I use DeepSeek every day (via VSCode Insiders and Zed Editor). It's very affordable and, while it's slightly behind Claude (not sure how far behind Fable), it suits my working style well. I'm not using unsupervised multi-agent workflows and don't need a library of skills files - I'm writing most of the code and leaning on AI to help with mundane tasks - like;
- generating types for APIs
- generating boilerplate based on existing code
- improving existing code (adding error handling, timeouts, things like that)
- Writing SQL repository boilerplate / queries
- Creating implementations against hand written tests
- Helping me understand and implement APIs from third party libraries
- Writing documentation
I've spent like $2 in the last month and have used over 100 million tokens.
It's doubled my productivity and unlocked work that I could not have done before.
As an Australian, I'm not sure that I care about the safety of my data when it comes to LLMs. US companies already stole scores of data to train their models on and it's hard to imagine they suddenly grew some integrity. I'll care when regulators step in, until then it's out of my control so I'll just use the best price-to-productivity product available.
In terms of price-to-productivity, nothing will beat DeepSeek right now. For what you have described, all of the existing frontier models will perform well (probably about the same even).
If you expanded the list to very hard research tasks, Fable was so far ahead of the others that it doesn't even deserve debate. If you are a researcher doing something involving scientific computation or mathematics that wasn't rejected by the guard rails, and you were using Fable, that week was probably your most productive week ever. A couple of my PhD students effectively finished their current projects in that period by getting Fable to chew on it for 30 hours straight (not sure how I feel about that).
This was my experience too. I asked fable to implement a (quite complex, novel) CRDT engine. It did fantastic work for the 3 days I had access to it. The spec it wrote is exactly what I want. It used prototypes and examples to figure out some hard problems and answer a lot of complex design questions. Claude Opus has been lumbering along over the last week trying to turn what it did into a useful library. As far as I can tell, by trying to vibe together pieces of work fable did without understanding it properly. Opus makes mistakes constantly, misunderstands the spec, and it makes terrible engineering decisions. Earlier today it claimed it proved something was impossible. I asked it to think that through and it immediately backtracked and apologized. Was it right then, or is it right now? Claude has no clue.
I'm kicking myself for not using Fable more while I had it. Now that I've used fable, I'm not sure I even want opus any more. It might be more efficient in the medium term to just program everything myself until I have access to a similarly capable model.
I feel like Deepseek, opus, etc are only good at problems that have already been solved 100 times on github. They're like the iPhone 3G. Its exciting they exist at all. But subsequent versions make them seem like cheap junk.
I had a similar experience. For complex maths in my 3d engine around IK, made more progress in those 3 days than I would’ve done in weeks.
I recently discovered though that ChatGPT 5.5 Pro is almost as smart but only available to me via the ChatGPT app. I’ve been having it read my code and having Opus 4.8 use the ChatGPT app to collaborate. It’s a step down but its a temporary stopgap for complex problems.
It's been pretty cool (and a bit scary) to hear similar experiences from many other researchers and developers just how impressive Fable was for their particular workflow and the productivity it seemed to unlock. Personally, I have found that I can't let any of the other AI agents touch the code that Fable produced for me, since they consistently fail to understand the very delicate choices that it made to ensure optimal performance.
I'm an engine programmer in video games on a big project - recently I'm mostly just fixing bugs. And Fable has been absolutely phenomenal in those 2 days we had it lol. It was genuinely the first time I didn't really have to reproduce a bug to fix it - Fable was able to really understand the code and interconnections between the systems to reason why it might have happened and fix it. Opus really struggles with that in my experience, or it munches code for 2 hours only to come up with a completely wrong explanation, like, wrong if you think about it for more than 2 seconds. TBF I only had 2 days to play with fable, but it was incredible in that time - can't wait to have it back.
Something I wonder about in this domain is if this is the fault of the language.
Game development is usually in C++ or C#. With C++, bugs are a nightmare to find. C# less so, but there are still memory leaks possible.
I've been writing Rust for a few years now and it has been absolutely phenomenal. I get the performance of C++ and the only thing I think about when developing is the logic itself.
