Shane Melaugh posted a claim on X on Nov. 7: Opus 5.5 had been nerfed. He said he ran Google's Gemini 4 Argon over 1,400 of his own Claude Code sessions. The tool flagged Nov. 3 as the day Anthropic's coding model started reasoning differently1.
Melaugh's proof was his own account alone. "The model name didn't change, but the way it does reasoning took a sharp turn," he wrote1. The post pulled 2,712 favorites and 287 replies within two days.
Shane Melaugh@shanemelaughis Opus 5.5 nerfed?
I just had Gemini 4 Argon analyze 1,400 Claude Code sessions and it found that on November 3, something definitely happened.
the model name didn't change, but the way it does reasoning took a sharp turn. pic.twitter.com/6xnofvg1rv
One Man's Count, Graded by Another Company's Model
Melaugh kept his prompts, his script, and his session logs to himself. His evidence is one description of what a second AI model told him about his own chat history1.
That second model was Gemini 4 Argon, Google's flagship reasoning system. Google released it Sept. 30, built for coding, security work, and long multistep jobs9.
Melaugh asked Argon to grade his own transcripts, apart from Anthropic's own records. Only his own account has shown the sessions, the prompts Argon used, or the rule it graded by.
One Flagged Reasoning Pattern Rose From 20% to 61.3%, by One Private Count
Share of Melaugh's sessions showing one Gemini-flagged reasoning pattern, before and after Nov. 3, 2026, per Adnaan's relay
Each column is the share of sessions carrying one reasoning pattern Melaugh's own analysis flagged; both numbers come from his private count, apart from any shared or published dataset.
These two figures come from one person's unpublished session count, relayed in a second person's post, apart from any independently run benchmark.
Sources: Shane Melaugh on X [1]; Adnaan on X [6]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Before Nov. 3 | 20% |
| After Nov. 3 | 61.3% |
Adnaan, an X user who amplified the thread, added a specific number the next day. One pattern, he wrote, "jumped from 20% to 61.3%, while others dropped sharply"6.
Melaugh's own count is the only source for that 61.3 percent figure. Anthropic, like everyone else, has kept the underlying data private.
The Tracker Built to Catch Exactly This
Nathan Langley had a tool running before Melaugh ever posted. Langley studies AI as a graduate student at the University of North Carolina. He built LiveNerf to catch a model quietly getting worse after it ships2.
The project stays simple on purpose. "A small, boring, append-only benchmark for one question: does a model get worse after it ships?" Langley wrote in its README2.
LiveNerf runs the same 78 questions through Claude Code every day. The questions come from GPQA Diamond, MMLU-Pro, and competition math. Langley chose them because Opus 5.5 answers them right about half to seventy percent of the time2.
Langley set his rule before he saw a single data point. A change counts only when two straight ten-day windows clear a 99 percent confidence bar. The swing also has to reach three points or more2.
Opus 5.5's Score Fell Lower on Nov. 5 Than on Nov. 3
LiveNerf's daily score for Opus 5.5 on its fixed 78-question panel, run through Claude Code, Oct. 28 to Nov. 8, 2026
Each point is the share of questions Opus 5.5 answered correctly that day; higher is better, and the lowest point of the stretch lands two days after the date Melaugh's claim names.
A 78-question daily panel swings several points on its own; Langley's own decision rule needs two full ten-day windows past a 99 percent bar before calling any of this a real change.
Source: GitHub [2]. Chart by Kelly Dennis.
The numbers behind this chart
| Period | Value |
|---|---|
| Oct. 28 | 60.3% |
| Oct. 30 | 57.7% |
| Nov. 1 | 70.1% |
| Nov. 2 | 60.3% |
| Nov. 3 | 56.4% |
| Nov. 4 | 62.8% |
| Nov. 5 | 50% |
| Nov. 6 | 62.8% |
| Nov. 7 | 57.7% |
| Nov. 8 | 65.4% |
Opus 5.5 scored 60.3 percent on Nov. 2, 56.4 percent on Nov. 3, and 62.8 percent on Nov. 42. Two days later, on Nov. 5, the score fell to 50 percent, lower than Nov. 3 itself.
The same panel read 57.7 percent on Oct. 30. It climbed to 70.1 percent on Nov. 1, its high for the stretch, then closed the run at 65.4 percent on Nov. 82.
