Anthropic is a bit nuts, I had $260 of credits on my max account for the extra usage the other night. It was expiring, so I figured I'll fire up an agentic swarm to deep dive and make some deep changes to some old cold bases.. literally 25 minutes or less, $260 burnt, it didn't get get into the implementation, just wrote a ton of useless plans for the most part. It really opened my eyes to what they expect to charge people.. wayyyy overpriced.
I managed to lose around $300 in credits I had saved for some emergency /fast sessions the following way: switch to Fable. Work on the design. Downgrade to Opus for the build. If any of other parallel Opus session has /fast enabled it seems to enable it for the newly spawned session by default. Before I knew it, the $300 was gone. I think the bug is now solved, but it was rather unpleasant. I dont ever remember bugs that would drain my wallet - with claude code its just another Tuesday. Still love it.
They have something like 80% gross margins, are at a $100B/yr ARR, and are growing at 10x per year... If that keeps up, they're going to be doing more revenue than Google in a year ($400B ARR, 20% per year growth)
I clicked through and it showed Qwen at the top at 55.4 compared to 55.3 for Opus Max. I have a screenshot.
Then I clicked away and back, and now it goes Qwen second, with 58.4, to Opus Max at top with 59.2.
I have screenshots of both. The description above the chart is the same in boh cases:
> Artificial Analysis Agentic Index
> Represents the weighted average of agentic capabilities benchmarks in the Artificial Analysis Intelligence Index (GDPval-AA v2, ³-Banking)
What happened? How can the scores change so much in a few seconds?
I believe it. It's extremely good at troubleshooting. I gave Qwen and Kimi K3 the same annoying, complicated, intermittent bug to track down. Kimi did a bit better in understanding the existing code, but Qwen built some diagnostic tools and did an excellent statistical analysis on the log data. Qwen got way closer to the truth.
I'm very much looking forward to their forthcoming smaller model Qwen 3.8 releases. A version that can easily run locally would be great.
I use https://pi.dev/ which works fine out of the box but is fairly minimal and intended to be customized. There are many extensions.
OpenCode or oh-my-pi might make more sense if you just want a batteries-included agent. You can also make Claude Code work with other models without too much work, but I think that's asking for headaches.
I used claude with GLM and it's easy to set up, just hard to find the documentation. No headaches really, unless you want to use it against multiple different APIs.
They didn't run all benchmarks. It's the best in AA agentic index (GDPval-AA v2, ³-Banking) but not coding index (DeepSWE which is missing, Terminal-Bench v2.1 they have 81% vs 90% for Sol, SWE-Atlas-QnA missing).
Does "artificial analysis" mean what it says? Dubious.
But: I've been very impressed by the larger Qwen Models, and a brief try of Kimi also impressed me.
A lingering sense of quality degradation when going deep remains.
But that's not an accusation: they seem to be hitting the compute/quality tradeoff extremely well.
And on-prem capability is simply irreplaceable.
Apart from all the innovations that were driven by the strive for this optimization: quantization, "distilling" (without obvious mad-cows-disease)... I think China was an invaluable player in this progress. Intuitively, I'd even go so far to speculate that LLaMa wouldn't exist without the competition.
> Artificial Analysis Agentic Index: Represents the weighted average of agentic capabilities benchmarks in the Artificial Analysis Intelligence Index (GDPval-AA v2, Tau³-Banking)
> Artificial Analysis Coding Agent Index v1.3 incorporates 3 benchmarks: DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA
Qwen3.8 Max is 55.4 on the Agentic Index but hasn't been tested for the Coding Agent Index.
Looks like coding agent is model+harness. There are far fewer models represented on that page. I believe "agentic index" is still the metric to look at for coding performance. I could be wrong about that though.
I'm dumbfounded to see Opus 5 making SO MANY mistakes in coding simple stuff. Most times, Fable 5 comes out to be cheaper because it nails so many things much quicker than Opus 5.
I have Fable plan and Opus implement. I haven't had any major issues working this way; however, Opus does seem plain fucking stupid compared to what I experienced with Sonnet previously.