There aren't any mature game engines that use Rust (Bevy is rapidly improving, but it's no Unreal/Unity) - though you can compile Rust to a dll so it can be consumed by Unreal/Unity.
Due to the language constraints of Rust, I have found that LLMs need to work _way_ less to figure out bugs and render out code. The compiler gives very specific error messages and if it compiles, it works.
When we started getting access to Claude last year, it was barely competent at C++. Like someone who read a book about C++ but never actually wrote any, or had to debug any issue. It would confidently declare things which any C++ programmer could tell you were bollocks straight away.
But....now it's able to reason about things like memory management, thread safety and second order effects when multithreading very very very well. It's honestly incredible how well it can understand C++ code and figure things out. Opus is already pretty good(although far from perfect) but Fable was just something else. I could give it a description like "It was reported that only sometimes(1/20 repro) when the user opens the inventory and closes it, they can't move the camera afterwards" and it was able to find out both the inventory camera control and input code and pinpoint the exact race condition that could lead to this situation.
To train people to become researchers. If the definition of what a researcher is changes over time (due to AI or otherwise), then the nature of the PhD will also change over time.
I would say to push frontier of human knowledge forwards. Its not a high school anymore where you learn to learn. PhD is already a place where you should deliver value. (By "value" i don't mean commercial product)
> I would say to push frontier of human knowledge forwards.
Not sure if this is what you meant, but to be clear: pushing human knowledge forward is a practical prerequisite of a PhD (given it's pretty hard to convince people you're adequately trained to do such a thing without actually demonstrating it), but it's not the value of the PhD itself. The value is producing a researcher -- that is, someone who has the skills to continue accomplishing this in the future.
The difference is that if you happen to expand human knowledge by a stroke of dumb luck (or smart AI...) without actually having acquired skills to continue doing so in the future, then you're not really earning a PhD.
Depends on the field. There aren't really all that many research tasks that can be completed in a week. One of my favorite stories from the past few decades was the finding of accelerating expansion of space. It took three decades of scheduling telescope time, looking at exactly the right places in space at exactly the right time, to find enough supernovas to have data to even analyze. Not a whole lot of non-trivial science is just "read existing literature, think hard, and produce text."
But honestly, there's a lot that is (particularly for simulation like stuff which can be implemented in code). We should expect to see a bunch of extra output here, which might improve scientific productivity (even though the real bottlenecks are gonna be the academic publishers and their fixation on holding on to their copyrights).
In most cases it seems (at least to me and colleagues) to be turning out that picking best intelligence is a better option than picking better intelligence / dollar, assuming you can afford the cost. At least on interesting problems. If you’re doing generic web dev work, probably not the case.
What are interesting problems? Deepseek and some other open source models do decently well in the gpu kernel benches. https://kernelbench.com/hard Most people think they’re working on “very hard” problems and they really aren’t
Absolutely agree with this. As Louis Rossmann recently pointed out, it's the difference between a correct answer and a wrong answer; the correct answer is worth a good amount, while the wrong answer is worth nothing. Under this metric, for harder tasks, the most intelligent model is best per dollar.
That's why experience + cost effective model is IMO the best combination.
Experience allows you to design the skeleton where the implementation details are often inconsequential. There are relatively few scenarios where an LLM would need more guidance to render an outcome, but even dumb local models can do that.
Building a simple UI component vs an efficient multi-threaded bidirectional socket implementation (both examples of things I did recently with DeepSeek flash).
Angular can only be written in one way so the UI was trivial. I know the architecture for the socket implementation and the trade offs for various approaches, so I sketch out the implementation and get DeepSeek to complete it (error handling, keepalive messages, timeouts, etc).
I don't think I would have saved much time if I just asked Fable to "make the socket implementation" but even if it got it right the first time, it would have only saved me a few minutes given that's how long it took to write any way.
Even in more conventional applications, Node.js / React CRUD applications - "write a graphql query for blah" "add an endpoint to run query" "add validation to endpoint" are all trivial for DeepSeek flash. In most of these cases, I have found you're constrained by context window size because these are rarely well architected applications.
> Experience allows you to design the skeleton where the implementation details are often inconsequential.