Langley's own reading stayed calm after reviewing the week. "We haven't found convincing evidence of a regression in the behavior we've measured," he wrote3.
What Alarmed People Enough to Post Early
Theo Browne, who posts as Theo - t3.gg, raised a related doubt two days earlier. "This is a conspiracy theory but I'm like 95% sure I'm right," he wrote on Nov. 6. "Reasoning traces for Opus 5.5 read like a shitty quantized open weight model then SOMEHOW the right answer comes out in the end"4.
Browne's doubt pointed at the visible reasoning text, apart from the final answer. He suspected Anthropic had changed what Claude Code shows, apart from what the model actually computes.
Anthropic's own documentation already names one standing difference between the models. Opus 5.5 defaults to medium reasoning effort, a step below the high default Opus 5 used. That setting has held steady since the Sept. 22 launch8.
A six-week-old setting falls short of explaining one sharp jump on one day. It still shows how Claude Code's visible reasoning style can shift with a flag Anthropic controls. The weights underneath can stay untouched the whole time.
Anthropic's documentation lists five effort levels and other settings like it on that same page. Each one can change what a session looks like on the screen, apart from touching the weights underneath.
The debate reached Hacker News too, where commenters traded theories about A/B testing and quantization, apart from settling on one7.
He asked his own logs to settle it once Melaugh's post appeared. "Kick off a job in the fleet repo on bb-1 using opus 5.5 xhigh to investigate this for me," he wrote to Melaugh on Nov. 7. "Dig through my cliproxy logs in clickhouse to see if things changed"5.
Both posts point the same way. Something about what Opus 5.5 shows a user may have shifted, apart from whether its answers held up.
What to do
Check a tracker's method before trusting a claim that a model got worse. Langley published his rule, his questions, and his daily scores before he drew a conclusion. Melaugh's claim rests on one private session log instead.
Save your own transcripts if you suspect a coding tool changed. LiveNerf only works because every run gets logged the same way, every day. A single rough session settles very little on its own.
Read the visible reasoning text as closely as the final answer when you judge a coding agent's output. Browne's doubt and Melaugh's claim both point at what Claude Code displays on screen, a layer LiveNerf's own scoring leaves untouched.
By the numbers
- 1,400: Claude Code sessions Melaugh said Gemini 4 Argon analyzed for his claim1.
- 2,712: favorites on Melaugh's post, within two days of posting1.
- Opus 5.5 scored 56.4 percent on Nov. 3 on LiveNerf's panel, the date in Melaugh's claim2.
- Its lowest score that week landed two days later, at 50 percent on Nov. 52.
- Langley set his confidence bar for calling a change real at 99 percent, before he saw any data2.
- Adnaan said one reasoning pattern reached 61.3 percent after Nov. 3, up from 20 percent6.
What to watch
Watch whether Langley's panel clears his own bar across the next two ten-day windows. A real change needs two windows in a row, apart from one noisy day. Claude Code's own release notes are worth a second look too, since they skip Nov. 3 entirely as a date anything shipped8.
Sources
- Shane Melaugh, X post, Nov. 7, 2026, https://x.com/shanemelaugh/status/2119192075505893695
- Nathan Langley, "livenerf: Benchmark for tracking model capability after release," GitHub, accessed Nov. 9, 2026, https://github.com/ninjahawk/livenerf
- Nathan Langley, "Was Opus 5.5 nerfed? Looks like it was not, according to LiveNerf.," NinjaHawk's Nest, Nov. 4, 2026, https://blog.nathanlangley.dev/posts/opus-5-5-nerfed-livenerf.html
- Theo Browne, X post, Nov. 6, 2026, https://x.com/theo/status/2118634626679775549
- Theo Browne, X post, Nov. 7, 2026, https://x.com/theo/status/2119193502852755520
- Adnaan, X post, Nov. 8, 2026, https://x.com/adnaannasir_/status/2119415636032184803
- "Is Opus 5.5 nerfed? Thinking patterns changed on November 3," Hacker News, Nov. 9, 2026, https://news.ycombinator.com/item?id=50422069
- "Claude Opus 5.5," Claude Platform Docs, Anthropic, accessed Nov. 9, 2026, https://platform.claude.com/docs/en/models/opus-5-5/overview
- "Gemini 4 Argon: our next era of frontier intelligence," Google, Sept. 30, 2026, https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
