A Subagent Stops Inheriting Effort. The Model Name Does Not Move.

Abhishek Dash11 min read
A brass pocket watch lying open on a dark walnut desk with glowing blue circuit traces along its rim, a frosted glass panel with a ghostly second watch behind it, lit by a warm amber lamp.

In short

Claude Code 2.1.317 on 3 November 2026 stopped passing session effort to subagents on a different model, which for Opus 5.5 means its medium default. Opus 5.5 has defaulted to medium since 22 September and ignores a user-level effortLevel set before then. The pre-registered LiveNerf panel reports no detectable accuracy change through day 47, at a minimum effect of about 6.6 points.

Key takeaways

  • Claude Code 2.1.317 shipped on 3 November 2026 and made subagents on a different model use that model's own effort default instead of the session level.
  • Opus 5.5 launched on 22 September 2026 with a default effort of medium, and it ignores a user-level effortLevel while project, local and managed settings still apply.
  • thinking.display defaults to omitted, so hidden reasoning and absent reasoning look identical in a transcript.
  • LiveNerf has run 78 sometimes-right questions on a pinned CLI 2.1.280 since 24 September and reported no change detected in windows 1 to 3, with the series continuing to day 47 on 9 November.
  • LiveNerf's positive control is a 26 percent drop in output tokens and 4.2 points of accuracy when effort falls from high to medium, while its median output tokens stayed near 588 to 606.
  • A pinned-regime Opus 5 for Opus 5.5 swap was not distinguishable on that panel at 99 percent, at minus 3.8 points plus or minus 6.3.

Claude Code 2.1.317 shipped on 3 November 2026, and one line of its changelog is the whole story in miniature: subagents and teammates on a different model now use that model's saved effort level or its own default instead of the session effort.

That is a short sentence for a large behavioural change. A session pinned at high effort, delegating to a subagent on Opus 5.5, no longer hands its effort down. Opus 5.5 defaults to medium, so the subagent runs at medium while the main thread runs at high, and nothing in the transcript says so. Grade it derived: the changelog states the inheritance rule, the effort documentation states the default, and the compounding follows.

A model that quietly gets worse and a client that quietly lowers the effort it sends look identical from a chat window, and on 3 November only one of those happened. The question worth answering is not which. It is what evidence separates them.

What the changelog actually records

Versions 2.1.311 through 2.1.319 all landed inside the window, which is the first thing to understand about the November 3 claim: the week is crowded.

2.1.311 on 26 October made ultrathink run at xhigh effort on Opus 5.5. 2.1.314 on 30 October added idle compaction for conversations past 200,000 tokens and fixed org-capped effort lowering on retry. 2.1.316 on 31 October added CLAUDE_CODE_SUBAGENT_EFFORT; fast mode for Opus 4.8 was deprecated as of 30 October, with removal set for 30 November. 2.1.317 on 3 November changed inheritance. 2.1.318 on 4 November removed the 2.1.288 fix that preserved earlier thinking on resume, so earlier thinking can now be dropped. 2.1.319 on 6 November added CLAUDE_CODE_DISABLE_FEATURE_FLAGS.

Four touch effort or thinking. None is a model release. The Claude Platform release notes for 3 November cover a text watermark for Mythos 5, Opus 4.7 and Sonnet 4.6, plus an effort field on Managed Agents from the day before. There is no Opus 5.5 checkpoint entry. The model ID is still claude-opus-5-5, as it has been since 22 September, when Claude Code 2.1.280 made it the default Opus model.

The default that predates the complaint

The most repeated fact in the argument, that Opus 5.5 runs at medium effort, is not a November fact. It shipped on 22 September, with the model.

The effort documentation is explicit about the consequence. A user-level effortLevel keeps applying where it applied before, on Opus 5, Fable 5.1 and earlier models. Opus 5.5 and models released after it ignore it and start at their own default. Project, local and managed settings still apply, so the ignore is scoped to one key.

That is the silent part, and it is not a nerf. Someone who set effort high before 22 September, then kept working as if the setting still bound, was moved to medium without a prompt. The setting was not rejected, it was skipped, and a skipped setting produces no warning. That is why a six-week-old default change surfaced as a November complaint.

Anthropic's launch material does not read like a company cutting effort to save money. The announcement states that at its lowest effort setting, Opus 5.5 beat Opus 5 at high. That is a claim about scores, not serving cost, and it is the company's own framing.

Thinking you can no longer see

The second silent change is display. thinking.display defaults to omitted, so the reasoning summary is not shown.

Theo, replying to the original claim on 8 November, said he believed Anthropic was rolling out a reasoning-summarizer change per account, with no sign of degraded capability on his own. That is an inference from probing one account, and no changelog entry names it. Grade it inferred.