Expanding on this thought a little more: it is possible to set up scaffolding that make incorrect implementations inconsequential. If the LLM can detect when the implementation is wrong, it can retry with the errors feeding back into the loop. This is shored up by up-front investments in tests, API definitions, strong types, linting rules, etc. the various cheap, fast Flash models do not need to 1-shot solutions if capable of autonomous reiteration.
This is a great point. In the past this is how I learned to architect software in general as you are never the only contributor and often must accommodate the contributions of contributors who are early in the careers. Good software architecture reduces the damage possible by low quality contributions.
I have had great success with contract driven development workflows in this regard.
LLMs being effectively very knowledgeable jr engineers, slot right into those patterns.
Another interesting thing is that Rust _really_ shines with LLMs. The compiler gives very helpful messages which the LLM bounces off very effectively. If it compiles, it works, you just need to double check the logic is correct.
By contrast, with something like Go, you'd need to be on the lookout for nil pointers, race conditions and so on. Node.js / TypeScript is less dangerous, but you still have memory leaks that aren't possible in Rust and lack pattern matching so must be vigilant about type discohesion.
This is mostly true and a month ago I would have agreed wholeheartedly. However, there are still situations where almost every line of code has to be deliberately chosen in a delicate way. Think research code or settings which demand a high degree of optimization. Opus or GPT-5.5 are absolutely miserable to guide in these settings.
"Well, don't use an AI for those purposes!"
That used to be true (and is now again true I suppose). Fable was a peek at something different, it seemed to be actually able to start to tackle these kinds of problems. That saves a lot of time, since checking these programs is often far, far easier than writing them. Experience is needed to check and scaffold still, but something like Fable becomes a prerequisite for these settings.
People are wasting tokens with perfectly fine models by not planning effectively. Anyone who says Opus is not capable of doing what Fable can probably are not planning effectively and just praying to the one shot gods.
What fraction of your work is "interesting problems", and what field do you work in?
> If you’re doing generic web dev work, probably not the case
I have a feeling the bulk of most people's work is "generic $X work." My take is people should figure out their mix of interesting vs boring work, and optimize accordingly. Flash models also tend to much faster,
Yep. After moving Deepseek for a long time I also moved to Mimo 2.5 pro. It's not only similarly cheap, but it's very good with 1m context window and gigantic tasks. Produces very lean code.
IMO US vs Chinese data vacuuming is a false choice. One of the main benefits of these open weights models is that you can get the privacy and cost savings by hosting your own model in cloud infrastructure.
Open weights models are only 4.5 months behind closed weight ones.
The fact that US considers propping its flagship technology by blacklisting competitors demonstrates how small the US competitive advantage really is.
Hosting DeepSeek Pro yourself is gonna be wildly expensive though?
You have a wide choice of providers available, so if you can find one you trust you can get inference without data harvesting and it's still very cheap. But dedicated HW is insanely inefficient.
Opus estimates you can do it for $13k/month if you get committed pricing on the HW.
The USA has an entire economic system to financialize your personal data and obfuscate your privacy. China can just sniff on their citizens API Calls, and use that to distill models.
They could just be flagging that they're considering blacklisting as a way to get some more Trump coin sold or other bribes. Or indeed to get further money from USA companies - which way will Trump decide, only the numbers in his family's bank accounts can give us a clue.
Using laws to ban competitors is just economic warfare thats all. Its got nothing to do with "national security", thats just the reason they give us normies. You should be safe in Australia since they actually need to be friends with China.
> You should be safe in Australia since they actually need to be friends with China.
As a non-Australian, I enjoyed very much reading about Australia in a book by Tim Marshall titled "The Power of Geography". I didn't quite realize just how vulnerable they become with China's ambitions and expanse in the South Pacific due to their reliance on vital sea lanes for trade with its Asian partners. After reading that book, I can appreciate your comment much more.
Economic safety can be guaranteed in many ways. It's not economic safety that these moves are for. It's for blocking anyone from preventing them from maximizing the value the Western oligarchs extract from both their own citizens and the entire world.
