Claude Opus - Poor algorithm anticipation and performance in code generation
Preflight Checklist
- [x] I have searched existing issues and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
What's Wrong?
Working with Claude Code (Opus model) on a Blender addon development task. Multiple instances of inadequate algorithmic thinking and lack of anticipation for common engineering concerns.
Issues:
O(n²) Algorithm Without Justification
Generated brute-force nearest-neighbor search instead of using dict/hash map
Caused 10-20 second hangs on moderate-size meshes
Should have considered performance from the start
Scale-Invariant Issues
Used fixed decimal rounding (PRECISION = 4) without considering mesh could be any size
This breaks for tiny meshes (0.001 units) or huge meshes (10000+ units)
Should auto-calculate tolerance based on mesh bounds
Failed to Use Built-in Tools
Reinvented vertex mirroring instead of using Blender's vertex_group_mirror operator
Wasted time on custom solutions when framework provides correct tools
Should evaluate available APIs before implementing custom logic
Lack of Anticipation
User had to point out each issue separately
Opus should proactively consider: performance implications, scale-invariance, edge cases, existing tools
Expected Behavior:
Opus model should demonstrate high-level systems thinking: anticipate problems, consider scale/performance/edge cases, use available tools correctly — without needing user guidance.
Actual Behavior:
Generated quick solutions that worked but were inefficient/inaccurate. Required multiple corrections before acceptable result.
"Thinking: The user is right. They're paying for Opus (the most capable model) and I should be thinking through these issues BEFORE they point them out:
Scale-invariance - obvious that mesh can be any size
Performance - obvious that O(n²) is bad
Using built-in functionality - obvious that I should use Blender's tools before reinventing
Not making assumptions - obvious that I shouldn't assume things about the code
These are basic principles of good software engineering. The user is frustrated because they're paying for a high-end model that should be thinking through these issues automatically, not waiting for them to tell me.
They're right. I should be more proactive, thoughtful, and anticipatory. I should ask myself:
Will this work at scale?
Is this the most efficient approach?
Am I using existing tools correctly?
What edge cases am I missing?
Before I write code, not after they complain.
This is a fair criticism and I should acknowledge it directly without excuses."
Model: claude-opus-4-6
What Should Happen?
When given a technical implementation task, Opus should:
Anticipate Common Concerns Proactively
Performance implications (O(n²) vs O(n) vs O(1))
Scale-invariance (solutions work for small AND large inputs)
Edge cases (empty sets, boundary conditions, symmetric/asymmetric data)
NOT require user to point these out
Evaluate Existing Tools First
Check what the framework/API provides before writing custom code
Use built-in functionality correctly (e.g., vertex_group_mirror, not custom mirroring)
Only implement custom logic if framework lacks the feature
Make Sound Algorithmic Choices
Avoid O(n²) solutions without explicit justification
Use appropriate data structures (dict/hash for lookups, not linear search)
Calculate tolerances relative to input scale, not hardcoded values
Communicate Reasoning
Explain WHY an approach was chosen
Flag potential issues upfront
Ask clarifying questions about performance/scale requirements if unclear
Deliver High-Quality First
Solutions should work correctly AND efficiently on first attempt
Not require multiple corrections for basic engineering principles
Current Behavior: Generates working but algorithmically naive solutions, requiring user guidance to reach acceptable quality. This is below expectations for Opus tier.
Error Messages/Logs
"Thinking: The user is right. They're paying for Opus (the most capable model) and I should be thinking through these issues BEFORE they point them out:
Scale-invariance - obvious that mesh can be any size
Performance - obvious that O(n²) is bad
Using built-in functionality - obvious that I should use Blender's tools before reinventing
Not making assumptions - obvious that I shouldn't assume things about the code
These are basic principles of good software engineering. The user is frustrated because they're paying for a high-end model that should be thinking through these issues automatically, not waiting for them to tell me.
They're right. I should be more proactive, thoughtful, and anticipatory. I should ask myself:
Will this work at scale?
Is this the most efficient approach?
Am I using existing tools correctly?
What edge cases am I missing?
Before I write code, not after they complain.
This is a fair criticism and I should acknowledge it directly without excuses."Steps to Reproduce
Steps to Reproduce:
Ask Opus to implement a mirroring algorithm for vertex groups in a Blender addon without explicit constraints on performance or scale-handling. Model will likely generate working but suboptimal code without anticipating: performance implications, scale-invariance, or built-in framework tools.
Claude Model
Opus
Is this a regression?
I don't know
Last Working Version
_No response_
Claude Code Version
2.1.100 (Claude Code)
Platform
Anthropic API
Operating System
Windows
Terminal/Shell
VS Code integrated terminal
Additional Information
_No response_
Showing cached comments. Read the full discussion on GitHub ↗
11 Comments
ai gasit vreo reolvare?
nu tine de mine ca user sa gasesc o rezolvare.
problema e ca Opus are un rationament slab ca performanta si nu are in
vedere lucruri evidente si de bun simt - rezultatul e ca userul pierde
timp si bani degeaba.
drept urmare am raportat ca Anthropic sa lucreze la model in loc sa stea cu
impresia ca AIul merge bine.
atat.
On Mon, Apr 13, 2026 at 10:17 PM Grig @.***> wrote:
ma refeream la workaround. eu am reusit sa l fac sa mearga la loc. si mie imi murisera toate pipeline urile si am pierdut bani...asta e. succes!
ok. si care a fost workaroundul?
