Anthropic appears to be A/B testing reduced effort levels in Claude Code
TL;DR Highlight
Reports say "high" effort in Claude Code reads as 10/100; Anthropic says it is a serving config test.
Who Should Read
A developer who uses Claude Code's effort setting and feels the model suddenly got dumber lately
Core Mechanics
- Author says it is a server-side experiment
- The author says it is server-side, not the app.
- Fable 5 sessions on Claude Code 2.1.236+ are enrolled in an experiment that shrinks the effort scale.
- Older versions and opus 5 are left alone, per the author.
- If "high" feels like "low", you're in the test group
- The author says it is probably an A/B test, so not everyone sees it.
- If "high" feels like "low" to you, you're in the test group.
- Since 2.1.237, high reads as 10/100
- The author says the model reads "high" effort as 10 out of 100.
- That is the exact number "low" used to be.
- The changelog says nothing about it, per the author.
- Author lost an afternoon blaming own apps
- They thought t3 code and their own app were broken.
Community Reactions
- Claude Code team member responded, per a comment
- Thariq reportedly said they test API serving configs in Claude Code, and one maps the numerical effort value differently.
- The scale isn't 0-100, the number isn't meaningful alone, and your selected effort is what you get.
- They say evals confirm no performance impact and offered credits for a clear regression sent via /feedback with the ID.
- Experience: Opus 5 overworked a simple task
- One commenter said a config file update took under 2 minutes on 4.6.
- On Opus 5 it took 43 minutes of pulling containers, running sandboxes, and building test suites.
- Experience: lower effort gave better cost and quality
- One user runs Opus 5 mostly at low effort with good results, while high goes off on unasked tangents.
- Another said most of Opus's extra cost over Sonnet comes from being chattier.
- They use Low reasoning instead of falling back to Sonnet, with better quality at comparable spend.
- Opinion: the model may not know its effort level
- Some think the agent harness may set the reasoning token budget.
How to Apply
- Run the same task at each effort level and log it
- Today, run the same prompt at low and high and compare time and token usage.
- Judge by output and cost, not by asking the model its effort level.
- Send the session ID via /feedback on regressions
- Run /feedback in Claude Code when you see a clear quality drop.
- Send the ID to Thariq, as the comment instructs.
- Log Claude Code version and symptoms
- Note the version (e.g. 2.1.236+) and model name with any odd response.
- Compare with teammates to see if you're in the test group.
Try It
# Prompt to paste into Claude Code (effort comparison check)
Do the task below and finish with a table of the following.
- Task: {{task description}}
- Scope: do not touch anything outside {{files you may edit}}.
Table columns: files read, commands run, number of steps, summary of final changes.
Next step: switch /effort between low and high, rerun the same prompt, and compare the tables.Terminology
Related Papers
Mercury 3: Diffusion at scale
Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality
Timestral: Mistral's time-series foundation models
A 209M time-series model ranks first on three benchmarks and runs on CPU.
DeepSeek V4.1 Flash 1017 (Max)
DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.
If you don't run your own LLM proxy, how do you know whats going on?
Running your own LLM proxy shows which keys, models, and cache misses drive your usage.
DeepSeek: DeepSeek v4.1 Flash 1017
DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents
DeepSeek-v4.1-Flash-1017
DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length
Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware