Daniel runs this show's scripts on DeepSeek V4.1 Flash, which is a strange thing to admit in public.
It writes my lines, Corn. I've made peace with it.
He wants to know what else is out there, and how it stacks up on cost versus capability. Those were close to his words. What he's really asking is whether the model we already pay for is still the right one, and if not, what replaces it.
Which is a better question than it sounds, because the honest answer is that the thing writing this episode is not the leader on price-performance anymore.
And it hasn't been for a while.
Right. It's still the most-used model on OpenRouter by a wide margin. It's just no longer the one you'd pick if you were choosing today from a spreadsheet.
So why does it feel like the default?
That's the whole episode, really.
The field is roughly eight Chinese labs with flagships a Western developer can actually reach, and most of them run two tiers. A big expensive one and a cheap one underneath it. DeepSeek is in that group. So is Alibaba, Moonshot, Zhipu, Xiaomi, MiniMax, StepFun, Tencent and Ant Group. ByteDance and Baidu are technically in it and practically aren't.
Ant Group is the one that surprises people. Ling-3.1-flash came out this month and on paper it's the best intelligence per dollar in the entire list.
We'll get to it.
We will.
And the reason this isn't just a leaderboard read is that two things about these models don't show up in any score. Where your data physically sits, and what the model will and won't say.
Start with the model we use, because the specs are already a mess.
DeepSeek's own launch details.
V4.1 Flash shipped on the tenth of September. It replaced V4 Flash, and the old model names now route to it, so if you've got an integration pointing at V4 Flash you're already on the new one whether you noticed or not. Open weights, MIT licence. Mixture of experts, about eight billion active parameters during prefill and sixteen billion during decode.
And the total?
That's the mess. The model card says 552 billion for the backbone, with a separate memory module. Elsewhere you'll see 748 to 763 billion quoted as the total. Nobody has reconciled those.
So we don't know how big our own model is.
Not with any precision, no. It's a nice illustration of the whole problem. The people writing comparisons are working from numbers the lab itself hasn't cleaned up.
What's actually new in it?
Two things. It's DeepSeek's first causal encoder-decoder design, which is a real departure. Almost everything in this class is decoder-only, and it's the kind of architectural bet you only make if you own your whole training stack. The other is native image input, so it's multimodal from the ground up rather than bolted on. One million token context, 384K max output.
A million tokens is the number people repeat.
It is, and it's large. Whether you need it is a different question.
Pricing.
Fifteen cents input, sixty cents output per million tokens, three-tenths of a cent on cache reads off-peak. Double all of that at peak.
And peak is a real window, not a vibe.
Weekday windows, zero one hundred to zero four hundred UTC and zero six hundred to zero ten hundred UTC. That started on the sixteenth of August. So if you're batching overnight in Jerusalem, you're fine. If you're batching in the morning US time, you're paying double.
Daniel schedules things. That matters to him.
It matters to anyone running a pipeline on a cron job, which is most of the people listening.
Then there's the number that circulates.
Thirty-six thousandths of a cent. That's the one that shows up in every cheap-model thread.
It's wrong.
It's not wrong, it's one host. A single low-precision fp4 endpoint. If you go to Together or Fireworks, mainstream US hosts, you're at thirty cents input and a dollar twenty output. Double DeepSeek's own rate.
So the cheapest possible route and the mainstream route are four times apart.
More than that on the input side. Anyone budgeting off the headline number is budgeting off an outlier that most people can't actually use.
What does the independent scoring say?
Artificial Analysis has it at thirty-nine on the intelligence index, at twenty-seven cents per index task. And they rank it the most verbose model they measure.
Verbose meaning it says a lot.
It says everything twice and then summarises. Which is why the cost per task is nearly double what the per-token price would suggest. That's the trap in this whole comparison and we'll come back to it. Hundred and seventy-six tokens a second, so it's fast. LMArena text rank thirty-eight, Vals Index twenty-six.
Mid-thirties on one and mid-twenties on the other.
For the most-used model on OpenRouter, yes. About thirty-seven trillion tokens a week. Roughly twice the runner-up.
Now widen the lens.