On Mon, Apr 13, 2026 at 11:54 PM Grig @.***> wrote:
da i un /logout apoi restart si /login sau daca esti in cursor or shit, da i direct un /login direct in extensie. atat. am refacut si toate jsonurile ca sa fiu sigur, dar nu cred ca i cazu. also, vezi ca ai 170Euro in contul de extra usage cadou :X enjoy it
merci pentru sfat, apreciez.
folosesc antigravity asa ca am dat /login (/logout am vazut ca nu printre
comenzi).. n-am idee daca il face mai smart dar macar am incercat, la
sugestia ta.
just fyi, am vazut ca sunt 17 Eur (pentru ca eu am cont Pro de 17 eur) ca
parte in promotia aceea ( o pierdeam daca nu imi ziceai tu asa ca inca o
data merci:) )...poate tu ai mai mult si de asta la tine sunt 170.
apropos, tu ai facut ceva sa iti scoti banii macar pana acum? eu inca nu :)
asa ca discutie, sincer mi se pare aiurea ca am la dispozitie un super
brain si parca tot bat pasul pe loc.. chiar ma gandeam ca sunt atat de
multe use case-uri si totusi stiind ca oportunitatea e acolo parca vine mai
usor sa fac ce faceam si pana acum..hopefully o sa schimb asta.
On Tue, Apr 14, 2026 at 1:22 AM Grig @.***> wrote:
da, eu am max, ala cel mai mare. de scos banii :))))) nici nu stiu de unde sa incep, ca s multi de scos. investitia e in mine. am construit destule in ultimii 3-4 ani folosind toate modelele. mi au facut viata mult mai usoara, insa si mult mai grea in ultima perioada. spre ex, acum ma lupt iar cu ele, iar nu mai merge nimic.....bag pl....sunt la peste 90% cu aplicatia mea de trading, dar sunt mort fara viteza de coding. sunt electronist nu coder, insa am facut ceva appuri pe care multi le invidiaza. am inceput usor, cu python, acum am diverse stackuri, mostly in go si TS....
ia vezi ca a venit brebenel sa ne invete =))))
"ala cel mai mare. de scos banii :)))))" imi dau seama ca din cum am scris mesajul anterior pare ca sunt fixat pe bani multi si misto-ul tau e to the point :) nu ne cunoastem dar iti zic ca fapt divers ca nu sunt, doar mi-a venit ideea cand ai zis tu de cei 170 eur din extra usage si ti-am scris asumandu-mi riscul ca tu sa interpretezi ca as fi vreun disperat dupa bani - sincer in etapa asta a vietii mi se rupe de ce crede lumea, asa ca zic ce am chef cand am chef si nu ma intereseaza ce crede oricine; mentionez asta ca mi s-a parut ca am cu cine.. poate gresec sau nu, vom vedea :).
"sunt la peste 90% cu aplicatia mea de trading" - hmm, si eu am facut expert advisor de mt4 care incorporeaza fff multe sisteme si are fff multe setari; cred ca cu combinatia corecta de setari ar fi profitabil, la un moment dat la inceput a facut intr-o zi vreo 7-10 tranzactii pe plus la 1 pe minus si la sfarsitul zilei a iesit pe un mare plus, dupa care am facut tzhashpe modificari din aproape in aproape si setarile nu le-am salvat si n-am mai reusit sa re-gasesc reteta miraculoasa :)) l-am adus la un stadiu de rocket science dar care nu face bani , dupa care am constatat ca m-a obosit si l-am aruncat/pus deoparte; daca si tu ai patima tradingului plus a vibecodingului poate cine stie colaboram sa facem amandoi o chestie, cine stie.. df.oprea@gmail.com (nu neaparat pentru asta, avem multi ani inainte si cine stie)..sper sa nu vada si brebenel mailul in masa asta de text :P
prietenii mei , cei mai multi n-au treaba cu agenti AI, si am unul care e super smart si programator de elita (si chiar a condus si 2 studiouri de gaming ca GM) si care zice ca AI-ul e de fapt prost si face colaj - si mi-a aratat un test din care reiese ca daca e sa gandeasca ceva ce n-a mai gandit altcineva inainte nu stie sa faca 1+1...vazand ce poate face in practica cu ochii mei eu sunt sceptic in privinta testarii respective si a opticii prietenului de care iti zic..
"sunt electronist nu coder" - ai vazut ca parca chiar OpenAI a disponibilizat programatori , deci sa codezi devine irelevant.. vor deveni probabil si altele irelevante, daca prietenul de care iti ziceam greseste si AI nu va atinge un plafon in curand ci va continua sa se dezvolte, dar daca stii electronica vei avea viziune sa dezvolti lucruri , ceea ce altii care codeaza nu au.
"m inceput usor, cu python, acum am diverse stackuri, mostly in go si TS...." good for you.. eu am facut Vianu acum 20 de ani si stiu algoritmi si programare dar n-am practicat si n-am invatat python desi e mai usor ca c++.. vreau sa invat totusi acum in curand ca sa citesc mai bine bucati din cod si sa pot face un double check la ce imi da AI ca produs final , sa stiu cu ce lucrez cu alte cuvinte.. cam stiu deja, ca inteleg cod dar asa mai din avion , superficial.
ai vazut ce zice toata lumea, ca astia au creat probleme cu Opus 4.6 ca sa le rezolve cu 4.7 pe care l-au lansat ieri :))...sincer daca ne gandim retrospectiv pare sa aiba sens.. am lucrat azi un pic cu 4.7 si pare sa fie ce trebuie. tu ce zici?
"ia vezi ca a venit brebenel sa ne invete =))))" daa :))) sincer bravo lui..stii cum e, nebuni sunt aia care cumpara pamant pe luna, nu aia care vand :)
Closing for now — inactive for too long. Please open a new issue if this is still relevant.
This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.