuinely different point of view, and it’s refreshingly short-winded and clear, something even Opus struggles with. It’s not Opus 5.5 level smart (44 on the Artificial Analysis index vs 58 for Opus 5.5), but for a model whose weights came out nearly three months ago (an eternity in AI years), it really holds its own.

GLM 5.3, full and Flash, is the workhorse. GLM 5.2 was in an awkward spot back in part 1, but 5.3 is great, and Go Plus finally gives you enough of it to lean on.

DeepSeek 4.1 Flash is good, but if price is all you care about, OpenAI’s GPT-6 Luna beats it, especially through a ChatGPT subscription. That’s the original 4.1 Flash, though. DeepSeek likes to upgrade its models in place under the same name, and on October 17 it shipped a new 4.1 Flash checkpoint called 1017 that it says is faster and cheaper per task. That could change things, but I haven’t tried it yet.

Hy4 from Tencent is surprisingly effective for how cheap it is. It’s good enough for a lot of work, and it feels smarter than DeepSeek 4.1 Flash.

MiMo 2.6 looks great on paper, but it felt slow when I tried it. That may have been the provider, so I’ll probably give it another shot.

Two didn’t work out. Qwen 3.8 Max, which I was excited about in part 1, thinks for so long that it’s hard to use. DeepSeek V4 Pro 0813 is slow too, and since its August price hike there’s no reason to pick it over 4.1 Flash, which is cheaper and scores higher anyway.

And if you’re worried about censorship, most of them will happily talk about Tiananmen Square. I tested that in Chinese models aren’t actually that censored.

The catch: how fast your usage comes back

The catch with flat-rate plans is how long you wait once you run out. In part 1 I said rate limits are as much a psychology problem as a math problem, and the reset schedule is a big part of that:

Plan	Resets	If you run dry, you wait
rout.my	Daily, at midnight UTC	Less than a day
Claude Max	5-hour session limit, plus a weekly limit	Hours, up to a week
ChatGPT Pro	Weekly	Up to a week
OpenCode Go / Plus	5-hour and weekly windows, monthly total	Up to a month
Ollama Cloud	Monthly	Up to a month
Command Code	5-hour and weekly windows, monthly total	Up to a month

OpenCode Go lets you use 20% of your month in any five hours and 50% in any week, so two heavy weeks can burn the whole month. Ollama’s new allowance doesn’t even have the shorter windows, so you can burn a whole month of it in a few days. If you code regularly like I do, that’s a long time to sit on your hands, and it’s a uniquely frustrating feeling when inspiration strikes and you’ve got no tokens left.

The best open-weights stack I can come up with

If I had to go all open weights today, I’d lean heavily on rout.my. A big daily quota is just hard to argue with. The $25 Plus plan gets you about 6M Kimi tokens a day (or twice that in GLM 5.3), and it all comes back every night.

For work code, or anything else where privacy matters more, OpenCode Go Plus ($40) is the pick instead. Run both and you’re at $65, still well under a $100 Max plan.

Sooooo close

Much as I’d like to run mainly on open-weights models, the American subscriptions still win if you can afford them. I’m not convinced they’re better bang for the buck, but if you want to get work done without fiddling, Opus 5.5 on Max is the best deal going right now.

Opus 5.5 is more than an ideas model now. It’s cheap enough to use for real work, it’s really pleasant to work with, and it just “gets it” in a way the others don’t. As a friend put it, it’s the first model “that’s good at everything.” It’s also more efficient, so the Max plan feels like it goes further than it used to.

Sol 6.1 got a similar efficiency bump, and like Opus you can just tell it what outcome you want and it’ll figure out how to get there. It doesn’t mind meld the way Opus does (OpenAI’s models have always been the crazy smart but maybe a little surly guy in the corner), but it makes up for that by being really persistent and creative, and it’s great at code architecture. They even fail in opposite ways. Give Sol something measurable to optimize and it’ll chase the number and forget what the number was for, while Opus tries so hard to read your mind that it sometimes ignores what you literally asked for.

OpenAI’s Pro plan used to be my favorite because you got an absurd amount of tokens. Now they’re cutting it in half. Existing subscribers like me keep the old allowance until October 29, and if you want more there’s a new $500 plan.

OpenAI’s reasoning is that the models are more efficient now, and Tibo Sottiaux says you’ll “still get more work done than if you were on the Pro $200 subscription one month ago.”

Maybe so, but the frontier keeps moving, and paying through the nose for these plans is supposed to get you the best intelligence, not second-best but affordable intelligence. How I spend an efficiency gain should be my call, not theirs. This is the first time it’s felt like OpenAI has done me dirty, and Tibo following up the announcement with “That was yesterday, today is DevDay. And it’s all good news.” was a bit gaslight-y… I’ll decide for myself what counts as good news, thank you, and cutting my usage allotment in half definitely isn’t.

It doesn’t help that the efficiency gains are almost all in Sol 6.1, which comes within a point of Astra for under a quarter of the cost per task. Astra’s own 6.1 upgrade got pulled before DevDay, so OpenAI’s smartest model gets the same halving with none of the offset.

Anthropic is heading the other way. Opus 5.5 gives you more frontier for the same money, and Fable is at least on par with Astra.

Anthropic does have a five-hour limit on top of the weekly one which can grate. OpenAI’s Pro plan only has a weekly limit, and Tibo was “committing to not reintroducing the 5h limit,” so they picked half the allowance over five-hour windows. Given how much the reset schedule matters, I’m not sure that’s a great trade. Tibo also says he’s got a plan that “should result in significantly more usage for everyone,” so we’ll see what they come up with.

What I’m actually using now

• Opus 5.5 on Max for thinking and real work. I barely touch Fable anymore.

• Sol 6.1 for well-defined outcomes I want it to grind toward, and for code architecture.

• Luna for minions.

• Kimi K3 and GLM 5.3 through rout.my and OpenCode Go, for a second opinion and as backup.

I’d love to flip that list around. A plan with OpenCode’s privacy, rout.my’s daily refill, and enough Kimi to work on all day would do it. Nobody sells that yet. Ask me again in a month?