I do the same, and generally have good results, but it does stupid things with gusto.
I'd open a blog with "weird things Opus did". Today it launched a swarm of cpu-hogging processes to test if the widget showing machine and I/O load is rendering nicely and correctly. The test went fine, but it was no longer able to kill those processes since they were really effectively hogging the CPU in various ways - being diligent, some of them were hogging CPU, some were murdering the SSD, some were pounding on the network adapters. Took me 30 mins to recover the machine to a working state without killing the meaningful, messy, in-flight sessions i had going on on other projects.
Weird how different people's experiences are. If it's making simple mistakes something must be wrong in your setup/context I assume? It's been solid for me, beyond the usual LLMisms that all models have. But I keep context pretty minimal.
I have some internal tests I use for areas where one particular solution/paradigm is dominant but worse.
Opus 4.6 is the last model that's actually useful and can "adjust" its perspective to use the newer & better solution.
Where Opus 4.8-5 has over fit training on worse/older but "dominant" solutions it refuses to adjust.
Not only does this create an existential threat to adopting progress but it also means that if you have a code base that has rare but real world tradeoff the newest versions of Opus 4.7, 4.8 and 5 are worse than useless and become a major dev timesink.
It's my daily driver. I like it and find it noticeably better than Opus 4.8.
After I started reading complaints about Opus 5, I gave Fable the task of evaluating a bunch of code Opus 4.8 had written and compare it to Opus 5's code. Fable ran a dynamic workflow and the scores came back 15-20% higher for Opus 5's code in terms of quality, correctness and readability/conciseness. I did not tell Fable which Opus wrote which code, and I turned off memory as well to ensure there was no pollution from that angle.
My only complaint is that Opus 5's prose is annoying as hell. I wrote a custom skill for it for concise debriefs and it has been working pretty well for me.
"As you requested, I've finished task X. Honestly, task X turned out to require task Y, which I haven't actually done. Task Y is the next step if you'd like to continue along this route."
Hopefully this boils down to the smaller versions they've teased. In my experience, Qwen models are the closest to the "less knowledge, more intelligence" (yes, the two are hugely correlated!) ideal some tool-dependent tasks need. Even the 3.5 2B can be easily prompted to always lean on tools and not jump to false conclusions (although its actual coding skills are abysmal, as you'd expect).
People produce such models by over-RL-ing smaller models on math and coding tasks. I've found the results capable of neither innovative work nor thinking outside the box. They're straight-A students raised by tiger moments who never let them play freely for hours in the dirt.
Perhaps you could say such models are skilled --- but intelligent? Not by my measure.
People and AIs alike need diversity of experience and a broad liberal arts education to see hidden connections between fields and make real advances.
Why does an open weights model cost nearly the same as GPT5.6? $1.14 vs $1.23 on the cost index. Since you can't presumably run this on your own hardware given the model size and hence gain other things like privacy, I don't see any reason to move away from GPT at this rate.
Many providers will host it and will compete on price. It also can't easily be taken away because one company (or one government) decides they don't want it around any more. People can fine-tune it for particular workloads.
They cherrypicked benchmarks. The ONE weighed benchmark where is beats Opus5 by 0.1 points is what was linked because that's how propaganda works. The Agentic Index that includes the full benchmark suite has it in 5th place.
It's barely better, and barely cheaper, not really enough to challenge the status quo IMO. Half the price for basically the same performance would be a much stronger value proposition.
Things change radically month to month. Nobody is remotely close to capturing the market or having any kind of stability over time. People move around quite a lot, often to sidegrade within a generation. Just playing fly on the wall with discourse would be enough to tell you all of this, even without the data to back it up.
Why should open weights correlate with cost? Cost correlates with the expense of running the model more than it does to the expense of developing the model.
Qwen Max is their large model - over a trillion params. Similar to Kimi K3 in size. Qwen 3.8 27B is going to be more accessible to your own hardware. I'd say that Qwen Max is not approachable for the majority of people and companies to self-host.
