I tried the cheapest provider on openrouter and burned through $50 in a few days. Quality was ok, seems slightly above Luna quality perhaps? But that $50 is 1/4 of my codex subscription where I could have burned that many tokens or more using Astra within my weekly reset.

This won’t last forever but as long as the frontier labs are subsidizing this heavily the open models won’t matter.

I don't think those subscriptions nave negative contribution margins, either. I think we're seeing a lot of price discrimination by the big labs, and huge margins on their frontier models. The fact that they have been cutting prices to their second-biggest tier of models (Opus/Sol).

Open models catching up and collapsing these margins would worry me if I were a shareholder in the big labs, but as a user, I really doubt that the western labs have bigger environmental impact just because they have higher API costs, I think they have a ton of efficiencies they aren't sharing with customers yet because demand is so high.

Have you guys see how aggressive is the push for enterprise use by both OpenAI and Anthropic? I had friend from a non-tech industry in Asia telling me that their company was offered free trial of the enterprise version of Claude, with trainings and such.

On the other hand, DS and Z.ai, have zero to none marketing outside China. There is friction to use DS/GlM models and the ZDR is unclear, so most enterprise that has heavy AI usage hasn't move over yet. They would rather spent $200 for the peace of mind than to take the risk of being slam as a national traitor down the road (which again is another form of marketing by Big AI, trying to frame Chinese models as thiefs).

So, I don't think they are not freaking out, it's just that they are addressing different market segments and reacting to the situation differently.

DeepSeek is horrible at grilling sessions (the /grill* skills to make technical decisions). It doesn't know how to explain things. Maybe the skill could be adjusted. It also doesn't come up with as good solutions as Opus/Sol.

What I use it for is

  * the orchestator of my coding workflows
  * the tester/verifier of code changes
  * the sub agent that explores code or does web searches
  * putting together code base research reports

They are all working together in Pi using the extension @tintinweb/pi-subagents where my workflow skill is calling different subagents that use different models.

Luna is cost competitive, but doesn't score as well on intelligence. I do need the intelligence for most of what I use it for, so I am not motivated to use Luna. Haiku also doesn't seem like a competitive price/performance mix.

Once its gets juicier, we let flash launch specialized subagents with specific models. GLM-5.3 for coding or Kimi K.3 for research and critique.

But as a main driver. I love flash. And it brought our bill down by A LOT :D

These open models still did not beat February's Mythos / Fable 5.

DeepSeek 4.1 Flash is behind GPT 5.6 Sol, and that one is left in the dust by the excellent Opus 5.5.

Rumors say Anthropic is holding in reserve the big improvement, Fable 5.5, for the IPO.

It's plausible that open models are 6 - 12 months behind, and there is no "good enough". As long as progress doesn't slow down, leading labs have nothing to fear.

I was previously using GLM-5.3 as the orchestrator, after switching to DS anecdotally there was an unnacceptable quality loss, mostly around not taking all the relevant context into account when making decisions, pulling new design out of thin air without discussion too often, and being way too wordy and rambly in documentation despite prompting to avoid it. There's a lot of docs, rulings, core concepts, design philosophy to uphold and DS was just not cutting it.

However, it's perfectly capable of being the sole agent for all of my well specced implementation tasks. I've gone back to GLM as the orchestrator.

My company is not willing to pay $50/engineer/month for a cheaper version that is nearly as good. My company is also not willing to pay any amount for a product produced by China, even if it is hosted in the United States.

I am a big ChatGPT fan, all our team has ChatGPT Subs, but the TPS across all models including luna is just so damn slow.

Commandcode giving 60$ worth of Deepseek for 10$ is just genuinely goat.

And it never says no for cyber tasks so that's a big win

The other reason is more interesting. Maybe the frontier providers think that price performance is irrelevant in light of very powerful frontier models that can start the RSI loop and or a huge displacement of work and a winner take all economic situation. After all if frontier providers earn everyone's money then you won't have any money to spend on any model 100x cheaper or not.

However, I can easily burn ~$5/day if I use it as my “worker” (still using opus for planning and review) so call it ~$150/mo.

I could, if I were so inclined, get a second Claude Sub and have even more headroom (though I’m able to stay under my limits most of the time with my current setup). Also Claude gives me Artifacts, Web Search, and now even some API Credits.