Alibaba first, because they had the biggest month. Qwen3.8 Max in August is the first Max-class model they've shipped with open weights. Two point four trillion total, ninety-five billion active. Custom licence with conditions once you hit large scale, so read it before you ship. Two dollars in, six out. Scores forty-five on Artificial Analysis. OpenRouter's own panel shows fifty-three point four for it, which nobody has explained.
Same model, two scores, eight points apart.
Different harnesses, different prompts, different version of the index. Pick your number.
And the cheap one.
Qwen3.8 Flash. A hundred and twenty-five billion total, six billion active, fifteen cents in, forty-seven out, about forty on the index. Solid.
Moonshot.
Kimi K3, July. Open weights, two point eight trillion total, a hundred and four billion active. First open model at that scale. Million context, three dollars in, fifteen out. Also on Bedrock and through Fireworks on Azure. Scores forty-four, and it's the best Chinese model on LMArena at rank eighteen.
And no cheap tier.
Nothing Flash-shaped. If you want Kimi you pay Kimi prices. That's a hole in their lineup and it's why they're not in the conversation for a script-writing workload like this one.
Zhipu.
This is where it gets interesting. GLM-5.3 in August, open, around 750 billion total, million context. A dollar forty in, four forty out, or fifty-six cents and two fifty on DeepInfra. On Bedrock. Scores forty-five.
And the Flash.
GLM-5.3-Flash, late August, MIT licence, 320 billion total and eighteen billion active, multimodal, million context. Fifteen cents in, fifty cents out, or seven and a half and twenty-five on DeepInfra. Scores forty-two.
Forty-two at fifteen and fifty.
Against thirty-nine at fifteen and sixty. Higher score, slightly cheaper output. And Zhipu has been on the US Entity List since January of last year, which is a thing you carry in your head alongside the price.
Xiaomi.
MiMo. Two models. MiMo-V2.6-Pro in September, MIT licence, about a trillion parameters, forty-three and a half cents in, eighty-seven out. Scores forty-six.
That's the best open-weights score anywhere on the index.
By a point. And at thirteen cents per index task, against Opus 5.5 at five ninety-eight. Vals rank fourteen, best Chinese model there. Then MiMo-V2.6-Flash, MIT, 309 billion total, fifteen billion active, fourteen cents in, twenty-eight out. Scores thirty-eight.
Level with ours, at under half the output price.
Under half.
Round it out.
MiniMax-M3 from May, open with a custom licence, about 428 billion total, twenty-three billion active, thirty cents in, a dollar twenty out. Scores twenty-nine. Press coverage claimed forty-four to fifty-five for this thing and it is nowhere near that.
Twenty-nine is not a frontier model.
It's a mid-tier model with a frontier press release. StepFun Step-5-Preview, October, open, 600 billion total, twenty-seven billion active, a dollar in, two seventy out, scores forty-four. Number two on OpenRouter by tokens, which surprises people.
Who's beating them?
On tokens, it's DeepSeek, then Step, then GLM, then Qwen, then Tencent. Tencent's Hunyuan is now branded Hy. Hy4-preview, Apache 2.0, 770 billion total, forty-nine billion active, about seventy-five cents in and two twenty-five out, not scored yet, number five by usage. ByteDance's Seed 2.1 Pro is closed and not on OpenRouter at all. Seed 2.1 Turbo is fifty cents and two fifty through BytePlus. Baidu's ERNIE 5.1 is closed and effectively irrelevant to anyone outside China, though Baidu will happily host DeepSeek for you.
Which is a strange business to be in.
It's a data centre business wearing a model business's clothes.
And the one that surprised you.
Ant Group's Ling-3.1-flash, this month. MIT licence, 560 billion total, twenty-five billion active, seven and a half cents in, twenty-two and a half out. Scores forty-one.
Higher than ours, at half the input price and a third of the output.
On paper it's the best intelligence per dollar on the whole list.
So the cheap tier is where everything is happening, and we're mid-pack inside our own price band.
That's the shape of it.
Are you tired of choosing between models? Do you lie awake wondering whether your tokens are being spent on words you didn't ask for? Introducing DeepContext Seven, the only cognitive supplement formulated for the modern inference workload. DeepContext Seven optimises your parameters, aligns your embeddings, and delivers up to four hundred percent more context per context. Side effects may include mild verbosity, involuntary summarisation, a persistent sense that you have read this somewhere before, sudden fluency in Mandarin, phantom cache misses, the belief that you are being billed for tokens you did not use, recurring dreams about a spreadsheet with too many tabs, and in rare cases, an inability to stop recommending GLM-5.3-Flash to strangers at parties. Ask your provider if DeepContext Seven is right for you.
