I’ve been trying to get off the frontier labs for a while now, for reasons I got into a bit in “good enough” models. The short version is that the cheap models have gotten good enough that paying frontier prices for mechanical coding tasks feels silly. My first thought was to try out the open weights models that have been coming out recently, but this turned out to be harder than I’d expected. Not because of the models themselves, but because finding somewhere decent to access them is surprisingly complicated. What follows is a haphazard chronicle of roughly a month of my poking around at various coding subscriptions, and how each poke has worked out so far.

If you just want the (entirely vibes based and subjective) numbers, the scorecard is at the bottom.

Every score in this post comes down to three roles. They’re the roles I’m trying to fill, so when I try out a provider the question I’m actually asking is which of these it can cover, and how well.

The idea man: The one I brain dump (and brainstorm) at/with. It pokes holes in what I’m saying, helps me see how a design might be wrong or come up with a better one and acts as a creative partner. As of today, for me, Fable is by far the best at this. Kimi K3 and Opus do a decent job too, but aren’t in Fable’s class.

The polisher: Takes the plan, finds the places where the pieces don’t fit and fixes them. Sol is the best here, but GLM 5.2, OpenAI’s Luna on max reasoning do well, and DeepSeek V4 Flash does surprisingly well for its price too.

The minion: Takes a pre-digested slice of work and writes the code for it. I usually use Luna on max or DeepSeek V4 Flash for this. This is a recent development though, up until this month I used GLM 5.2 for this (more on that later).

There’s arguably a fourth role, the manager who supervises the minions, a “chief of staff” or CEO type role. Essentially an implementation manager, the thing that spawns the subagents and consolidates the worktrees. I’ve been using heavier models like Sol or Fable for this without thinking about it too much but am realizing that smaller models are actually quite capable here. GLM 5.2, for example, is surprisingly good at this. I haven’t used it as much as I should and need to test more, but it performs better in that slot than I’d have expected for its weight class.

The idea man is the expensive one. The polisher and the minion are cheap and more or less solved, but every model I’ve found that can do the idea man job well belongs to a lab that would prefer I live inside their app. Kimi K3 was the first open-weights model I’d used that could plausibly fill this role, so I went looking for somewhere to run it that wouldn’t cost a fortune.

Some background on how I code, so you can work out how much of this applies to you: I write software for a living and I’m at it pretty much every day. Some days a lot more than others, and I can be fairly bursty, but I’m a near-daily user. If that’s not you, adjust accordingly.

Ollama Cloud

Idea man 1 · Polisher 3 · Minion 4

I was already paying for Ollama Cloud before any of this started, and in practice I use it for exactly one thing: GLM 5.2. For $20 a month you get GLM 5.2 without having to watch a usage meter. And while it may be a one-trick pony, for this one job it’s been very solid.

Even though I use Ollama Cloud a lot, the scores are on the low end because of that narrowness. GLM 5.2 sits in an awkward spot these days. It isn’t quite heavyweight enough to make a good polisher, while also being on the expensive side for a minion since DeepSeek V4 Flash massively undercuts it on price with only a relatively small intelligence hit. So a provider whose main draw is generous GLM 5.2 access is a provider whose main draw is shrinking.

That said, there’s a version of this where Ollama still earns its $20 keep, and it runs on DeepSeek V4 Flash rather than GLM 5.2. They serve it on the cloud tier and they keep it current, though you have to ask for it by name: deepseek-v4-flash:0731-cloud is its own tag and the plain :cloud one is still the three-month-old build. On Pro that’s about as close to an unlimited supply of DeepSeek V4 Flash tokens as I’ve found at this price, which is a genuinely useful thing to have running in the background while a swarm of minions chews through work.

The catch is that OpenCode Go costs half as much and hands you something like a hundred and fifty-eight thousand DeepSeek V4 Flash requests a month, which I’m not going to exhaust either. Ollama’s best remaining argument is one somebody else is making for $10.