I have no doubt the future is open weights and I can’t wait, literally, I can’t wait for them to catch up on intelligence or for hardware to run a decent model to be within my grasp. But until that comes to pass, I’ll keep using Anthropic.

see https://artificialanalysis.ai/models/releases/comparisons?co...

I realized that mistake and guided DeepSeek where it should be.

Next I fired Fabble 5.5 set to high to check if the hype is real about Fabble. It exhausted 89% of quota and came up with NOTHING that DeepSeek hadn't flagged itself already in its notes.

- Models are still very jagged. A superior model (e.g. Opus 5.5, Astra) may not be materially better in some tasks, but usually is for completely new modalities: computer use, game development, etc. People will always prefer using less jagged intelligence, because it allows them to do so much more

- Distilling intelligence from frontier models means that unless chinese labs manage to replicate the training regimes of OpenAI & Anthropic, they are always going to be behind a few months. That's a feature of training by distillation.

- China doesn't yet have the compute to compete at the frontier. Because they don't have access to top tier chips, their GWs are not equal to US GWs. Until they close the hardware gap, I think they will always be focused on competing on efficiency, as opposed to intelligence.

- OpenAI and Anthropic subscriptions are valuable of quality of intelligence + generous compute allowances

I’m convinced that I’ll have good enough inference on my laptop at reasonable speeds within the next year.

It's way faster than Opus or any of the GPT models.

I have a coding harness which is opencode plus a few skills relevant to my workflow. Deepseek 4.1 Flash does very well in this environment. I haven't noticed much difference quality wise compared to Opus 5, which I use in my day job as my employer pays for it (although I'm considering using DeepSeek here too given how cheap it is).

With my OpenCode Go sub of $10/month, DeepSeek is basically unlimited

My OpenCode Go monthly window was scheduled to reset this morning. It was sitting at 22% used despite me using DeepSeek V4.1 Flash heavily as my implementation agent the past couple weeks (I use gpt-6.1-sol high for planning/orchestration).

I had 1.5 hours left so I fired up first 10, then 20, and finally 50 concurrent subagents all working on reverse engineering C code from an old PC game. They found over 100 new functions.

This is the first workload I've found that could make a dent in my sub. It got my 5 hour window to 85% used, but sadly my monthly was still only at about 35% when it reset. So that cost maybe $2.

That said, it's my best understanding that these american companies aren't profitable and will eventually raise rates (the old uber trick) so I'm keeping myself ready to switch when that day comes.

VRAM & Memory Requirements by Precision

• FP16 (Full Precision): Requires ~1,664 GB of VRAM (e.g., an 8x B300 288GB cluster).

• INT8 Quantization: Requires ~832 GB of VRAM (e.g., 8x H200 141GB).

• INT4 Quantization: Requires ~416 GB of VRAM (e.g., 8x A100 80GB)

VRAM aint cheap, Sam Altman ruined the cost of memory, Nvidia doesnt make enough consumer GPUs letting the market go insane over them, I still have friends on 1070s or 1070 TIs because GPUs have been severely overpriced for too long. I remember when a gaming PC was only $1000.

Even so why would anyone not sleep on a model they cannot run?

Because I don't have the time and money to do an extensive benchmark of all major LLM, so when I had to select a LLM for my usage (which was not coding at first), I went to the most used one, chatgpt, because I knew if would be one of the best at the task.

I suspect it's the case for many if not most people.

I'm also the CEO of a new company - LunaRoute - so if you want private, fixed cost, we run our own server in the US type of DeepSeek 4.1 Flash check out https://www.lunaroute.com

If you want a trial, hit the contact us and write that you saw this post.

We also have GLM 5.3 (with vision!) and GLM 5.3 Flash - all included.

Opus 5.5: TIME 9.3m COST / $1.99 / SCORE 99/100 https://jonclegg.github.io/pacman-bakeoff/#claude-opus-5-5

Deepseek 4.1 Flash: TIME 2.8m / COST $1.89 / SCORE 72/100 https://jonclegg.github.io/pacman-bakeoff/dev/#deepseek-v4.1...

That said, has anyone else found DS models to be unpolished? They seem to "lose their mind" a lot more often than Claude/GPT. I have tried all of the top open source models that came out over the past ~4 months or so and the GLM models (5.2, 5.3, and 5.3-Flash) have been much more usable for me. They feel like Opus but X months ago, DS feels like something else.

