HN Debrief

Anthropic appears to be A/B testing reduced effort levels in Claude Code

  • AI
  • Developer Tools
  • Open Source
  • Startups

The post started from a screenshot suggesting Claude Code was reporting a lower-than-expected effort level, which sparked a broader complaint that Anthropic’s coding experience has become worse even when users pick the same model and settings. The strongest repeated claim was not “the slider number changed.” It was that Opus 5 and Fable now burn time and tokens on tangents, verbose explanations, and repo-wide overengineering for tasks that used to be quick and local. Several people said high and extra-high effort are the worst offenders, while low or medium effort often produce better coding results because the model stays focused instead of inventing work.

Do not treat a hosted coding model as a fixed tool just because the model name stayed the same. Track output quality, latency, and token burn in your own workflow, and keep a fallback stack ready because model providers are clearly changing serving behavior underneath users.

Discussion mood

Strongly negative toward Anthropic and especially Claude Code’s recent behavior. People are frustrated by verbosity, overengineering, unpredictable regressions, and the sense that serving changes, safety routing, or cost controls can alter the product without clear notice.

Key insights

  1. 01

    Low effort often beats high

    Higher reasoning settings appear to be the trigger for a lot of the bad behavior, not the cure for it. People using Opus 5 at low or medium effort reported better focus, lower spend, and fewer runaway tangents than high or extra-high, which suggests the current tuning can reward unnecessary exploration instead of task completion.

    Treat effort level as an empirical tuning knob, not a quality ladder. For coding tasks, start low and only raise it after you see a concrete need, otherwise you may pay more for worse output.

      Attribution:
    • onlyrealcuzzo #1
    • ricardobeat #1
    • perching_aix #1
  2. 02

    The real backlash is about silent experiments

    Anthropic’s explanation confirmed something many people already suspected. Claude Code users are being exposed to serving experiments before full rollout. Even if this specific remapping does not change capability, the bigger issue is that paying users were not told they were in an experiment and had no opt-out, which turns every unexplained regression into a trust problem.

    If you rely on hosted AI tools in production workflows, assume live experimentation is part of the service. Build monitoring and escalation paths around that reality, and do not promise internal teams that a vendor’s named model is behaviorally stable.

      Attribution:
    • trq_ #1
    • lobsterthief #1
    • cube00 #1
  3. 03

    Safety rerouting can look like model drift

    One commenter connected perceived quality drops to Anthropic’s own disclosure that some Fable requests can be rerouted to Opus 5 when biology safeguards fire. The concern is not just blocking. It is undisclosed fallback behavior that may degrade answers for legitimate work without clearly telling the user that a different model handled the request.

    If your use case touches regulated or sensitive domains, test whether safety systems are changing which model actually answers you. You may need separate validation and user messaging for any workflow where hidden rerouting could affect output quality.

      Attribution:
    • areoform #1
  4. 04

    API users see a different problem set

    Several comments separated Claude Code the product from direct API usage. The claim was that API behavior is easier to reason about because you control the request, can set spend limits, and can model token costs at scale, while chat and coding harnesses may add routing, policy, or serving-layer behavior that users cannot inspect. That makes anecdotes from Claude Code much less transferable to teams building directly on the API.

    Do not generalize from a consumer coding agent to the underlying model API. If model reliability matters, evaluate the API and the vendor harness separately because the harness may be the unstable layer.

      Attribution:
    • rrr_oh_man #1
    • taude #1
    • boredumb #1
    • Aurornis #1
  5. 05

    Token billing is manageable at scale

    The most useful pushback on billing complaints came from people running enough volume to know the difference between theoretical opacity and operational reality. Their point was that token costs are noisy per request but predictable in aggregate, and the major providers already expose enough knobs and telemetry to estimate spend, cap responses, and catch changes. The chaos is worse for interactive subscription products than for API-backed systems with instrumentation.

    If you are building on LLM APIs, budget from distributions, not single-request intuition. Instrument token usage early, and use provider limits and your own guardrails instead of waiting for pricing surprises to show up in the invoice.

      Attribution:
    • boredumb #1
    • sroussey #1
    • daishi55 #1
    • Aurornis #1
  6. 06

    Harnesses and model routing are now differentiators

    People who switched to Codex, GLM-5.3, DeepSeek, or multi-provider tools were not claiming one raw model had solved intelligence. They were praising faster routing, better defaults, and less need to fight the agent. That shifts the competitive battle from pure benchmark scores toward harness design, task dispatch, and when the system decides to use a cheaper or more focused model.

    When you evaluate coding agents, score the full stack, not just the flagship model. Routing logic, session management, and model selection policy can outweigh headline capability gains.

      Attribution:
    • rshnotsecure #1
    • surgical_fire #1
    • cynerx #1
    • deathmonger5000 #1

Against the grain

  1. 01

    The original A/B claim was likely overstated

    Anthropic’s team said the current test only changes how a numerical effort value is mapped and displayed, not the effort tier users selected, and another commenter noted that the viral framing took off partly because people were ready to believe the worst. That does not erase the larger dissatisfaction, but it does mean the specific screenshot evidence was weak support for a claim of deliberate effort reduction.

    Separate a bad product experience from a bad diagnosis of its cause. If you are escalating vendor issues internally, anchor the complaint in reproducible regressions rather than a screenshot the model generated about its own settings.

      Attribution:
    • enraged_camel #1
    • trq_ #1
  2. 02

    Competition limits token abuse theories

    A few commenters argued that deliberate token-padding makes little business sense in a market where users can switch providers quickly and compare cost-performance constantly. Their view was that model vendors are more likely struggling with tuning tradeoffs, capacity, or benchmark incentives than intentionally making outputs bloated just to extract more spend.

    Do not default to fraud as your operating assumption. You still need monitoring and vendor leverage, but the more useful question is which optimization target the provider chose, not whether they are cartoonishly trying to waste your tokens.

      Attribution:
    • Rudybega #1
    • gdudeman #1
    • hn_throwaway_99 #1

In plain english

Anthropic ↩
An artificial intelligence company that makes the Claude family of language models and products.
API ↩
Application Programming Interface, a defined way for software to expose functions or data to other software.
Claude Code ↩
Anthropic’s coding-focused command-line agent tool, used here through Emacs integrations or terminal sessions.
Codex ↩
An AI coding model name commonly used for systems that generate or edit code from prompts.
DeepSeek ↩
A Chinese AI lab and family of language models that compete aggressively on price.
evals ↩
Evaluations, usually repeatable test cases or benchmarks used to measure how well a model performs on specific tasks.
Fable ↩
A Claude model or mode referenced in the comments as a premium coding or reasoning option.
GLM-5.3 ↩
A version of the GLM language model family that commenters cited as an alternative coding model.
open-weight ↩
A model released with downloadable parameters, so others can run it or build on it, though the full training code and data may still be closed.
Opus 5 ↩
A higher-end Claude model variant discussed here as a coding model with adjustable effort levels.

Reference links

Vendor statements and documentation

Alternative tools and harnesses

Market and ecosystem references

  • OpenCode data page
    Referenced as anecdotal evidence that cheaper open-weight model releases can quickly shift user traffic and token usage.