GPT5.6Sol completes the suite in 70M tokens, while Qwen3.8Max needs like 145M tokens. So this is a case where models like Qwen 3.8 and Kimi K3 use a lot more output (reasoning) tokens, go a good bit slower, so they can ultimately achieve a better intelligence score than if they went more quickly.
There are a couple of frontiers (ok bad word, maybe categories) in open weight models.
These Qwen 3.8 and Kimi K3 style models aren't trying to win on price, they're trying to compete on intelligence and capability.
Models like Deepseek V4 Flash (updated this week) are $0.03 a task, or 50X cheaper than Qwen3.8/Kimi K3, and 100X cheaper than Fable, while offering stunning intelligence. That's a different frontier for competition, and perhaps one more interesting for someone who wants to see them compete on cost.
Prefill is survivable if you cache well. But what kills me is the context. Qwen 27 needs a ton of room for KV Cache. I guess not an issue on a 128 GB Halo or Spark, but if you are running of consumer/prosumer GPUs it's miserable to be compacting every 120k tokens.
I interpret @syntaxing as meaning they are looking forward to running Qwen3.8-27B, but are frustrated by prefill times with other models, such as Qwen3.6-27B.
I meant Qwen3.6. Unsloth supposedly has early preview of the model and the VRAM requirement is the same so most people expect similar model size and type.
I've been trying it on several projects and have found it's pretty sloppy. It leaves stuff broken, doesn't reliably write tests to check its own work unless explicitly prompted, misunderstands the assignment, etc.
It is smart and reasonably quick but not reliable.
I've come to the same conclusion over and over with all of the Chinese models that have been claimed to be catching up with OpenAI's and Anthropic's frontier models (Deepseek 4, GLM 5.2, Kimi K3).
At their best, I think they're closing in on Opus and GPT, but they're incredibly inconsistent and the variance in output quality is much higher than the best from any of the Anthropic or OpenAI models from the last few generations. The only way I can describe it is that it feels like a lack of intuition with the models which means I find my self needing to write longer prompts or have more back and forth to get them to do what I want from them.
To give an example, I have a saved prompt that I use as a sanity check on some data I'm storing. It reads about 50 rows from a DB and matches them to the UI and makes sure the data is displaying correctly. I've been using this with GPT 5.5 and now 5.6 for a few months and running it a few times a week with no issue. Sometimes I'll run it multiple times in a single chat if I notice bad data (run it, fix thing, run again, fix another thing).
I recently tried to switch to using Deepseek v4 (first flash and then pro) and while both did the task just fine, both would do things like change the response format from one message to another in the same chat or randomly decide to omit things it didn't think were relevant. At one point I ran the prompt, fixed some bad data, and then said "Okay, I fixed row 7, run {prompt} again" and so it decided to leave row 7 out of the response. A few times the first message would contain a table and then the next run in the same chat would contain the data in a bulleted list.
None of those are major issues and all could be solved with a bit more rigor in my prompting, but for me it makes them harder to work with. Those examples are a bit trivial, I think they're the easiest way for me to illustrate the gaps I see with them.
It is also the most expensive open source frontier model, per task; cf. Cost per Intelligence Index Task. If it is as good as the benchmarks say, it is a positive indicator for Qwen and China, but I see no reason to use it.
Depends on what you want to do. Some task specific models can be trained with a few ten or hundred thousand training examples so you can use a bigger model to produce synthetic training examples and then fine tune a smaller student model. I think that's the usual process. Whether you'd get acceptable performance this way depends, as mentioned, on what you're trying to do and what you'd consider acceptable.
once you are able to get the full probability distributions per token you can distill it on specific domains. distilling without that isn't generally a good idea unless you have invested millions in the requisite infrastructure.
DS4 Flash Q2/Q4 mixed quant fits on a DGX Spark (a $4000 device which is not particularly unheard of expense for Apple customers), and is indistinguishable for me from Opus for my personal daily use/assistant benchmarks[0].
Haven't tried this yet, but going to soon! I have to wonder what happened at Anthropic. We've cancelled our subscription in favor of OpenCode & Codex. Sol is just so good & OC goes so far for every $ spent. Claude's become a pain to work with - average output with an annoying personality. Who knew this would be an issue even a year ago? In any case, loving the stuff from the Chinese models!