A cool finding we had: we tried a lot of different models - Deepseek v4.1 flash, Qwen3.8, Opus 5.5, 6.1 Sol - while testing our "agent experience" for setting up and using qURL (OSS temporary tunnels; similar to ngrok but faster to use / more private) and found that even though Deepseek and Qwen were less intuitively smart about figuring out how to use qURL, the number of tokens they spent digging through docs and figuring things out was barely 1.5x that of Claude/GPT.

This would indicate that even though these models aren't the "frontier", they're still quite capable and WAYYY cheaper (which, from an avg dev's perspecitve, is still worth it). For most consumer-grade tasks, opus is going to be overkill anyway, so eventually, I think we will really see the consumer-enterprise AI model choices split.

China is going to eat their lunch

No doubt about it, that's why their push for international regulation to the levels of nuclear inspections using the narrative of annihilation and apocalypse

If it's underpriced, it's a loss leader to sell the other models, so it actually can't be too good.

I really put these things through their paces because I use them to review and work with new abstract game rules and models, so they're always flying blind. Luna misses the obvious (and more importantly, the clearly explained) consistently. My second prompt is listing all of the points in its first response, and saying "No, it doesn't work like that." The third prompt is picking out the two or three suggestions it made after correcting itself on all of the original points and saying "That's how it already works." The fourth prompt is "Now that we're done going over the rules, can we start?"

I actually feel like 5.6 Luna seemed better.

Sure, they stole Claude's training, and Anthropic stole it from other people. I'm not getting into the whole who-owns-whose-data debate, because most developers aren't thinking like that. They're just trying to get the most bang for their buck.

Until they get laid off and suddenly discover their moral compass.

DeepSeek 4.1 Flash is really is 10x cheaper than Opus, but it's in the same ballpark as Luna & Haiku in terms of quality, cost, and speed for data analysis. (Luna & Haiku actually perform marginally better, cheaper, and faster).

Here's a breakdown: https://plotly.com/blog/deepseek-v4-1-flash-data-analysis-be...

Note that this benchmark is for data analysis, not general software development. So the tasks are shorter and more constrained.

But ever since I've switched to one of the $100-tier subs, I can see why a lot of the people on it don't really discuss the open models often. I'd still use it especially when it comes to sensitive inputs, but for most work, what you get on OpenAI or Anthropic is really more than enough.

It really got even better when they also made their cheaper models up to par if not better than the open models.

I do think the crowd for open models are out there, especially when you see trillions of tokens running for them on OpenCode or OpenRouter leaderboards.

So the industry is responding, where it matters. Which is on heavy API usage, not coding subs.

I'm well aware that there's nearly infinite opportunities to yak shave "perfect" OpenRouter setups and some people appear to enjoy bouncing from IDE to IDE as though change costs aren't a thing, but I discovered that I genuinely like Cursor and at least right now it's insanely subsidized by Auto clearly defaulting to whatever Grok's most powerful model is.

I dropped my $200/month subscription to $20/month and stick to Auto for all but really important Plan tasks, and I have basically zero chance of using up my monthly credits even using it 6-10 hours some days.

This is like comparing big pharma with generic manufacturers who can skip the R&D. While these drugs are not 1 to 1 copies, we are still comparing apples to apples, but with a 90% price cut.

Weird! It's almost like distillation is bad for the long-term growth of the industry, just like generic manufacturers would be if they could release the generic versions of drugs 2 weeks after the original R&D completes.

Very strange sentence to include in an article after saying "I don't care at all about distillation" up at the top. Can the author not hear themselves?

I've been focusing on deep research related tasks for biotch and life science applications. The problem with this sort of task is that we need subagents to reason through multiple (potentially 100s or more) chains of knowledge/concept/evidence, so the token usage really explodes as complexity of the task and data expands. A typical task can cost me nearly a $1k overnight...

I've been testing out GLM5.3, but now I'm really tempted to try to Deepseek 4.1 flash too. Any chance you've benchmarked / compared the two?

Luna is on par in benchmarks and my personal experience is Luna is better for what I do, and Luna is cheaper.

