[BUG] claude-sonnet-3.5 is better for coding tasks than 4x

and quiet honestly, i dont need 128k token outputs to prove it

And its all about personality, right? Congrats you have a next gen "conversational" model. Problems are its a f* psychopath, with junior level devops skills with chain of thought to code gen task completion. Thats what claude code is about, right? The complex problem it solves... simplictically. Code gen? Exactly what your coerced MAX subscribers want, right?!

@blik616287 Hi!
I understand your frustration with the debugging time difference (5 hours vs 10 minutes is indeed significant). Your concern about the shift in focus from pure code generation to more conversational capabilities is valid and worth discussing.
While I didn't have the chance to work with Sonnet 3.5, my experience with Sonnet 4, Opus 4.0, 4.1, and Sonnet 4.5 has been somewhat different - though I acknowledge that experiences can vary significantly based on use case and coding style.
A few thoughts and questions:

The "personality vs capability" trade-off: You raise an interesting point about the models becoming more conversational at the expense of direct code execution. Have you experimented with more directive system prompts or custom instructions that minimize explanations and focus on code output?
Context management: The 128k tokens might actually be part of the problem if not managed properly. Have you tried working with smaller, focused contexts or creating specialized sub-agents for specific tasks with clean contexts? Sometimes less is more.
Model-specific optimizations: Different approaches might work better with 4.x models:

More explicit step-by-step instructions
Breaking complex tasks into smaller chunks
Using specific formatting requirements in prompts

Alternative workflows: Since you mentioned Claude Code - have you tried using the checkpoint feature or different models for different stages (planning vs implementation)?

Could you share specific examples of tasks where 3.5 significantly outperforms 4.x? This could help identify if there's a pattern to the regression or if certain optimizations could help.
I agree that maintaining 3.5 access (or open-sourcing if discontinued) would be valuable for those who have optimized workflows around it. The community would benefit from having options rather than forced migrations.

its just not usable. every single task that i have run 3.5 outperformed 4.x. every single one.

its not about smaller / larger tasks. its horrible at even small tasks. even asking it to run a script with no previous context sees 4.5 change directory without an instruction to do so, run an entirely unspecified script. 4.5 cant handle stdout/stderr appropriately. i had a simple task bug yesterday that was removed. asked 4.5 to simply run a terraform plan on a directory. it returned no-op, when in fact the plan was set to create 50+ cloud resources. none of these issues are present in 3.x

i 100% refuse to believe that spending more time debugging or altering previously work prompts is the answer primarily because you cant seriously claim that 4.5 is better at coding when it take 5x the amount of time and tokens to achieve a comparable result.

Hi - You can continue to use Sonnet 3.5 in Claude Code by setting an env var (export ANTHROPIC_MODEL='your-model-name') or passing it as a param during startup (claude --model your-model-name)

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.