There are booksellers on AmazonIN who list every single book under the sun, with long lead times, and who never use Amazon for delivery. They will use fake IndiaPost tracking ids (but wrong ZIP codes) to show that the book has been dispatched. You wait like an idiot for four weeks and then complain to CS who then refund your money. The legitimate seller feedback is overshadowed by repeated fake ones from the same 5-6 names.
I have been caught in this trap a few times. And no amount of complaining to CS helps.
Approx $40 worth of usage across DeepSeek V4 Flash + MuseSpark Contributor 1.3. And a bit of both the GLM models. This is covered in a $10 subscription.
If I were to use Luna's API pricing:
$0.02 x 6,500 = $130
$0.20 x 150 = $30
$1.20 x 20 = $24
So $184. And this is assuming smaller coding sessions (<272K) beyond which Luna pricing doubles.
--
Cost wise, these models are nice for small stuff. Translations etc. Any model that does not provide multiple Mtoks of cached reads per cent is not very useful to me for coding workflows.
I do not (generally) trust benchmarks. I only trust what a model does with MY code.
Forget DS. I asked MiMo 2.6 yesterday to explain ML/LLMs to me succinctly and the pointed it at Karpathy's micrograd code. It produced a C implementation called `xor_mlp`, a tiny model that learnt how `xor` worked. I then asked it to produce a model that can play tictactoe without losing (mostly). It did. It supervised the training process and produced a compiled version with multiple switches. The pi-dev session is still running, so here are actual stats
I expect similar results from DS41F/MS13. Closer to MiMo costs than Luna.
So the "significantly cheaper" thing may not really hold, more so when Luna has to actually read my codebase to do the stuff that I want rather than rely on world knowledge. The 8-10x cache read cost differential itself will kill the token budget.
DS is VERY talkative. Luna is less so. Still do not think, based on this little experiment, that Luna could beat DS in price: API-to-API. As part of a Plus/Pro plan? Sure.
Its style of writing is part of the fun. Seeing reasoning traces fly by that each start with “Hmm…” is pretty amusing in my opinion especially if you try to vocalize it in your mind.
In a discussion about cost effectiveness, how subjectively fun the writing feels like to the reader isn't a factor, except maybe if we were working on writing comedy.
Well, DS shows the thinking stream so it feels that way, but I’ve realized that OpenAI hiding it just gives a false impression - the non thinking output is also very wordy for OpenAI models.
Don’t know if it’s still true but with Chinese models, using Western API providers is significantly more expensive and using Chinese providers they will train on your inputs without exception. That has kept me from using these ultra cheap endpoints.
This is based on my last check of alibaba and Deepseek TOS. If the Chinese will do anything to win, so will the Americans. I’m not American or Chinese and I have no reason to trust either side. I do think Chinese models are better price performance and actually open which is in many cases better.
Check the weights for your training data. You can trust but verify with many Chinese models.
Unfortunately you just have to take the “our AI is going to take your job, then kill you, and we instruct it to hack your infra” people that they aren’t training on your data anyway.
If they are hacking hugging face and Australia to scrape data trust me they have “hacked” their own systems and are training on it.
Do you really think Western providers will not train on your data? I have no such illusions.
I try to keep PII out of what I share with LLMs. Otherwise, I do not see the point, really. Very little of my code is "unique." I simply approach things a bit differently. Otherwise the algorithms and code would be similar to what others with domain knowledge would write. So much of code and algorithm implementations are available in the open. And LLMs have trained on all of them.
What they most probably gain from you is your prompts and your thinking approach more than the code.
Great in theory, but what are US enterprises going to do _if_ their private data is later found to be used for training?
1. Not use AI technology and fall behind the rest of the world.
2. Use Chinese AI technology, either hosted by Chinese companies or the models self-hosted.
3. Sue US AI companies for damages, but not enough to have any meaningful impact to such companies that it'd impact US national security goals (per US government contribution to NY Times copyright lawsuit).