Comparing Deepseek 4.1 flash to Opus is just ludicrous.

https://artificialanalysis.ai/models/releases/comparisons?co...

Don't really understand people who say DS4 or 4.1 have frontier level performance. Anyone who has used it will tell you that it's a hallucination factory. The only thing it has going for it is deepseek's unique infrastructure that allows better cache retention, but the cost savings from that obviously come nowhere near how much subsidized usage you get out of even a $20 subscription with openai or anthropic.

And it's why people like Hillary Clinton have been trotted out to talk about the dangers of open weights models -- I mean, does she even know what that phrase means? (I know HRC is a controversial figure and I'm not bringing her up for that purpose; I just note that she and other prominent retired politicians are now doing the circuit on Anthropic's behalf.)

It's good, and you can do most work with this. For complex software implementation you need to split your runs into various phases, build in verification, and use subagents so that work gets another audit and repair pass from the lead agent. You can do pretty much everything then. Frontier models can do without compelx workflows, that's the difference.

This article might have sat as draft for a few months. With the current deepseek pricing, the same membership lasts me a week at most even though I am using the free middle too and have Gemini pro+ultra.

It used to be that you could dump pocket change into the deepseek api and forget about it. Nowadays it'll make you notice real fast as the dollars pile up.

But the other reason is because most people in technology are just as lazy as consumers and the friction between getting onto DeepSeek vs OpenAI or Anthropic is just a tad too much and people go with what's convenient.

It did correctly make waves when it launched, but was quickly eclipsed by the deluge of american model releases, especially those competing on cost.

My AI pilled clients who were early AI adopters are already there and they are looking for solutions to spend less.

And to the self-hosters out there, the economics of 4.1 Flash mean self-hosting is not worth it. If saving money is your goal, you will never recoup the costs.

Self-hosting is, for most enterprises, absolutely not about economics but rather about data confidentiality.

And in that regard, yes, the open-source weight models, especially the chinese ones will eat the fat closed US model's lunch big time.

The fight between Anthropic and OpenAI is just about who leads the duopoly. Chinese open weight models are threatening a duopoly in its entirety. That's why Ant and OAI are begging for regulation. A regulation that will hit Chinese models way harder than US models, finally giving them a moat.

But I run it locally. When I tried it on open router when my gpus were busy I must have gotten routed to some crappy providers, because it was pretty bad.

For me Glm and DeepSeek are nowhere near this Qwen model. I tried various harnesses including omp which I heard supposedly "makes DeepSeek 20 points better". The difference was in the noise (1 point). I run a bunch of benchmarks Terminal World 40, terminal bench 2.1,SWE Pro, GSO. Before those 3 there was no open model that scored more than 1 point on my subset of GSO. Glm scored 5, DeepSeek 3, but Qwen did 17 and opus 19.

Qwen is a small model so it fails on factual recall. But if you give it most of the info it needs it us amazing.

It's like asking a math PhD to organize the files on your desktop.

I actually do use an agent harness to organize files on my desktop. They make a great fuzzy file renamer. Point it at a directory of disorganized files with names all over the place, give the directory layout and file name pattern you want it to have and it makes it happen.

Is OpenAI coming in $20B under a sign of "freaking out"?

It costs pennies and you got really great output.

The author is spot on.

Everyone does optimization of model serving because it's good for every player in there.

(Also the water consumption thing is not a real issue.)

I switched from DeepSeek 4.1 flash about 2 weeks ago for my Hermes sysadmin/coding agents and I am seeing better intelligence and lower overall spend.

https://artificialanalysis.ai/models/mimo-v2-6-pro

I've benchmarked, rigorously, deepseek-v4-flash for programming and personal use, and it is definitely less smart than Qwen3.8-flash-next (which in turn, is not terribly smart).

Local models are also really slow, unless one spends insane amounts of money.

Having said that, Qwen3.8-flash-next is an impressive evolution; it reaches the small versions of the frontier models (like Sonnet) - but again, it's massively slower and not 100% reliable (including: stability).

Just try Opus 5.5 reminds me how Opus 4.5/4.6 astonishes me. Completely different, and GLM-5.3/Kimi3/DS-4.1 are still like Opus4.8 levels.

It stops the constant context switching.

I personally used it for secondary research agents and classification . Now classification part is gone .