What? Every country has the right to freely decide which companies are permitted to operate under its jurisdiction, and exercising that right isn’t “warfare” of any kind.
I think the country that talks the most how the market regulates itself and needs no state intervention isn't the most qualified to put tariffs and block other countries from trade. And yet here we are…
How are you accessing their API? Through OpenRouter, or direct? Are you using DeepSeek v4 Pro? $2 seems a lot cheaper than my own experience accessing them through OpenRouter for over 100 million tokens, but I am using OpenRouter to access v4 pro.
Cache hit rate dominates your total cost calculation for long agent session, and it largely depends on the provider. Deepseek's native deployment is probably much better than third party in this regard. For v4 pro it's a whopping >100x price difference between normal input vs. cached input tokens.
I am using Flash and accessing the API directly via vscode insiders and occasionally Zed (it's buggy but I keep coming back to it because I want it to succeed).
Unless you need enterprise multi-model management, I don't see the point in OpenRouter as it just adds cost overhead and you can just self-host an open router alternative (LiteLLM, Bifrost, etc). Running an LLM gateway locally is kind of nice as it allows you to normalize your configurations against your internal gateway - but I haven't really needed to.
Deepseek direct is atleast 2x cheaper than other providers serving it. Maybe their caching strategy is just significantly better or they’re subsidizing api pricing. I think it’s the former. $2 is closer to 50 million v4pro tokens in my experience
Ensure you're hitting the DeepSeek provider via OpenRouter - they have the massive cached tokens discount. If you're hitting any other provider you're paying an order of magnitude more.
Pro is substantially more expensive than flash. In addition, there's wide variance in price with DeepSeek themselves providing the cheapest tokens last I checked (but they train on them). Caching policy also varies by provider. TTL can be as low as 5 minutes or as high as 24 hours and reading from the cache might or might not reset the timer. Whether or not you get a hit makes (IIRC) a 10x (edit: it's actually 50x) price difference in the case of DeepSeek themselves.
I second this. I’ve been using it a lot (with OpenCode) for personal projects. It’s intelligent enough at a tiny fraction of the cost of Claude or Codex.
The fact that America has yet to figure out the importance of soft power (and TRUST in particular) in the AI/information age is mind boggling to me.
The CLOUD act, FISA rulings, the Snowden leaks, and now the aggressive tech oligarch push to weaponize the unholy combination of AI, MAGA, and social media algorithms in an absurd and patently obvious attempt to impose (or maintain?) "world domination" seems likely to cement their downfall.
"Authoritarian" China's low-cost and open source approach looks downright democratic by comparison.
It seems increasingly clear to me that AI is forcing a reckoning in terms of how we interpret authority and control in light of new ways that information is evolving. The old political labels seem grossly insufficient to describe the present reality and the ones whining about "democracy", "freedom", and "liberal values" increasingly sound (and act) like bitter, out of touch old men desperately clinging onto a world that is rapidly outgrowing them.
Last time I checked, the closest approximation of liberal values are "liberty, equality, fraternity". They seem to think that "liberty" should only apply to them, "equality" is a threat to their consolidated power, and "fraternity" is something that should be weaponized to turn communities against one another to distract from the ways the oligarchs have been and continue to abuse and consolidate their power.
> As an Australian, I'm not sure that I care about the safety of my data when it comes to LLMs. US companies already stole scores of data
I'm not sure I understand this. I'm not defending the US, but isn't your data being in more hands worse?
Also, isn't Australia in a more contentious situation with China? Them being more allied with the US and all? Not to mention the whole nuclear sub issue. Having data stolen is shitty either way but isn't data taken by an adversarial country a worse situation?
If I spoke freely about how I feel about the US right now and the direct personal impact of policies implemented by the recent administration, I'd be denied entry into the country to visit friends later this year.
It is inconsequential if the US or China have my data, both will misuse it and I am powerless to protect myself from that fact.
Not using LLMs presents a bigger threat to my career than protecting my data.
Correct - As an Australian I feel free to say anything. But as an Australian who might like to travel again to the USA at some point in the future, I do not feel I have freedom to share openly online.
Whereas, China does not ask for my social network logins, or for me to be pre-approved to travel there. So that is unlikely to be affected.