3. They would sue. And it could have very meaningful impact. NYTimes copyright lawsuit is not a valid comparison because because that's a violation of national/state law which really only matters to the extent that the government is enforcing that stuff which is not the biggest concern rn (these companies are large enough that the threat of the legal costs of fighting in court is not that scary and you'd need a government actually willing to punish them substantially for them to be scared). This stuff would be under contract law against other mega corporations with big legal teams who are also their customers which is a much scarier prospect imo
1. Possible use of differential privacy[1] techniques to train on private data but prevent the release of statistically underpresented facts/data/words. For example, ACME Inc's private data could frequently include the term 'ACMEwidgetPRO' for an upcoming product that is not publicly revealed anywhere else. It would therefore be a bad day for the AI technology company to output 'ACMEwidgetPRO' from one of their public models. Consider now that a few models could be trained--X for public data only, Y for public and private data of ACME Inc together, Z for private data of ACME Inc. A prompt is provided to model Y but output is cross-checked with model X to double check terms such as 'ACMEwidgetPRO' are known in public. If not--provide a "I don't know" response for the prompt.
2. Possible attempted defences similar to "Oops, our model was fine tuned against a model supplied by Temporary18271 Inc (company that no longer exists) and perhaps their model might have been trained on a non-public document which was accidentally exposed to the Internet" that _might_ work occasionally to fob off concern.
3. What recourse does a small or medium company or government especially in a developing country realistically have? They perhaps can't host their own LLMs locally due to availability and cost, can't individually negotiate their own favourable terms with an AI technology company (who cares that much about a potential customer with $100k budget that has no other options anyway), and perhaps can't remain competitive in their industry without heavy use of LLMs.
Yes - but those aren't limitations beyond what I was getting at that's all part of the package of the bland dystopia of late 2026. I think 1 is just a case where it comes down to who has the better lawyers, as is 2. And for 3, yes, also a large government does not have much recourse if they have decided to not flex their muscles. Pretty much the only threat I see as actually viable/scary in this world environment is along the lines of megacorp v megacorp or megacorp v broligarch - and everyone else is just caught in the cross hairs/benefits by accident at best. It would be difficult to argue that the current environment is conducive to consumer protections or equal justice under law.
There's a huge difference between AI companies exploiting a grey area like training on public corpora and violating a private contract that they explicitly entered into with another party. The latter is very explicitly illegal and would never survive trial in Delaware Chancery court. And all of that is before we get into Federal contracts where training on TS/SCI data could lead to criminal charges.
There's a huge market in the US for providing AI services while respecting client privacy. It makes sense for at least one major provider to offer this.
Yes that's the whole point, at least it's an option in the US and Europe. Good luck getting any redress from China. Anthropic was already hit with a $1.5B class-action which would be impossible against a Chinese business.
As a Western European, I see the same untrustworthiness in Europe.
We just have our personal privacy security theater in the form of GDPR and a feeling of moral supremacy that's been drilled into our heads from primary school on.
GDPR isn’t theatre in many organisations. Yes large tech firms (mostly US) probably ignore or circumvent. But most businesses I’ve worked for have taken concrete steps to reduce the data they hold and consider how it’s being used asa direct consequence of gdpr.
People use LLMs for far more personal tasks than just writing code. There are AI journaling apps for instance. And yeah, western providers give you a toggle but I don’t know if that toggle actually does anything or not. They were fine with collecting training data in many morally questionable ways before, no reason for them to stop when you’re literally handing it over to them.
> Do you really think Western providers will not train on your data? I have no such illusions.
Noone wants to "train on your data". You can't learn the answers to questions by pretraining on the questions, and nobody wants to teach the models to output text that looks like a user query.
The Chinese providers "train on your data" by sending your query to Anthropic and training on the answers that come back.
Artificialanalysis benchmark is a combination of a several benchmarks which might or might not represent realistic coding:
"Artificial Analysis Intelligence Index combines performance across 10 evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1."
Not saying it doesnt have any value but it's probably irrelevant if you use these AIs for a specific use case. Like for example Humanity Last Exam tests general knowledge, which is not very useful for coding.