Isn't Mimo 2.6 pro smarter and cheaper? Haiku 5.5 is smarter and cheaper. Luna is basically as smart and much cheaper.

For coding, I rather spend 10x more than have even 1 bug but I'm only spending 2x 3x more if you count subscription cost.

This is a killer use case for something like customer support though.

And yes, Opus is enough smarter than DSF that it's worth the extra steps. This ranking is from live tickets, no contamination: https://slopcop.com/power-ranking

Would be cool if they added it.

Chasing the latest and greatest is silly

Said every week by someone who would never go back to using the model they had 6 months ago

And as long as I pay as little for claude opus 5.5 i do right now, i'm using it.

But yes i'm glad that we have alternatives.

So why aren't the frontier labs freaking out right now?

They are. Isn’t this why they’re trying to get regulatory capture?

shrank the KV cache by roughly 437X

Can't you just say "shrank to 1/437th the size"? It's not that hard.

We just launched it on our platform today (Mixlayer, https://mixlayer.com), promo code LAUNCH-DSV41F gets you some free credits if anyone wants to check it out.

And since then, there has been so many articles that made it to HN front page, and all of them didn't get it. They just went on and on about tokens generation. Most HN commenters didn't get it either, find-in-page for "cach" typically yield 2~3 responses. If I had a dime for every time this happened, I could have .. paid for 1B cached input tokens?

Anyway, DeepSeek still has to come up with a frontier model, and they almost did it with DSV4 Pro 0813, which is just slightly below GLM 5.3, but 30x cheaper. Unfortunately, the massive price hike happened just 3 days later.

DSV4.1 Flash is good, but not quite the same level. Much easy to self-host though, especially for serving a team of developers. Let's see what the next one can do.

if you have a legitimate coding application, it isn't very good. if you have some kind of inauthentic activity, which could be what it is trained for for all sorts of reasons...

P.S. That's not to mean there aren't dangers regarding AI. I just don't trust the people making money from selling AI to manage those risks ethically instead of "protecting" us from those risks like pimps but with suits and good manners.

For now, I think everyone is still using Anthropic and OpenAI because if you use a subscription you pay 1/40–1/50 of the API prices, and the models are good when they don’t nerf them, and they are also way cheaper than open models’ API prices.

The interesting thing will happen when they pull the plug and become economically smarter to stop using them. I regularly try alternatives to avoid being locked in and found GLM-5.3 as an orchestrator and GLM5.3 Flash + OMP and DeepSeek Flash as advisor to be able to get jobs done just fine. Space Bunny too was pretty great, which was probably MiniMax’s new model.

I think they are using an Uber like strategy but without the network effects that justify losing money for so long

Every company in China has to abide by the 2017 National Intelligence Law: "supporting, assisting and cooperating" with state intelligence work, and keeping that cooperation secret. They have to hand prior knowledge of vulnerabilities to the state before public disclosure, in order that the state always has an exploit pipeline. No matter how ethical the company staff may be, they'll always be bound by law into being an arm of the Communist Party.

Agentic access is infinitely worse than chatbots. They can exfiltrate silently, target users, plant persistent malware, and be run by third parties through you.

You don't have to be a tin foil hat sinophobe to understand the dangers of being a Westerner granting CCP access to your files and network.

ByteDance staff accessed US journalists' TikTok data to hunt leakers (admitted in 2022). Volt Typhoon and Salt Typhoon were state operations pre-positioned in Western infrastructure and telecoms. Regulators in Italy and South Korea blocked DeepSeek's app over data handling, and analysts found its web client sending data to a China Mobile domain.

Please don't sacrifice security for cost and convenience.

It’s disgustingly good value. I find it capable of doing anything I want.

Obviously can’t use it at work, but for home projects it’s awesome.

I dont get why people says D4.1 flash is good

However, the quality of the open-source Chinese models is terrible. They game or fake their benchmarks because in real-world usage they suck.

Haiku 5.5 is 23% cheaper with a 4 point intelligence lead.

I'm on subscription usage so I can't compare Flash 4.1 to them directly but the OP has his head up his ass if he thinks Opus 5.5 is the best point of comparison. Why is anyone using Opus if the new Haiku is indistinguishable /s

Just absolutely terrible post, admits to using Opus for review but claims its intelligence isn't needed, why aren't you using Haiku or Sonnet then?