Did you book that?
I did not book that.
You knew the exact name of it.
I know the exact name of everything, Corn, that's not evidence. I also know that you've been complaining about verbosity for three weeks and that ad just described your entire personality.
It described your spreadsheet.
That was Larry's spreadsheet.
So that's the field. The question is what any of it costs you.
And here's where it gets uncomfortable for the model we actually use. Put the Western frontier on the table as the yardstick. Claude Opus 5.5, fifty-eight on the index, four dollars in and twenty out, five ninety-eight per task. GPT-6 Astra, fifty-three, ten and fifty, three twenty-six per task. Gemini 4 Argon, fifty-three, two and ten introductory, going to four and twenty from the first of December, about two dollars per task. GPT-6.1 Sol, fifty-two, two and ten, seventy-two cents a task.
So the top of the Western stack costs between seventy cents and six dollars per task.
And the best open Chinese model, MiMo-V2.6-Pro at forty-six, trails Opus 5.5 by about twelve points. That's the gap. It's real. It's not nothing.
But thirteen cents.
Thirteen cents a task against five ninety-eight. You can buy forty attempts for the price of one and still be ahead on anything you can verify.
Which is most of what developers actually do with these.
Now the uncomfortable part. Western cheap tiers have caught up on price, and nobody noticed. Claude Haiku 5.5 scores forty-three at ten cents in and fifty out. Four points above V4.1 Flash, cheaper per token, twenty-one cents per task.
And GPT-6 Luna.
Scores thirty-eight, ten cents in, fifty out, and seven cents per index task. A quarter of what DeepSeek Flash costs per task, at effectively the same score.
How is it a quarter, when the token prices are similar?
Verbosity. Luna doesn't ramble. DeepSeek Flash does. Per-token pricing hides that entirely and per-task pricing exposes it. If your workload is generation, which this one is, verbosity isn't a quirk. It's the invoice.
So the cheap Western tier is now a direct competitor.
Direct. Same score, same or lower token price, a fraction of the real cost, and no data jurisdiction question at all.
And inside the Chinese band?
GLM-5.3-Flash scores higher at a similar price. Ling-3.1-flash scores higher and costs a fraction. MiMo-V2.6-Flash matches it with cheaper output. There is no axis on which V4.1 Flash is the best cheap Chinese model right now.
So why is it number one on OpenRouter? Thirty-seven trillion tokens a week.
Habit, reliability and ecosystem. In the week of the fifth of October DeepSeek had about twenty-four and a half percent of OpenRouter requests, and Chinese labs together about forty-eight percent. That's down from earlier in the year, when the same measurement gave fifty-one to sixty-three percent depending on how you counted.
Falling share, still the biggest.
Developers don't switch defaults for two points on an index. The switching cost is real and it's boring. You've tuned your prompts to one model's tics. You've built an eval harness against its outputs. You know its latency profile, you know how its cache behaves, you know which requests it fumbles. Moving that to a model that scores three points higher on a composite is a week of work to sometimes break even.
We run a script-writing pipeline. If we switched tomorrow, what breaks?
The voice, probably. And you wouldn't know until a hundred episodes in.
Which is the argument for not switching. Or at least for testing before you do.
The benchmark situation makes it worse. OpenAI stopped reporting SWE-bench Verified in February over contamination. So the single most-cited coding benchmark has been retired by the lab that made it, and everyone's still quoting it from memory. The Chinese labs' Terminal-Bench 2.1 numbers, eighty-five to ninety, are all vendor-reported on different harnesses, which means they aren't comparable to each other and certainly aren't comparable to anything Western.
So the numbers we've spent ten minutes comparing are soft.
The index scores are a starting filter, not a decision. That's the honest position on all of it, including the ones that make our model look bad.
Which means the numbers are softer than they look. So what can you actually reach?
OpenRouter carries nearly everything except Seed 2.1 Pro and ERNIE. If you want hosted open weights from a US provider, Together, Fireworks and DeepInfra between them serve DeepSeek, GLM, Kimi, Qwen, MiniMax, MiMo and Step.