> If I spoke freely about how I feel about the US right now
If you read my comment as defending the US then you've misread. Also, you're probably just as pissed as 60% of Americans
> It is inconsequential if the US or China have my data
Sure it does. The way each distributes data between government and industry has some differences. So too does the different disinformation campaigns each country is running against Australia.
But the main point is really that 2 > 1. 1 country scraping your data is bad. 2 is worse.
To be fair, most of their votes don't actually count.
And even with our westminster system, One nation has swept the right wing. Considering how many seats they will likely get in the next election, my high horse has shrunk a fair bit.
Why do you think I like them or support those actions?
Do you support everything the Klingon Empire does? That doesn't seem like the Worf I know. The Worf I know can distinguish the people of a civilization from the leaders of their societies. To conflate them just creates the problems you are criticizing
The same fact is true in a democracy. A democracy specifically is set up so just a single person doesn't rule. That's called a dictatorship.
But you turn a democracy into a dictatorship through tribalism. So forgive me if I get mad at you trying to place me into a nice well defined box. You're helping erode the democracy I (and so many others) desperately want to maintain (and improve).
Sorry that you're being blamed. People have a pretty negative sentiment towards Americans in general right now because we have been watching helplessly in horror as our best friend tears things down.
I have friends in blue states who didn't vote because it wouldn't have changed the outcome given the electoral college - it's unfair to blame them for not participating as they were never able to in the first place.
Not that democrats are angels. You guys are damned if you're blue / (extra) damned if you're red - given the state of both options.
Consider that USA's government logic is: "if we don't like a government we will kill civilians until they rebel and put a government we like". See Cuba for example.
So, by their own metric, they aren't helpless at all and should appreciate other countries aren't as savage as they are with civilians, instead of complaining.
Look Worf, I'm just one human. You don't know me. You don't know what I'm doing or not doing. But I know you're making our situation worse and I know you're not living up to the ethics of your namesake
As an Australian.. politically I need to worry about business data touching China. It will come up at a Risk Advisory Committee meeting as a serious issue.
In actual personal practice, no. China having my data presents no actual impact to me, America will do things that impact me.
China is mostly interested in geopolitical stuff and getting an economic advantage, plus they have no jurisdiction in the US. Your data in the hands of the US government however could potentially land you in prison.
Most Aussies aren't really worried about China... China remains our largest trading partner, and polling shows less than half of Australians think the AUKUS alliance (which includes the nuclear subs) makes our region safer.
Latinamerican here.
When you talk about "adversarial country" I think of the USA (they can kidnap a president, kill people on boats without a trial, etc) and not China.
YMMV for different regions.
China is not digging through my social media in order to find a reason to cut my grant funding. China isnt going to pull me over for speeding in montana only to examine my phone to see whether i am maga enough to get off with a warning. And china isnt at o'hare security scanning for anyone with skin darker than freshly fallen snow. China may be evil, but it is a far away evil that doesnt have a physical impact on my day to day. The other evil is much closer to home. Even if it is not the biggest, the crocodile closest to the canoe is always more concerning than the one still on the bank.
While having data/code stolen isn't ideal, there is a certain point where you need to assume it's already out there. There's actually more probability of harm from shady US companies imo, because people are less suspicious about data sovereignty
Historically speaking, the US might even be a larger risk to Australia than China is. The US alliance goes back a long way, and so does the opposition to US influence. Since the Whitlam government, MPs are generally fearful of US retaliation, rendering much of our politics hostage to US influence. Of course, some PMs have openly embraced the US so this feeling isn't universal. But many of our issues are directly tied to the US.
China has been far more beneficial to Australia by comparison, with the downside being the encroaching influence of CCP propaganda. Many of our strengths are tied to our relationships with Southeast Asia.
Paul Keating has famously declared US as an "aggressive ally", "our colonial masters", AUKUS as our "worst international decision", and that "our future is in Southeast Asia". This was under Biden too.
So the situation is much more complicated, and the feelings on the ground right now is that the US are not our friends (of course, the CCP is not great either).