To top things off, their Kimi K3 rollout really tarnished my view of them as a company. Ollama’s announcement tweet made it sound like Kimi K3 had arrived on the platform, and while technically true, the caveats are so absurd their tweet got community noted as misleading. Basically you can only use Kimi K3 on Ollama if you both buy a subscription and buy extra API usage credits on top of the plan you’re already paying for. So in short your $20 buys you the privilege of buying Kimi K3 tokens for the same price you can buy them anywhere else. That isn’t a Kimi K3 plan, it’s just misleading advertising.

synthetic.new

Idea man 2 · Polisher 3 · Minion 3

synthetic.new does have Kimi K3 now, but I burned through the allowance fast enough that it barely counted as a trial. The interesting thing about Synthetic isn’t the quota size though, it’s the way they handle rate limit refills.

You get $24 of credits a week, and instead of resetting on a timer they regenerate incrementally: two percent every 202 minutes, which works out to a full week to climb back from zero. Synthetic frames this as the friendlier option and the logic holds up if you’re a light user. You never hit a wall and wait three days, there’s always another tick coming in a couple of hours.

I found it miserable anyway. If you code in bursts, hammering it Saturday and ignoring it midweek, the drip works for you the whole time you’re away. If you code most days you never get back to full. You live at fifteen percent, permanently in the red, where any real session digs the hole deeper. Objectively that’s less restrictive than a hard reset. Subjectively it’s token anxiety every single day. Rate limits turn out to be a psychology problem as much as an arithmetic one.

DevPass

Idea man 2 · Polisher 4 · Minion 3

DevPass has the most legible pricing here, which is genuinely a perk. You get roughly three times your subscription in usage, metered at provider list rates. $29 buys $87 of models, $79 buys $237. No mystery, you can work out what you’re getting before you sign up.

The problem is just that it isn’t enough. Three times a big number is a big number, and $87 of Kimi K3 at $3 in and $15 out disappears pretty darn quickly. The bill stays flat either way, which is the whole point of a subscription, so the failure mode isn’t a surprise invoice. It’s running dry a week into the month.

For a mid-priced daily driver it’s a decent place to live though. Two hundred some models, one key, two environment variables and it drops into whatever harness you like.

OpenCode Go

Idea man 2 · Polisher 5 · Minion 5

$10 a month, and the best value in this post. If somebody sold a $50 version of OpenCode Go I’d probably buy it twice.

The reason it works is that OpenCode negotiated bulk rates on the models everyone serves and passes the savings on, aiming for six times what you’d pay at raw API rates. On GLM 5.2 that comes out to around 4,300 requests a month. On DeepSeek V4 Flash it’s about 158,000, which is more than I could get through if I tried.

But Kimi K3 only gets 490 requests for the whole month.

Four hundred and ninety requests a month, against DeepSeek V4 Flash’s hundred and fifty-eight thousand, inside the same subscription. Some of that is raw model cost, but most of it is that Kimi K3 sits in a different bucket: most of the catalog comes with $60 of included usage and Kimi K3 comes with $15. To their credit OpenCode explains why right there in the docs. For some models they haven’t managed to negotiate a discount, either because the model is too new or because its public pricing is already low.

You can watch the same split play out across the rest of their table. The models stuck in the small bucket next to Kimi K3 are the closed ones and the brand new ones, while the generous bucket is full of open-weights releases that have been out long enough for a dozen providers to fight over.

So what’s next?

Qwen 3.8 Max is the one I’m actually excited about. It’s the closest thing I’ve found to Kimi K3 on the thinking side while costing a good deal less to run, $2 in and $6 out against Kimi K3’s $3 and $15, which makes it a plausible cheap idea man and, I suspect, a very good manager. Those are the two slots I’ve been overpaying to fill all month.

It’s early days, so it’s sitting in the same starved bucket as Kimi K3 at 810 requests a month on OpenCode Go, for the reason you’d now expect: the weights haven’t officially dropped, so the only way in is Qwen’s own API or somebody routing to it. That’s the pattern working exactly as advertised, which is the encouraging part. Once the weights land it should spread the way Kimi K3 did, and a model this capable getting cheap to host is the whole thing I’ve been waiting for. Qwen might quietly turn out to be the most useful model in this post.