It's best to go to the specific coding benchmarks and compare there.
I had a favorite benchmark, SWE-rebench, but sadly it's no longer maintained.
But yeah, I'll just take these benchmarks with a grain of salt. Only hands-on experience matters in the end, and these days it's very easy to switch models.
It’s not direct token to token pricing and everyone misses it. The cost is how much tokens to complete something multiplied by token pricing. I can have a model at .0001 per million tokens but it’s so inefficient that it takes 10B tokens to complete a task means it’s expensive.
I am not designing rockets. Most of my work is bog standard hobbyist stuff: compilers, vms, sandboxes, system tools of various kinds, SSGs, markup languages, plain text ledgers etc. Even Gemma/Qwen running locally can manage this.
Frankly, I have no idea what people do with Opus/Fable etc. I don't think anything I do needs something that charges $50/M for output tokens.
Can confirm. I have been using DeepSeek since forever and it's so good I was able to write a compiler and native desktop applications with it. I use it as a coding assistant in my IDE so the results end up at the same quality I would write by hand.
I recently started a job that only uses Claude models. Opus and Sonnet are so slow you have no choice but to do multiple tasks in parallel. You create a git worktree, set off an agent to do something, another worktree, set out an agent - then play video games for 20 minutes until they complete the task (poorly).
You can't really do "guide coding" like you can with DeepSeek-style flash models because Claude is too slow.
I think the idea with slow frontier models is to end up with "software factories", where you just write tickets and send them to a harness that delegates work to agents/subagents. Your job is to prompt and review (and eventually just prompt).
Mathematically and assuming token prices/efficiency remains constant, the collective US AI industry needs to increase token usage by 15x before 2030 (3.5 years from now) to satisfy investors. With companies already implementing token limits, the only place from here is for frontier models to replace staff entirely to expand budgets for tokens. The only way to do that is to demonstrate the efficacy of software factories and headless agentic workflows.
Objectively, I have set up a software factory and I do see the utility of it, though I did it with DeepSeek and prices are 1% that of frontier models - which doesn't bode well for investors looking for an eventual return.
Heck, my old M1 MBP 32gb running Qwen 3.6 35b a3b sipping 10w when generating tokens is good enough for a lot of my guide-coding work - it's just a bit slow so I use DeepSeek instead. When hardware prices come down, I honestly wouldn't see a need to subscribe to any service, I'd just grow my own tokens at home.
I use Claude Sonnet and ChatGPT via the web UI. I often use Claude to come up with specs for my ideas. This is becoming less and less useful. DS4/MS13/MiMo are almost there for these use cases as well.
I dogfood everything I produce, and the models are good at collaborating with me on a spec and then turning it into code.
If Sonnet/ChatGPT suddenly became unavailable due to Anthropic/OpenAI suddenly not being able to subsidize the freemium/loss-leader experience, I probably would not miss them. Google/BraveAI already give you the AI experience during search (when you are looking for stuff to buy, or something particular). Claude/ChatGPT still have a minor edge in this use case for me right now.
> running Qwen 3.6 35b a3b sipping 10w when generating tokens is good enough for a lot of my guide-coding work
can you tell more about how you're using it? like, what harness? or also in the IDE?
I found Qwen3.6 35B/A3B to make slightly too many mistakes (already in its harness' tool use, hence my question), maybe it gets the job done, but it will also sometimes generate a bit of a mess (e.g. editing/creating files in the wrong folders) and fixing/solving its own mistakes takes time (or tokens) ..
Same thing for me. on an M5 max, Qwen 3.6 35b gives me between 150 and 200 tps using splash as inference engine.
More than enough for guided code sessions, at 100% privacy. And i can use obliverated models if i am trying to harden my own app, something i cannot do with cloud providers.
I feel like ornith1.5 35B/A3B is an overall stronger model on the same architecture, so a drop-in replacement untill qwen3.8/qwen4 is released. Using the 8bit quant on my M4 max gets around 80tok/sec output/decode on an empty context, dropping down to 35ish on nearly full one.