The same Kim Dotcom who was arrested in Germany, extradited from Thailand to Germany, and most recently has been parroting Russian propaganda points about Ukraine? The same one whom the New Zealand legal system has allowed to exhaust every legal avenue and remains free to this day?
There are much better examples to use to showcase the US's extrajudicial/international reach than this guy.
Not jailable. But I'm absolutely free to disregard moral agency of mr. Kim Dotcom and stop paying attention to whatever he says or what happens to him as a result of his support for warring Russia.
Laws are ultimately defined by the boundary and edge cases.
The godly person, doing no wrong, that through unfortunate circumstances, is brought before the court does not advance our understanding of a law.
It's the messiest case, by unscrupulous persons, in unconscionable circumstances that ultimately decide if the execution of that law gives the public more or less rights than on initial interpretation.
Yes, but public minds are not made up by edge cases. Surely a better example could be used rather than a divisive criminal.
Also, not sure how anyone except bad actors or sloppy readers could have gotten that I was defending the US here instead of criticizing using Kim Dotcom as an example in characterizing the US's actions.
> Also, isn't Australia in a more contentious situation with China?
China does not run 'rendition flights' to kidnap other countries' citizens and hold them in gitmo style prisons for decades. China cannot compel and does not compel random countries to extradite foreign citizens to China to persecute them and lock them up. Guess who does.
I can second this. Deepseek is great for mundane work day tasks.
I use it via claude code, just pointing the api to deepseek.
It's also not a clear "Opus 4.8 >> DS 4 Pro", I've done 16 tasks in 4 days across the two, and while Opus was indeed on average better, both models performed well being able to handle most of my workload.
In fact DeepSeek was _significantly_ better on 3 task out of 16 and Opus was _significantly_ better only in 2 out of 16.
So why I still claim Opus 4.8 to be the winner? Because the few times that DS failed or got off the rails, it failed much harder and needed several prompts to be realigned on the actual tasks.
Another thing at which Deepseek is significantly behind is code reviewing. Opus is more intelligent/thorough, Deepseek will sometimes generate bogus or low quality feedback.
And the last thing at which Opus is better, period, is vibe coding. If you want to implement features end-to-end it handles ultracode flows quite better. I don't vibe code at work, but I do so with personal projects.
But the cost concern is real. I've spent sub 2$ in 5 days of work using DS 4 Pro, which is on average just 4 queries to Anthropic.
Give me a slightly better DS 4 (it is still in preview and training isn't finished) and I may ditch Anthropic for good.
Have you tried glm 5.2? It’s better than ds4pro and very close to opus4.6. Alot more expensive than ds4pro/mimo but significantly chesper than opus 4.6 for 90% of the quality
I've been using deepseek for some development at home and it is really good for the price. It is at the point where i am ok with using it as a tool that i can rely on and not an expensive gadget with flaky uptimes.
I'm not interested in explaining to you something immediately evident that you clearly disagree with. If you don't feel that trillion dollar companies training commercial products on the free labor of human writers/artists/engineers/etc. without their consent or compensation isn't "stealing" you're welcome to that opinion, but I believe that's what the GP was implying.
I'm in favour of sending as much data as possible to DeepSeek for them to train on. I'd happily zip up everything useful I have and send it to them if they asked for it.
I'd be fine if it cut both ways. It shouldn't be legal for works distributed by people under licenses that would prohibit derivative work whilst if I reverse engineered a Nintendo Switch 2 with an LLM I'd be sent to the underworld.
- generating types for APIs
- generating boilerplate based on existing code
- improving existing code (adding error handling, timeouts, things like that)
- Writing SQL repository boilerplate / queries
- Creating implementations against hand written tests
- Helping me understand and implement APIs from third party libraries
- Writing documentation
I've spent like $2 in the last month and have used over 100 million tokens.
It's doubled my productivity and unlocked work that I could not have done before.
As an Australian, I'm not sure that I care about the safety of my data when it comes to LLMs. US companies already stole scores of data to train their models on and it's hard to imagine they suddenly grew some integrity. I'll care when regulators step in, until then it's out of my control so I'll just use the best price-to-productivity product available.