I was with you until there. Qwen and the OpenAI models are great, aggressive agents, but they’re not as good as the anthropic models for human interaction. They just don’t have the subtlety, understanding, or attention to detail.
Really? I've heard so many other people complain about this recently. And maybe it's possible that it's the prompt style even. But interesting that it's not across the board.
Opus 5 is just a token burner.
I use fable plan and spawn opus 4.8 workflows which seems to work alright.
With the $200 subscription, I can have Fable on ultracode working for hours and not dent the usage limits.
Our leaderboard combines Arena ELO, AA Intelligence index, latency and speed and goes: #1 Opus 5 #2 Kimi K3 #3 Qwen3.8 Max #4 GPT 5.6 Sol
Source: http://pellmell.ai/leaderboard.
This jumps around a lot based on the top throughput and latency of whatever provider happens to be best at the moment.
Then I clicked away and back, and now it goes Qwen second, with 58.4, to Opus Max at top with 59.2.
I have screenshots of both. The description above the chart is the same in boh cases:
> Artificial Analysis Agentic Index > Represents the weighted average of agentic capabilities benchmarks in the Artificial Analysis Intelligence Index (GDPval-AA v2, ³-Banking)
What happened? How can the scores change so much in a few seconds?
https://artificialanalysis.ai/methodology/intelligence-bench...
I'm very much looking forward to their forthcoming smaller model Qwen 3.8 releases. A version that can easily run locally would be great.
OpenCode or oh-my-pi might make more sense if you just want a batteries-included agent. You can also make Claude Code work with other models without too much work, but I think that's asking for headaches.
But: I've been very impressed by the larger Qwen Models, and a brief try of Kimi also impressed me.
A lingering sense of quality degradation when going deep remains.
But that's not an accusation: they seem to be hitting the compute/quality tradeoff extremely well.
And on-prem capability is simply irreplaceable.
Apart from all the innovations that were driven by the strive for this optimization: quantization, "distilling" (without obvious mad-cows-disease)... I think China was an invaluable player in this progress. Intuitively, I'd even go so far to speculate that LLaMa wouldn't exist without the competition.
$0.36 per task, Intelligence Index score 56 -> Grok 4.5 high
$1.13 per task, Intelligence Index score 58 -> Qwen 3.8 Max
$0.81 per task, Intelligence Index score 59 -> GPT 5.6 Sol xhigh
$1.80 per task, Intelligence Index score 63 -> Opus 5 xhigh
> Artificial Analysis Agentic Index: Represents the weighted average of agentic capabilities benchmarks in the Artificial Analysis Intelligence Index (GDPval-AA v2, Tau³-Banking)
> Artificial Analysis Coding Agent Index v1.3 incorporates 3 benchmarks: DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA
Qwen3.8 Max is 55.4 on the Agentic Index but hasn't been tested for the Coding Agent Index.
https://artificialanalysis.ai/models/qwen3-8-max
Doesn't have the claim either. Clickbait?
Even then, this seems a much more marginal win than the headline suggested to me.
I'd open a blog with "weird things Opus did". Today it launched a swarm of cpu-hogging processes to test if the widget showing machine and I/O load is rendering nicely and correctly. The test went fine, but it was no longer able to kill those processes since they were really effectively hogging the CPU in various ways - being diligent, some of them were hogging CPU, some were murdering the SSD, some were pounding on the network adapters. Took me 30 mins to recover the machine to a working state without killing the meaningful, messy, in-flight sessions i had going on on other projects.
Infuriatingly so, in a way I don't remember Opus 4.8 being, but maybe I've just been ruined by Fable 5.
Opus 4.6 is the last model that's actually useful and can "adjust" its perspective to use the newer & better solution.
Where Opus 4.8-5 has over fit training on worse/older but "dominant" solutions it refuses to adjust.
Not only does this create an existential threat to adopting progress but it also means that if you have a code base that has rare but real world tradeoff the newest versions of Opus 4.7, 4.8 and 5 are worse than useless and become a major dev timesink.