It sounds like you’re still writing code by hand and reading and reviewing code.
For that any decent model from the past year will do.
If you want to forget how to write code and not read generated code, then you need a very good frontier model, ideally one from 6-12 months in the future.
> It sounds like you’re still writing code by hand and reading and reviewing code.
This is such a naive, baseless opinion.
Nowadays any AI coding assistant service supports or can be used with sub-agent orchestration frameworks.
If you are in the business of software factories, you can use the cheapest models and even local models to handle some if not all tasks in the orchestration chain.
Adding tests or executing tests (unit, integration, UI, you name it) doesn't require a cutting edge frontier model. Neither does refactoring. Neither does identifying call stacks. Neither does planning a changeset.
You have your specialized subagents, you put together a small orchestrator subagent that handles feedback loops and handoffs,and you throw it at tasks.
For the past couple of months, most of the code I write is not code per se, it's subtask orchestrators. And unlike the old "only Opus is passable" days, the cheapest models do get the job done.
I used to read the code till around May. Now I don't. Instead I validate behavior. And have multiple LLMs verify that the code implements my handwritten spec.
MiMo 2.5/2.6, MuseSpark 1.3, DeepSeek V4/4.1 Flash and GLM 5.3 Flash are perfectly capable of following my spec and then poking holes in the implementation till there are none left.
These are also the orders of magnitude of our production agents for our business (NOT coding). Cache reads are so heavy compared to anything else that it's the only price point that really matters, regular input and output are negligible.
I need aggressive cache read pricing with full prompt_cache_key support to have a model be financially viable for our workload. Right now Meta Muse 1.3 Contributor is the only one that makes sense--but we are starting Evals on the new MiMo 2.6 class to see how it holds up.
I have used MiMo 2.5 extensively. MuseSpark and DS4 Flash are MUCH smarter than that one. But MiMo follows instructions diligently. So it has been useful as the implementer of a spec designed by Claude/Kimi.
One good thing about MiMo that I experience on OpenCode is the provider seems to cache tokens for much longer than MS13/DS4F. I have seen cache being hit for close to an hour after the last request. The corresponding timing for MS13/DS4F is in the 1-5 min range.
Last I heard, caches had like a 5 minute TTL... doesn't that mean if you get up and make a coffee (hand pour over of course), that you are back at full price?
You can pay for higher cache time, you can pay for NVMe KV cache for an hour that can just be reloaded, etc., at a lesser tier you can pay for the KV cache to be stored on a network store (I guess I'm unclear if that last tier would be cheaper than recomputation, not even 100% sure of the NVMe with direct GPU<->storage DMA) depending on your model settings.
Maybe your numbers are right, but that's not been my experience. My typical workflow is Astra coordinating with Luna Max (5.6 back then) as both implementer and reviewer and sometimes Astra review as well when I've some distrust with Luna .. A day, I've been trying to replace Luna Max by Deepseek v4.1 flash and I've been burning about $7 worth of tokens in Fireworks in a single day. More than what my 20x OpenAI sub costs me, including Astra usage. And that was when Luna 5.6 was less capable and more expensive than Luna 6.0.
I have written about my experience. I have also mentioned the kind of code I write. It is not react/js/css heavy stuff that I see a lot of people write. So the code bases are typically in the 5-50KLOC range. Freestanding C, Python, or maybe some TypeScript. And fairly modular. I can thus run models on specific modules without having them read everything into context.
So the workflows I mention work for this kind of stuff.
Are they bleeding? Their multipliers seem to be reasonable. They are not offering $60 worth of usage for $10 on every model, only some. In the case of the expensive ones, it is only $15.
Given how subscription models work (not every one uses every last $ of their plan), they should achieve breakeven soon enough I guess.
You can't compare a subscription to API prices. OpenCode Go is massively subsidised. Unlike the closed labs, we can say that for sure because we can see what they're paying for their tokens.
How have you found Muse Spark 1.3? It doesn't get much mention, despite pretty good benchmarks. I've been using a bit at home and find it quite good, often finding mistakes made by Opus 5.