ClinePass

Not rated yet, just bought it this morning

ClinePass is promising, but there’s no score yet because I’ve only had access for a few hours. On paper it’s the best Kimi K3 deal I’ve seen yet though: for $10 a month you get 2-5x the standard API rate limits, and no Kimi K3-specific carve-out of the sort that hollows out OpenCode Go’s plan.

OpenCode already gave me a number to test it against: 490 Kimi K3 requests a month. If ClinePass beats that comfortably I’m done looking. If “2-5x” turns out to mean 2, it’s DevPass again.

The other thing I want to know is whether they meter Kimi K3 the same as everything else or quietly bucket it the way OpenCode does. If it’s uniform then Cline is eating the Kimi K3 margin themselves, which is great for me and probably won’t last. They already paused new signups once when Kimi K3 demand spiked.

The ones I haven’t paid for yet

Two providers I want to try mostly because their pricing is strange. CheapestInference sells time instead of tokens: reserve a daily eight-hour block and run unlimited tokens inside it, from about $15/mo for one block, or take all three for round the clock. Outside your hours the key just stops working, and somebody in another timezone using your GPU while you sleep is what makes it cheap. I like this a lot, since it matches how coding actually feels. You don’t ration thinking by the word, you sit down for an afternoon. Sadly the Kimi K3 pool is its own $149 tier and it’s currently sold out, and the cheap tier is DeepSeek V4 Flash, which I’ve already got covered several times over.

NeuralWatt bills by energy. Flat rate per kilowatt-hour, same for every model, and every response tells you what it drew. The consequence is that sparse MoE models get dramatically cheaper, because they’re priced on what they physically burn rather than on how big they nominally are. Whether that’s good for Kimi K3 in particular I have no idea, which is exactly why I want to find out.

The scorecard

	Idea man	Polisher	Minion	Price
Ollama Cloud	1	3	4	$20/mo
synthetic.new	2	3	3	$30/mo
DevPass	2	4	3	$29/mo
OpenCode Go	2	5	5	$10/mo
ClinePass	?	?	?	$9.99/mo
CheapestInference	?	?	?	$15–149/mo
NeuralWatt	?	?	?	per kWh

So which one to buy?

If you’re not chasing Kimi K3 and just want the one subscription that does the most for the least, buy OpenCode Go. $10 gets you more DeepSeek V4 Flash than all but the most dedicated tokenmaxxers could plausibly use, plus an occasional (small) GLM 5.2 swarm which should cover your polisher and minion outright.

After that it depends on what you’re optimizing for. If all you want is GLM 5.2 without having to do much meter management, Ollama Cloud does that one job and does it well. If you disappear for a week and then code for eleven hours straight, Synthetic’s drip refill is built for exactly that (it just isn’t built for me). And if you’d rather know precisely what you’re buying before you buy it, DevPass is the easiest to calculate directly.

Unfortunately if you just want lots of cheap idea men and managers, I still haven’t found a great answer. For now I stack OpenCode Go with something else and spend my Kimi K3 tokens carefully. I’m optimistic about ClinePass and Qwen 3.8 though! Ask me again in a month.

So what do I actually use after all that?

Fable for the idea man on Anthropic’s $100 tier, with Kimi K3 as backup when I’m running low on Fable tokens. The limit I hit there is the five-hour session cap rather than anything monthly, because Fable burns quota at a pretty absurd rate. OpenAI’s Sol on medium or high is my goto polisher, with GLM 5.2, Luna, or Deepseek v4 Flash for less important work or when running low on OpenAI tokens. And almost exclusively Luna and Deepseek (both on max reasoning) for the minions now.

The harness I use is oh-my-pi, a batteries-included take on the pi coding agent with very good subagent support, which makes it easy to plan with one provider, manage with another, and hand off to minions with yet another from within the same chat thread. Anthropic does their best to make this difficult since they want you to use Claude Code, but I work around this by having oh-my-pi drive Claude Code over ACP (this took some fiddling to get working reliably, details in a future post maybe).

Sometimes I’ll have Fable manage the subagents directly, which gets great results but burns through tokens quickly (still cheaper than letting Fable implement directly though). Overall Fable is unbeatable as the idea guy, but doesn’t seem worth the token cost for management work. You can get 80% of the performance for a fraction of the token cost with GLM 5.2, Sol, or Opus