@PawelHuryn claude sonnet better than opus?
OrigineelDus ik heb GPT-6.1 Sol getest op echt werk: 2 repositories, 105 ingebouwde bugs. Vind en los op wat je kunt. In tegenstelling tot GPT-6 Sol, wat gewoon een generfde GPT-5.6 Terra was, is deze echt. De resultaten: - GPT-6 Astra (max): 45 voor $33 - GPT-6.1 Sol (max): 44 voor $6,56 - GPT-5.6 Sol (max): 43,5 voor $95,35 - Opus 5.5 (max): 41,7 voor $58,53 - GPT-6 Sol (max): 29,3 voor $9,33 Met zo'n model maakt de korting van 50% op het abonnement van $200 niet uit. n=1. Meer runs en resultaten op andere inspanningsniveaus verschijnen in deze thread ove
0
0
0
0
Reacties 42
-
-
@PawelHuryn Sonnet 5.5 in first, what?
Origineel -
@PawelHuryn @grok explain
Origineel -
@PawelHuryn GPT 6.1 sol has been absolutely stupid for me. https://x.com/cyberpeterg/status/2105313537543303596?s=20
Origineel -
@PawelHuryn what was the 1 bug Astra caught that Sol missed?
Origineel-
@TheHasnainK The answer key and specific bugs aren't public, but you can see a bug coverage here: https://bughunt.productcompass.pm/
Origineel-
@PawelHuryn @TheHasnainK Why is Fable 5.1 not included?
Origineel-
@tf3y @TheHasnainK You can include it on the website. It's barely better than Opus 5.5. No point in using it right now, IMO.
Origineel-
@tf3y @TheHasnainK @tf3y, it's a single run, so the coverage looks less advanced. This single score is 1.3 points higher than the average for Opus 5.5 (n=3)
Origineel
-
-
-
-
-
@PawelHuryn Nuts
Origineel -
@PawelHuryn 44 bugs, $6.56. Cost-per-fix is the ladder.
Origineel -
@PawelHuryn WOWIE! amazing results :)
Origineel -
@PawelHuryn @grok sol 6.1 is clearly cutting into OAI's margins. Is this "race to the bottom" reminiscent of early 2000's internet prices? When did this race start? When did it end? Outcomes?
Origineel -
@PawelHuryn Просто на подписке за 200$ можно было бы бесконечно пользоваться 6.1 sol ultra и не покупать кредиты, openai это непонравилось😂
Origineel -
@PawelHuryn do the fixes get checked for new bugs too, or just whether the planted ones are gone?
Origineel -
@PawelHuryn Opus 5.5 at 41.7 for $58.53 against 6.1 Sol at 44 for $6.56 is the line that matters here. Once you log a few more runs, I'd love to see whether the misses overlap or whether Opus catches bugs Sol never touches, because that decides routing versus just picking one.
Origineel -
@PawelHuryn how much you got from openai? "With a model like this, the 50% cut to the $200 plan doesn't matter." all these are just for this sentences to plant the idea it's ok. it's not ok bro.
Origineel -
@PawelHuryn Short: 6.1 Sol finally sits on the right shelf. Not the bug king. The king of cost per bug found.
Origineel -
@PawelHuryn Interesting
Origineel -
@PawelHuryn There is alternative 3D visualization of the results, can make 3 axis comparison easier. https://icauro.com/toabundance?benchmark=bug-hunt
Origineel -
@PawelHuryn The only problem? "remove the xx button from the page" ... 'valutating bla bla bla' ..... 10 minutes later 'button removed, i'll now verify if the button is not visible' ..... . 15 minutes... 'the button is removed and not present on the page' (but still in temp page) 30m tot
Origineel -
@PawelHuryn thanks for running this and sharing the test results!
Origineel -
@PawelHuryn A bit useless to show the highest level of AI. The real challenge comes when you test on medium.
Origineel -
@PawelHuryn is it just me or is gpt 6.1 sol slow? what tps are you getting?
Origineel -
@PawelHuryn In terms of criticality (identifying more impactful bugs), was GPT-6.1 Sol on par with GPT-6 Astra and Fable 5.1? Btw, thanks for this, this is gold! 🤝
Origineel -
@PawelHuryn https://x.com/SlovenskiGlas1/status/2105265982948180341
Origineel -
@PawelHuryn GPT 6.1 Sol is basically a telenovela: the cast is great, young, and gorgeous, the costumes are stunning, but almost nothing actually happens. You could easily skip four episodes and still not miss a beat.
Origineel -
@PawelHuryn TY
Origineel -
@PawelHuryn Glad you're adding n=3. I'd put a bug-by-bug stability view next to the totals: fixing 44 different bugs each run creates a different review workload from reliably fixing the same 44.
Origineel-
@jakestone2306 It's not 100% deterministic. Bugs differ a lot between runs. This view: https://bughunt.productcompass.pm/?preset=featured&view=coverage
Origineel-
@PawelHuryn @jakestone2306 GPT-5.6 Sol + GPT-6.1 Sol looks like an interesting combination or is it just a coincidence that one finds bugs at the end the other one missed ?
Origineel-
@PawelHuryn @jakestone2306 Uh, just saw that GPT-5.6 Sol is horrible expensive, I would get away cheaper with Astra ! ... but combination of multiple models might be a cost effective approach to try.
Origineel-
@SvenMeyer It could be complementary coverage, or just run-to-run variation. I'd compare two passes of one model with one pass each of two models, at the same total budget. Count newly fixed bugs on the second pass, not just the combined score.
Origineel
-
-
-
-
-
@PawelHuryn For finding bugs, does it make sense to first run Sol 6.1. Then run Opus 5.5 with the knowledge of known bugs from Sol. Can it lower the total cost while catching more bugs?
Origineel -
@PawelHuryn i used 6.1 for bug findings. it was much better than opus 5.5. but opus is much better in performance, design and smart code
Origineel -
@PawelHuryn 跟我在實際工作中使用6.1 sol xhigh 的體感一樣，有好又省！好像有點慢？
Origineel -
@PawelHuryn I would advise against testing every model on Max reasoning. Most tend perform better on Medium or High, like my test here where Terra and Luna both suffer on Max:
Origineel -
@PawelHuryn The real story is the price ratio, not the score — 44 bugs for $6.56 vs 45 for $33. At one-fifth the cost you can just run five passes and brute-force the gap. Makes the whole "Astra got shelved for safety" line look more like a pricing problem.
Origineel -
@PawelHuryn 6.1 sol’s performance for the price is honestly wild.
Origineel -
@PawelHuryn Great overview! Are you planning to also check gpt-6.1-luna?
Origineel -
@PawelHuryn it matters. this model ships far more slop than 6 sol did
Origineel -
@PawelHuryn Would be nice to see amount of slop and regressions introduced. Gemini for me was terrible for fixing things, it wasn't using good architecture, patterns and introduced a lot of regression. But the results may actually vary by how good your agentic context is in the workspace.
Origineel
스크롤하면 더 불러옵니다