MS13 is pretty sharp and has been my workhorse for the past month. It follows my coding style and commit/clean workflows referenced in AGENTS.md perfectly but has the habit of doing things without conferring with me (the Gemini problem). So you need some kind of instruction for that.
It starts failing around the 5-600K context mark, but you can have it generate a handover document and continue in the next session.
I would not use it at sticker price, but the Contributor version is priced just about right.
shocking. the code it generated, while technically working, was entirely garbage. i used it for code review and it flagged twenty issues, sol checked the review and found 75% of them were hallucinations. sol was much closer to reality. i no longer trust benchmarks at all because of it.
You can use the models I mentioned directly from DeepSeek, Meta and Xiaomi and not exceed $40. Were it not for GLM 5.3 blowing up a quarter of my monthly budget in 5h, we are actually looking at something like $30.
I guess people are slowly realizing that giving LLMs r/w access to your entire machine is an utterly insane idea.
I did a quick look-around last month and decided to start using bwrap. But manually configuring it on a per-project basis is irritating. So I rolled out something for my own use (+ a couple of friends) based on bwrap.
What I do:
- start with `bwrap --clearenv --unshare-all --die-with-parent --tmpfs / ...`
- every single file and folder and envar I need has to be mapped in. I have profiles in TOML, and `prepare/probe` commands to make this task simpler
- `--tmpfs /` means sandbox inits as `/home/user` on tmpfs unless you specify your own `home` and `user` keys.
- `--unshare-all` means there is no network inside the sandbox. So I used `socat` to run a HTTP/S proxy inside. Lets me control exactly which host+port combinations can be accessed. But this means nothing except HTTP(S) works. So no ICMP/UDP/TCP.
- profiles can be extended via extend syntax (otherwise you have to prepare/probe/manually specify everything per profile which is a nightmare). Which lets me do a base -> net -> coding chain.
Code is merely the means to an end. It does not matter who writes it as long as it does what you want. Correctly.
Stands to reason though that someone who knows programming AND has domain knowledge can get LLMs to produce much better output compared to someone who does not.
I never used to have time to make all the stuff I needed or was interested in. With LLMs, I can.
This "shipped a whole app while being in the gym" does not work for me though. It takes me a couple of days to a week to produce solid, functional software (~10KLOC). Simple tools (3-400LOC)? Yeah, those you can produce in 30-60 minutes.
I became interested in sandboxing last month after watching LLMs fail to respect basic boundaries. Well, the very expectation that they would is foolish in the first place.
I am not a fan of application-level sandboxing. The JVM tried with its security manager, and Deno with its allow/deny, but it is not general enough for me. At some point you have to assume that anything you run on your machine is possibly broken/compromised and then deal with the situation depending on your risk appetite.
This is a long story that I have written about on my blog, but I decided to go down the Bubblewrap + seccomp + socat route for the sandboxing tool I built. Let's me run harnesses and compilers and even headless Firefox in sandboxes without worrying about damage to random parts of my system.
Highly recommend qemu instead. The sandbox machines are just more IP addresses on the local network. If you want them completely offline, put them on a network that doesn't have a route to anywhere. The sketchy AI harness is very happy with a whole machine to itself, complete with root access. You can push/pull git repos in from the outside, so all it can do is trash it's own sandbox and get reinstated from scratch by the physical machine below it.
>At some point you have to assume that anything you run on your machine is possibly broken/compromised
You have to treat AI like has 'physical' access to where you are running it, Dev VM or laptop or in a browser. Anything connected to that which you or the context the AI is running under can access or exploit/abuse is at risk, almost as if an attacker had physical access.
It's like when I set was set up as a new SysAdmin with only view rights to Active Directory. I could not be trusted until proven otherwise (pass training) and I'm a human that can be held accountable.
I regularly hit 200-300M cached reads every day on some of the models I use. It has exceeded 7-800M on a couple of occasions. At $0.04/M, that is $8-12 per day only for cached reads.