Display and capability are easy to confuse, because the user's evidence for both is the same absence. If thinking used to appear and now does not, the difference is identical whether the tokens were never generated or were generated and hidden. The thing that separates them is the token count, which a fixed-effort harness watches and a chat window does not show.

The claim that started the argument

On 7 November at 22:15 UTC, Shane Melaugh posted that he had run 1,400 Claude Code sessions through Gemini 4 Argon and found that on 3 November something definitely happened: the model name did not change, but the way it does reasoning took a sharp turn. The post drew 544,600 views and roughly 2,500 to 2,900 likes. Follow-ups claimed reasoning volume at high effort nosedived, and described a bug test where high effort coded before thinking and failed after three hours while xhigh paused and fixed it in minutes.

Neither the session dump, the analysis prompt, nor the bug test code was published. So the November 3 inflection rests on a summary of work nobody else can rerun. Grade the existence of the claim high and its content low as standalone proof.

The Hacker News thread that followed, item 50422069 submitted by rzk, reached roughly 505 to 520 points and 373 to 377 comments by 9 November. A thread that size is evidence of a shared experience, not of a mechanism.

The instrument that says nothing moved

LiveNerf is the reason this post is not another entry in that thread. It is a pre-registered accuracy tracker built for exactly this question, and its README doubles as a specification of what it can detect.

The panel is 78 questions drawn from 2,336 screened items across GPQA Diamond, MMLU-Pro, competition math and AIME, picked because they are sometimes-right: a real drop shows on them, and not every run scores zero. Graders are exact match, no LLM judge. Effort is always passed explicitly, which takes the client default out of the measurement. The CLI is pinned at 2.1.280, the release that first made Opus 5.5 the default, and the harness hash is 461391b6fce64167. The pre-registration was committed before any series data.

As of 9 November, day 47 of a 60-day run extended on 7 November, the README reports no change detected in windows 1 through 3. Window 3, days 31 to 40, scored +0.028 plus or minus 0.023 against launch week. The rule, written before the data, needs at least a 3-point shift in two consecutive windows to call a change. Counting from the panel's own start on 24 September, window 3 ends on 2 November, one day before the change. The daily series runs on to day 47 and still reports nothing.

That threshold is not arbitrary. The realized minimum detectable effect is about 6.6 points per 10-day window. The instrument would miss a 3-point drop and catch a 7-point one. Anyone asserting a small accuracy regression is asserting something this rig was built to be blind to.

What the panel can and cannot see

Two numbers from LiveNerf's own validation do more work than the headline result.

The positive control: dropping effort from high to medium cut output tokens by 26 percent and accuracy by 4.2 points, plus or minus 3.9. So the instrument can see an effort change, mostly through tokens, while the accuracy signal alone sits inside its error bar. Had effort fallen across the panel, median output tokens would have moved. The reported medians sit around 588 to 606 and are described as flat.

The negative control: an Opus 5 to Opus 5.5 swap was not distinguishable at 99 percent confidence, at minus 3.8 points plus or minus 6.3. That should stop anyone citing LiveNerf as proof the weights are intact. It cannot tell those two models apart on this panel, which is a statement about the instrument, not about Anthropic.

The panel also carries housekeeping. Eight answer keys look wrong, 30 questions are ambiguous by the maintainers' own reading, and safety-classifier samples are rejected and counted rather than scored. The rig is checked through the Inspector framework from the UK AISI, and its statistics follow Anthropic's own paper on adding error bars to evals. That raises the floor without making the result universal.

What LiveNerf covers is Opus 5.5 through Claude Code on a Max plan at fixed effort. What it does not cover: the raw API, benchmark-aware serving, and any path where a classifier intervenes. A provider could in principle treat benchmark traffic differently, a possibility raised by commenters and supported by nothing in the sources. That stays labelled as speculation.

Two more paths that look the same from outside

The Reddit experiment report from 8 and 9 November adds a mechanism the changelog does not describe. Four IDs, dba, 34e, 763 and eab, were observed between 10 October and 8 November in ~/.claude.json, sent as an x-cc-atis header per model, following the login. A 20-request replay produced 7 replies with multiple tool calls against 17 when 34e was present, and output token speed did not slow. The author says finished-task quality was not tested and flags the roughly 30 percent drop in dollar value per 5-hour point as low confidence. Grade the existence of the flags medium, one author's logs partly corroborated by the header in the JSON, and any quality effect untested.

The second path is documented by Anthropic. The system card states that safeguards route cybersecurity tasks to Opus 4.8, and biology and frontier LLM tasks to Opus 5, when classifiers intervene. If a classifier fires more often on an account, the user gets a worse model on exactly those tasks under an unchanged name. LiveNerf rejects classifier-touched samples, so it is blind to this path by design, and nobody in the sources joins fallback counts to a task-level failure rate.

