Do Claude Get Quietly Nerfed After Launch..(We Have Proof)

Is Claude getting worse, or is Claude Opus 5.5 just thinking less than the benchmark table says it does? Every week someone posts proof that Claude got nerfed or ChatGPT got dumber: a launch-day screenshot next to today's, a Reddit thread at 2,700 upvotes, a NerfBench reading of 94.2%. This video goes through the actual evidence on Claude Opus 5.5, Claude Code, GPT-6 Astra and Codex, and finds the degradation in a different place than the screenshots do. Anthropic's Opus 5.5 launch page runs every benchmark at max effort. The API and Claude Code both default to medium, two levels lower, and Anthropic's own docs say a request that leaves effort out "runs one level lower than it did on Claude Opus 5." On CursorBench that is 57.8% in the headline and 52.5% at the default, on the same page. LiveNerf, the most careful daily tracker, can only detect a 6.6-point drop, and moving effort from high to medium costs 4.2 points, so the score barely moves. The output tokens drop 26%, which is why the LLM benchmarks keep missing what users feel. Then the paper trail. Four written postmortems from Anthropic and OpenAI, ten documented incidents, from the August 2025 routing bug (16% of Sonnet 4 requests at the worst hour) to the March 2026 Claude Code effort change, the January 2026 thinking-time cut, and the three GPT-6 Astra faults OpenAI fixed on September 12. None of them touched the model weights. Marginlab's Claude Code tracker caught a five-day drop from 65% to the low 50s that tracked the Claude Code version, not the model. And on a subscription, the tokens a lower default saves go to the lab. WHAT THIS VIDEO COVERS - Why a launch-day vs today screenshot cannot detect a nerf, and what LiveNerf, NerfBench and Marginlab measure instead - The Opus 5.5 footnote: max effort on the chart, medium by default, 5.3 points apart on CursorBench - How "40% cheaper than Opus 5" splits into a 20% price cut and a quarter fewer tokens per task - The ten admitted degradations across Claude Code, the Claude API, ChatGPT and Codex, and why none was quantization - Safety fallbacks that move a Claude Code session to Opus 4.8 at medium effort, and the setting that stops it - Model degradation vs harness bugs: the May 2026 Marginlab catch and the tool-call spike that gave it away - Three numbers to log per task so you can tell a default change from a real regression CHAPTERS 00:00 The feeling that your model got dumber 02:28 The screenshots: Astra, Opus 5.5 and NerfBench 03:32 LiveNerf: accuracy is noisy, tokens are not 05:06 Max on the chart, medium by default 08:01 Where the 40% saving comes from 10:50 Ten admitted nerfs, zero weight changes 13:07 When a different model answers you 14:00 Marginlab: when the harness is the nerf 16:25 Why your side-by-side proves nothing 18:21 Check the dial before you blame the model 20:53 The weights never changed, the tokens did SOURCES Claude Opus 5.5 launch: https://www.anthropic.com/news/claude-opus-5-5 Effort docs: https://platform.claude.com/docs/en/build-with-claude/effort Model IDs and versioning: https://platform.claude.com/docs/en/about-claude/models/model-ids-and-versions Pricing: https://platform.claude.com/docs/en/about-claude/pricing Claude Code model config (defaults, fallback): https://code.claude.com/docs/en/model-config Anthropic postmortem, Sept 2025: https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues Anthropic postmortem, April 2026: https://www.anthropic.com/engineering/april-23-postmortem ChatGPT release notes: https://help.openai.com/en/articles/9624314-chatgpt-release-notes GPT-6 Astra fixes: https://yellow.com/news/openai-defects-gpt-6-astras-sudden-slump ChatGPT Pro pause: https://runtimewire.com/article/openai-pauses-200-pro-subscriptions-as-astra-strains-capacity Claude Code weekly limits: https://getcoai.com/article/anthropic-called-it-a-25-raise-claude-code-users-get-17-less/ LiveNerf: https://github.com/ninjahawk/livenerf NerfBench: https://bridgebench.ai Marginlab Claude Code tracker: https://marginlab.ai/trackers/claude-code Marginlab, May 2026: https://marginlab.ai/blog/claude-code-degraded-before-opus-4-8/ Zapier AutomationBench: https://x.com/wadefoster/status/2102436862501813983 LLMs Get Lost In Multi-Turn Conversation: https://arxiv.org/abs/2505.06120 Is Claude nerfed, is ChatGPT getting worse, is Claude Code getting worse, Opus 5.5 nerfed, Codex nerfed. The answer is the same for all of them: the model ID is pinned, the effort, the router and the harness are not. Measure tokens, tool calls and the app version before you switch. #claude #opus55 #claudecode #anthropic #gpt6astra #llmbenchmark