Short answer
The quickest way to check whether the Claude you're paying for has been "nerfed" is the pelican test: ask the model to draw a pelican riding a bicycle in SVG and compare the result with what a genuine Claude produces. An endpoint that quietly swaps in a smaller model or cuts the thinking budget draws a visibly worse pelican. TeamoRouter runs the same prompt every 30 minutes on Claude Opus 5.5, Sonnet 5.5, Fable 5.1 and the GPT-6 family, and publishes every result on the pelican benchmark page, so you have a reference to compare against.
The pelican test only measures one thing: capability. Whether your cache works, whether thinking gets truncated and whether first-token latency is unreasonable show up in the API response instead. Each is covered below.
What a nerfed Claude looks like
The usual reports: answers to the same question get shorter and lazier; thinking is enabled but the response contains no thinking, or a high thinking budget still comes back in seconds; long conversations get more expensive because cache reads stay at zero; even simple prompts take tens of seconds to start streaming.
The cause is usually the endpoint, not the model. Some route requests to a cheaper model, some lower the thinking budget, some prepend their own hidden system prompt, and some don't keep sessions sticky, so the cache never matches.
What the pelican test is
Simon Willison has been running this test since 2024: every time a new model comes out, he asks it to "generate an SVG of a pelican riding a bicycle". Pelicans and bicycles are both hard to draw, and combining them takes code, geometry and spatial reasoning at once, so the differences between models are obvious at a glance. His pelican-bicycle repository keeps results from successive model generations.
Our page uses an animated version: an HTML page with an SVG 2D animation of a pelican riding a bicycle. The wheels have to turn and the legs have to pedal, which widens the gap between strong and weak models.
How to check your own endpoint
- Run the same prompt on the same model through the endpoint you use.
- Open the pelican benchmark page and compare with a few results from the same model around the same time: is the pelican's body, beak and wings complete, is the rider attached to the bike, do the wheels turn?
- Don't judge from one image. The same model draws better and worse pelicans from run to run, so run it two or three times and look at the overall level.
One caveat: some endpoints recognize well-known test prompts and return a canned answer for them. Change a few words in the prompt, or swap the pelican for another animal, when you test.
What the pelican test can't catch
Trimmed thinking: a request with extended thinking enabled should come back with thinking content. If it's always empty or a sentence or two, or the maximum thinking effort still returns in a few seconds, the thinking budget is probably being cut.
Cache that doesn't work: from the second turn of a conversation on, cache_read_input_tokens in the response's usage should make up most of the input. If it stays at zero, every turn re-reads the context at full price, which is both slower and more expensive.
Hidden system prompts: in a fresh conversation, ask what instructions it received before your message. This isn't reliable every time, but if it mentions an identity or rules you never set, be suspicious.
First-token latency: long thinking is allowed to take a while. A simple prompt with thinking off that still takes tens of seconds to start is a sign to switch endpoints.
What we're doing about it
From October 20 to 31, TeamoRouter is running a Claude Crackdown: one Claude experience improvement ships every day, and on any day we miss, all Claude users share a US$20,000 credit.
Day 1 published an audit of dozens of live Claude routes and 13 candidate suppliers: 2,187 test requests and more than 57 million tokens across real-session replays, blind ratings, reasoning and long-context retrieval tasks, parameter-compatibility checks and thinking-signature round-trips. Five suppliers failed (trimmed thinking, injected hidden prompts, first-token timeouts and more); three were removed and the two candidates won't be onboarded. From Day 2, the platform covers 80% of cache-rewrite charges caused by upstream fluctuations, deducted automatically at billing.
Daily progress is on the TeamoRouter home page, and the pelican results keep updating.