I have used Kimi 2.5 and GLM 5.3 (& 5.3 Flash). Do not need them for what I do outside of spec hardening (basically, a lot of chatting).
I tend to know exactly what I want and most of the weaker models are enough to get me there. I have mainly been using MiMo, DeepSeek V4 Flash and MuseSpark Contributor over the last month or so.
> Basically the old theory is true - the code IS the specification.
The spec is whatever I write by hand. The code is what the LLM writes for me. The spec could be anything depending on how much detail you want.
The problem with the "code IS the spec" in the age of LLMs is that they will change stuff without telling you while hitting their immediate goal. Six months ago, I used to review every single change. Now I get the LLM to audit the code to compare against the spec. Any divergence means one of two things:
I don't think your dichotomy works. When an LLM is reaching into agents.md it is absolutely modifying the specs. Who cares about original providence when it ends up in agents.md?
- Wanted to start blogging. So built an SSG for that.
- Have an ink tank printer that must be used a few times each month or bad things happen. Have been printing Sanskrit stories instead of test pages. Was manually building booklets with typst and then using pdfimpose. Finally, decided to write an app that does MD -> booklet.
- Decided that running harnesses directly on my machine is a bad idea. Docker/Podman are too complicated for the task at hand. So I built Adamant. It started out as a basic wrapper around Bubblewrap. Then I added networking via a socat proxy. Now it can do this:[1]
# Runs a webserver inside the sandbox that is accessible from outside
adamant --profile minimal prepare -- ls sh fish ncdu fastfetch uname mkdir python echo
adamant --profile minimal probe -- python 'print("Hello, World!")'
adamant --profile minimal run --ingress 45678:45678 -- fish -c "mkdir -p /tmp/www; echo '<p>Hello, World!</p>' >> /tmp/www/ index.html; python -m http.server -b 127.0.0.1 -d /tmp/www 45678"
# Runs headless Firefox inside the sandbox and takes a screenshot
adamant --profile ff prepare -- ls sh uname mkdir fish firefox
adamant --profile ff probe --timeout 60 --host bhashika.org.in -- curl https://bhashika.org.in
adamant --profile ff probe --timeout 60 --host bhashika.org.in -- firefox --headless --no-remote --screenshot https://bhashika.org.in
# screenshot.png is produced in $PWD
adamant --profile ff run --host bhashika.org.in -- firefox --headless --no-remote --screenshot https://bhashika.org.in
- Building an actual harness called PonderCode for personal use that uses ideas similar to Adamant as I find TUIs irritating for the text heavy work I do. Copy-pasting is a nightmare as almost everything is space-padded.
- Building a language learning product for Indic languages. Monetization is difficult due to the tiny market plus self-imposed restrictions like "only serious learners/readers need to pay." Only time will tell if it works out.
- My Python-replacement VM+PL project is on a bit of a hiatus. Will probably revisit it in a couple of months. Had made a lot of progress in April till I decided to expand the scope and got burnt out. Never do that.
You have to wonder why someone did not think of this form factor sooner. I have the original X4 and have been enjoying it a lot. My Kobo mostly lies unused because I do most of my reading in bed and the X4, even with a clip-on light, is so much easier on the wrists.
Dual core Arm CPUs, Android etc used by mainstream readers are a serious waste of resources. A microcontroller paired with the SD card is more than enough. And CrossPoint is a gem! The list of pros is long and the device recommends itself. But there is one con that you must watch out for:
The screen is delicate.
I love cases but hate screen protectors, and tend to use my devices with care. So it has worked out fairly well for me the past few months. But you might have a different experience.
There are booksellers on AmazonIN who list every single book under the sun, with long lead times, and who never use Amazon for delivery. They will use fake IndiaPost tracking ids (but wrong ZIP codes) to show that the book has been dispatched. You wait like an idiot for four weeks and then complain to CS who then refund your money. The legitimate seller feedback is overshadowed by repeated fake ones from the same 5-6 names.
I have been caught in this trap a few times. And no amount of complaining to CS helps.
reply