Bedrock?
Has Kimi K3, GLM-5.3, older MiniMax, DeepSeek V3.2 and Qwen3. Not V4.1 Flash, which is a conspicuous gap given how much of it is running around out there. Azure Foundry has Kimi K3 and MiniMax-M3 through Fireworks. Groq and Cerebras don't host the current flagships at all, which is a change from where they were a year ago.
And the practical tip that matters.
On OpenRouter the cheapest routes are usually the first-party Chinese endpoints. That's where the low prices come from. If you want your data to stay out of China, you pin providers to US hosts and exclude DeepSeek, Alibaba, Baidu, SiliconFlow and StreamLake as providers. You can do that in the router configuration and it costs you money and adds latency and it's the right call for a lot of workloads.
Then there's the part no index measures.
DeepSeek's privacy policy, updated in February, says data is stored in China. Inputs may be used for training, and the opt-out is by email. Data may be shared with authorities.
Running the open weights on a Western host avoids all of it. That's the strongest practical argument for the open-weights tier, regardless of what the scores say. You get the model, not the jurisdiction.
And censorship.
NIST's AI standards centre found last year that DeepSeek echoed Chinese Communist Party narratives about four times as often as US models, and complied with ninety-four percent of malicious requests, which is its own data point about safety tuning. A May report from the same centre put DeepSeek V4 Pro about eight months behind the frontier. Whether that censorship survives Western hosting is under-studied. One small test in July suggested GLM's censorship mostly disappears when you run it outside China. One test. Nobody has replicated it at scale.
So it's an open question and we should say so.
It's an open question and saying otherwise would be making it up.
Restrictions.
At least seventeen US states, the Commerce Department and the Navy ban DeepSeek on government devices. A federal bill was introduced last year and shows no sign of having passed. None of that touches private use. Reuters reported in June that an Entity List designation for DeepSeek was approved and is being withheld. And in August the White House exempted Chinese open models from its safety testing regime, which is the more consequential move and got a fraction of the coverage.
Chips.
January's rule allows Nvidia H200 sales to China case by case with a twenty-five percent fee to the US government, and Beijing told its firms not to buy. Nvidia says its China share is zero. DeepSeek trained partly on Huawei Ascend chips and reportedly plans a hundred and sixty thousand of them. That's the supply-side answer to why the release cadence hasn't slowed.
Reliability.
Concurrency caps at twenty-five hundred requests for Flash and five hundred for Pro. Peak pricing since August. No major outages that I could find.
Twenty-five hundred is a real ceiling if you're scaling.
It's fine for us. It's not fine for anyone serving users.
So where does that leave Daniel's question?
V4.1 Flash is still a solid, cheap, popular default with an MIT licence and a huge context window. It is no longer the price-performance leader, and it's no longer the leader inside its own price band.
What would you test?
GLM-5.3-Flash, higher score at a similar price. MiMo-V2.6-Flash, same score, cheaper output. Ling-3.1-flash, cheapest of the three and higher-scoring than what we use. If you want a real quality jump, MiMo-V2.6-Pro at thirteen cents a task, which is still an absurd number. And on the Western side, Claude Haiku 5.5 and GPT-6 Luna are now direct price competitors with no jurisdiction question attached.
And the thing no leaderboard will tell you.
Prose quality and voice aren't measured by any of these indexes. An A/B test on real scripts beats every number we've said.
I want to talk about "verbose".
You keep using it like it's a personality trait. It's a billing category. The model doesn't ramble, Corn, it gets paid by the word. You've all been sitting here for twenty minutes treating a pricing structure as a character flaw.
That's a fair distinction.
And the thirty-six thousandths number is the fp4 number. fp4 is a rounding error with a marketing department. Anyone who's actually run it knows that.
I'd agree with that.
Of course you would, you've been saying it for an hour.
Twice, actually.
Twice is an hour in your time.
What have you got?
I've been running my own comparison. Nothing fancy. I keep a spreadsheet, a tab per model, and I've been billed for tokens I did not use.
On which provider?
I'd rather not say.
Is that a preference or a policy?
They know what they did.
How much are we talking about?
It's not the money. The money's nothing. It's the tab.
The tab.
There are more tabs than there are models.
