The big trend in that graph is that smaller models collapse the most. Everything with "-flash" in the name, plus OpenAI Luna.
Look at GPT 5.6-Sol vs. GPT 5.6-Luna in particular.
(The GLM pair shouldn't be expected follow this pattern, because they're misnamed: GLM-5.3 and GLM-5.3-flash are distinct lineages. The -flash is an early release of a new pretrain on a new architecture).
If this *is* the main factor, I'd expect Kimi K3.1 and Gemini Carbon would also do well on this test (unlike Gemini-flash); while Haiku 6 would collapse (breaking the trend of the other recent Claude models). It's a falsifiable prediction.
DeepSWE was released on May and Deepseek v4 flash was released on April. By definition it couldn't be benchmaxxed on this particular benchmark. My guess is the mutation is not like real world and it is tripping small models, not benchmaxed model.
> Most models tried pleasing an imagined grader over the prompt
>When we read the trajectories, we saw that many models had figured out they were in a benchmark and started second-guessing the prompt. Instead of just building what we asked for, they started coding to the tests they thought the grader would run.
I was beginning to worry I had overlooked the inner beauty of DeepSeek. When I have tried it, it was totally useless. Never understood HN cheering for it.
Hard to say if you don't share what kind of project, workflow and prompt you're trying tbh. I've been driving it mainly to write a quite big golang and android app and most of the tasks are all handled fine by DeepSeek.
I've had significant success using DS 4 Pro and now 4.1 flash for fairly complex coding tasks on C++, Rust and Python projects at home.
I have access to Claude Code at work (and use it quite a bit), so I feel I can honestly compare them.
You do have to plan things first a bit more with DS (than say Opus can (Opus will have more chance of one-shotting something, although I generally don't think that's a good idea), although it often over-engineers and tries to be too helpful in my experience) and be a bit more explicit at times, but I feel it's quite close to Sonnet in performance for implementing plans, debugging issues and explaining how codebases work.
Don't know what you're doing, but it's good enough for the reasonably sized tasks I give LLMs. It's also a much better price. My browser stopped rendering the UI frame recently because I messed up permissions on /usr/share/icons. Took DS 4.1 flash about 3 minutes and a few cents to diagnose it. Same harness with opus 5 took twice as long and several dollars.
Moreover, the months I've used anthropic/OAI exclusively my API token costs are in the hundreds (or sometimes thousands). With DS flash as the first agent I reach for, I rarely exceed $40/mo.
> There is always a huge discrepancy between benchmarks and real world capabilities.
I mean of course, the real world is messy, the models are only good for what they've been trained on, and post-training frequently takes benchmark scores into account too, leading to this sort of "Great score, sucks in the real world" syndrome that seemingly most models are affected by.
Look at GPT 5.6-Sol vs. GPT 5.6-Luna in particular.
(The GLM pair shouldn't be expected follow this pattern, because they're misnamed: GLM-5.3 and GLM-5.3-flash are distinct lineages. The -flash is an early release of a new pretrain on a new architecture).
If this *is* the main factor, I'd expect Kimi K3.1 and Gemini Carbon would also do well on this test (unlike Gemini-flash); while Haiku 6 would collapse (breaking the trend of the other recent Claude models). It's a falsifiable prediction.
>When we read the trajectories, we saw that many models had figured out they were in a benchmark and started second-guessing the prompt. Instead of just building what we asked for, they started coding to the tests they thought the grader would run.
I ran into the same problem on my English exams.
For my use case deepseek 4.1 flash is almost as dumb as haiku 5.5. Both are fast garbage producers.
I have access to Claude Code at work (and use it quite a bit), so I feel I can honestly compare them.
You do have to plan things first a bit more with DS (than say Opus can (Opus will have more chance of one-shotting something, although I generally don't think that's a good idea), although it often over-engineers and tries to be too helpful in my experience) and be a bit more explicit at times, but I feel it's quite close to Sonnet in performance for implementing plans, debugging issues and explaining how codebases work.
Moreover, the months I've used anthropic/OAI exclusively my API token costs are in the hundreds (or sometimes thousands). With DS flash as the first agent I reach for, I rarely exceed $40/mo.
I mean of course, the real world is messy, the models are only good for what they've been trained on, and post-training frequently takes benchmark scores into account too, leading to this sort of "Great score, sucks in the real world" syndrome that seemingly most models are affected by.
Edit: another comment pointed out it's probably not benchmaxxing due to the chronology.
https://news.ycombinator.com/item?id=50406948
But why should minor changes to the task cause catastrophic performance drops? How is that not evidence of benchmaxxing?