After I started reading complaints about Opus 5, I gave Fable the task of evaluating a bunch of code Opus 4.8 had written and compare it to Opus 5's code. Fable ran a dynamic workflow and the scores came back 15-20% higher for Opus 5's code in terms of quality, correctness and readability/conciseness. I did not tell Fable which Opus wrote which code, and I turned off memory as well to ensure there was no pollution from that angle.
My only complaint is that Opus 5's prose is annoying as hell. I wrote a custom skill for it for concise debriefs and it has been working pretty well for me.
People produce such models by over-RL-ing smaller models on math and coding tasks. I've found the results capable of neither innovative work nor thinking outside the box. They're straight-A students raised by tiger moments who never let them play freely for hours in the dirt.
Perhaps you could say such models are skilled --- but intelligent? Not by my measure.
People and AIs alike need diversity of experience and a broad liberal arts education to see hidden connections between fields and make real advances.
Many providers will host it and will compete on price. It also can't easily be taken away because one company (or one government) decides they don't want it around any more. People can fine-tune it for particular workloads.
Might as well use gpt-sol.
It's barely better, and barely cheaper, not really enough to challenge the status quo IMO. Half the price for basically the same performance would be a much stronger value proposition.
Things change radically month to month. Nobody is remotely close to capturing the market or having any kind of stability over time. People move around quite a lot, often to sidegrade within a generation. Just playing fly on the wall with discourse would be enough to tell you all of this, even without the data to back it up.
That said, it's a fair point. For me, it boils down to things covered here: https://earendil.com/posts/session-portability/
Things like obscured reasoning traces.
There are a couple of frontiers (ok bad word, maybe categories) in open weight models.
These Qwen 3.8 and Kimi K3 style models aren't trying to win on price, they're trying to compete on intelligence and capability.
Models like Deepseek V4 Flash (updated this week) are $0.03 a task, or 50X cheaper than Qwen3.8/Kimi K3, and 100X cheaper than Fable, while offering stunning intelligence. That's a different frontier for competition, and perhaps one more interesting for someone who wants to see them compete on cost.
I've been trying it on several projects and have found it's pretty sloppy. It leaves stuff broken, doesn't reliably write tests to check its own work unless explicitly prompted, misunderstands the assignment, etc.
It is smart and reasonably quick but not reliable.
At their best, I think they're closing in on Opus and GPT, but they're incredibly inconsistent and the variance in output quality is much higher than the best from any of the Anthropic or OpenAI models from the last few generations. The only way I can describe it is that it feels like a lack of intuition with the models which means I find my self needing to write longer prompts or have more back and forth to get them to do what I want from them.
To give an example, I have a saved prompt that I use as a sanity check on some data I'm storing. It reads about 50 rows from a DB and matches them to the UI and makes sure the data is displaying correctly. I've been using this with GPT 5.5 and now 5.6 for a few months and running it a few times a week with no issue. Sometimes I'll run it multiple times in a single chat if I notice bad data (run it, fix thing, run again, fix another thing).
I recently tried to switch to using Deepseek v4 (first flash and then pro) and while both did the task just fine, both would do things like change the response format from one message to another in the same chat or randomly decide to omit things it didn't think were relevant. At one point I ran the prompt, fixed some bad data, and then said "Okay, I fixed row 7, run {prompt} again" and so it decided to leave row 7 out of the response. A few times the first message would contain a table and then the next run in the same chat would contain the data in a bulleted list.
None of those are major issues and all could be solved with a bit more rigor in my prompting, but for me it makes them harder to work with. Those examples are a bit trivial, I think they're the easiest way for me to illustrate the gaps I see with them.
I wasn't able to find an explanation from them. Anyone knows what happened?
[0]https://humanparadox.org/local-vs-frontier-benchmarks-for-my... - note here I tested Q8 but have found no difference at lower quant.
I was with you until there. Qwen and the OpenAI models are great, aggressive agents, but they’re not as good as the anthropic models for human interaction. They just don’t have the subtlety, understanding, or attention to detail.