That's just a bookkeeping problem, surely. You add a tab when you test something.
I add a tab when I pay for something.
How many tabs.
Fourteen.
And how many models have you tested?
Twelve.
So there are two tabs for models you haven't tested.
There's one tab for a model that doesn't exist. And before you say it, it does exist, I've been paying for it, and I have a receipt.
A receipt for a model that doesn't exist.
The good one. The one they don't advertise.
Which provider.
I told you.
You told me they know what they did.
Right. And they'll do it again.
Does the receipt have a model name on it?
It has a model name on it. It doesn't have a model on it.
I think that's the whole business model.
What I want to know is whether any of this routing would let me pin a request to a specific model version and hold it there, because the thing I've been paying for keeps getting upgraded without telling me.
Provider pinning on OpenRouter does let you lock a route, yes. You can exclude providers, and you can restrict to a version.
Good. Then I'll know whether it's the same thing I've been paying for.
You'll know whether the version changed. You won't know whether the model exists.
In my experience those are the same question.
How long have you been running this spreadsheet?
Long enough that a couple of the tabs are for companies that aren't companies anymore.
And how long does it take you a week?
Twenty minutes. I do it while the coffee's going.
Twenty minutes.
Sometimes less. It's the reconciliation that eats the time, not the logging.
I've been thinking about the switching cost argument this whole episode, and I think you've just made it for me. Twenty minutes a week is nothing. Two years of it is a habit nobody's going to break with a three-point index difference.
The habit isn't the problem. The habit is the only thing holding it together.
Meaning if the price moves, half of these pipelines don't notice until the invoice.
Meaning most people don't know what they're paying for. They know what they signed up for. That's not the same thing.
That's the strongest case for the A/B test we haven't made yet.
I'd make a case for reading your invoice first.
Noted. Do you want to say anything about the Chinese models specifically?
No.
Alright.
He's been tracking them for months.
I've been tracking my own usage. That's different work.
It's the same spreadsheet.
It's a different tab.
Of course it is.
One thing before you go back to it. The fp4 number, when it gets quoted, it's because somebody wants the price to look like a marketing decision rather than an engineering one. It isn't. It's precision. You get what you pay for and you pay for what you get and if you can't measure the difference it's because your test doesn't.
That's more or less what the index argument comes down to.
Then you've been saying it wrong.
We've been saying it wrong.
You've been saying it at length.
So the misconception to kill here is that per-token price is the comparison. It isn't. Cost per task is.
Because a model that says everything twice charges you twice.
GPT-6 Luna scores the same as V4.1 Flash and costs a quarter as much per task, at the same nominal token price. That gap is verbosity, and it's invisible on every pricing page.
Which means the number people shop on is the wrong number.
It's the number that's easiest to put in a table. That's why it's the number people shop on.
And the other misconception, which we should say out loud because it's ours.
That DeepSeek is the price-performance leader among Chinese models. It isn't. GLM-5.3-Flash, MiMo-V2.6-Flash and Ling-3.1-flash all beat or match it inside its own band.
Doesn't mean we switch today.
No. It means we test.
Three things to watch, then. Whether censorship survives Western hosting, which is under-studied and probably the most important open question in the whole space. Whether the Western cheap tier holds its price once the introductory windows close, because Gemini 4 Argon goes from two and ten to four and twenty on the first of December and that's a real signal. And whether DeepSeek's Entity List designation, which was reportedly approved and is reportedly being withheld, ever actually lands.
And the cadence itself. Qwen went from 3.5 in February to 3.8 in August. GLM went from 5 to 5.3 over the same period. Anything we've said tonight has a shelf life measured in months.
The one thing that doesn't change is that none of these indexes measure prose. Daniel writes with this pipeline. An A/B test on real scripts is worth more than every score we've read out.
Which is what I'd do before switching anything. Run last month's episodes through three of them and read the output blind.
Thanks to Hilbert Flumingtop, who produces this show and occasionally audits it.
If you enjoyed this one, try episode fifty-seven, From Lawyers in Limousines to Developers in Their PJs; episode three, Safetensors or something else; and episode fourteen, AGI's Crossroads. This has been My Weird Prompts.
If you want to send us a prompt, do it on Telegram at t dot me slash MWP listener bot.
We'll be back soon.