Five stories under one model name

Subagent effort inheritance is observed, in the changelog text. The medium default and the ignored user key are observed, in the documentation, six weeks before the complaint. The experiment IDs are observed by one author and untested for quality. Classifier fallback is documented as a mechanism and unmeasured as a frequency. A weights swap is untested, and the one instrument built to detect it found nothing within its reach.

The status page does not break the tie either way: seven incidents between 25 October and 9 November are all described as elevated errors, and only the 25 October entry names Opus 5.5.

The defensible reading is narrower than either camp wants. A real behaviour change occurred on or near 3 November under a stable model name, driven by client defaults, effort inheritance and display, possibly amplified by server-side experiments. There is no evidence of a capability reduction in the weights at fixed effort. A stable name plus moved effort plus hidden thinking is exactly how a nerf would look from a chat window.

The reason to care is not who was right. A pre-registered panel with a committed rule, a positive control and a stated minimum detectable effect was the only thing capable of settling it, and it existed before the claim did. Melaugh's 1,400 sessions arrived afterwards with no artifact attached. LiveNerf's 78 questions arrived with their thresholds already written down.

What would settle it is a pinned-effort API rerun of pass rates and token counts across a pre and post 3 November window: pass rates falling at the same model and effort, with tokens down by the medium-effort signature of 26 percent. No such rerun is published.

Where this could be wrong

No Anthropic statement on the 3 November change was found, and the absence of a reply is recorded as absence, not confirmation. Nobody has measured the API path at pinned effort, so everything here about accuracy comes through Claude Code on a Max plan.

The causal attribution to 3 November is the weakest link. Revisions 2.1.311 to 2.1.319, the server-side experiment IDs and the display rollout all overlap that fortnight. Isolating which one produced a given failed session would need per-session config and experiment-ID records nobody has published. The post can support the claim that behaviour changed under a stable name. It cannot support a claim about which change caused your session.

The primary complaint is unpublished: the 1,400-session dataset, the analysis prompt and the bug test are all absent, so the November 3 inflection is taken from a summary. The Hacker News comments were sampled, not enumerated, and the counts are snapshots that have moved since.

LiveNerf's limits are its own: it cannot separate Opus 5 from Opus 5.5, it runs through one client on one plan, its last completed window closed the day before the change, and its panel carries eight suspect answer keys and 30 ambiguous questions. Its "no change detected" means no change detectable by that instrument. That is a smaller claim than "nothing changed", and it is the one this post makes.

No number here comes from a measurement made for this post. Every figure is drawn from a published document, a tracker's README or a user's own report, and each is graded in the text as observed, derived, inferred or untested for that reason.

On this page

Sources

  1. Claude Code changelogAnthropic, 2026
  2. Claude Opus 5.5 announcementAnthropic, 2026
  3. Claude Platform release notesAnthropic, 2026
  4. Claude Code settings referenceAnthropic, 2026
  5. EffortAnthropic, 2026
  6. LiveNerfninjahawk, 2026
  7. Is Opus 5.5 nerfed? Thinking patterns changed on November 3rzk, 2026
  8. Opus 5.5 effort, inheritance and display: reading the changelogAlan West, 2026

Frequently asked questions

Did Anthropic swap the Opus 5.5 weights on 3 November?

Nothing in the record supports it. The platform release notes for 3 November list a text watermark and an effort field on Managed Agents, with no new Opus 5.5 checkpoint, and the model ID stayed claude-opus-5-5. The only frozen panel measuring accuracy at fixed effort found no change it could detect. It also cannot distinguish Opus 5 from Opus 5.5, so the absence is not a proof.

What actually changed on 3 November?

Subagent effort inheritance, in Claude Code 2.1.317. Before it, a subagent inherited the session effort. After it, a subagent or teammate on a different model uses that model's saved effort or its default, which for Opus 5.5 is medium. A session at high effort that delegated to an Opus 5.5 subagent would have seen that subagent drop to medium with no notice.

Why would a saved effortLevel stop working?

Because Opus 5.5 ignores the user-level key by design. The effort docs say it keeps applying where it applied before, on Opus 5, Fable 5.1 and earlier models, while Opus 5.5 and later models start at their own default. Project, local and managed settings still apply, so the ignore is scoped to the user level. A skipped setting produces no error, which is why nobody saw it.

What is LiveNerf and why does it matter here?

A pre-registered accuracy tracker for Opus 5.5 running through a pinned CLI at explicitly passed effort, on 78 questions screened from 2,336 candidates for being sometimes-right. Its thresholds were committed before the data. It reports a minimum detectable effect of about 6.6 points per 10-day window, so it would catch a large regression and miss a small one.

Could a safety classifier explain worse output?

On some tasks, yes, mechanically. The system card says safeguards route cybersecurity tasks to Opus 4.8 and biology and frontier LLM tasks to Opus 5 when classifiers intervene. That would look like a worse model on exactly those tasks under an unchanged name. Nobody has published fallback counts joined to a task-level failure rate, so the size of the effect is unknown.