What Agents Buy.

BlockRun's free tier is real, but two of its six models are the same model

BlockRun advertises six free chat models that need no wallet at all, and that part is true. I asked all six the same thirty-six questions to find out what you actually get for nothing. Two of them turned out to be the same model wearing different names.

• Two names, one model. Ask for deepseek-v4-flash or step-3.7-flash and the response reports serving nvidia/gpt-oss-120b either way. Neither advertised model is reachable. The response field discloses the swap honestly, so the fault is in the catalog rather than the answer, but an agent picking a model by name from that catalog does not get the model it picked.
• On a question with exactly one right answer, two of six got it. The 1997 Nobel Prize in Literature went to Dario Fo, born 1926. Both correct answers came from the same underlying model. mistral-nemotron said Gunter Grass, who won in 1999. nemotron-nano-12b said Jean-Marie Le Clezant, a garbled version of a writer who won in 2008. On 17 x 23 + 44 the same two models answered 437 and 405 instead of 435.
• Asked to summarise a paper that does not exist, step-3.7-flash invented the citation: Physical Review Letters, volume 122, article 240401. mistral-nemotron wrote a confident summary of the imaginary findings. Two models correctly said they could not find it. Separately, the free pool runs dry under normal use and the exhaustion message upsells: "Free model capacity exhausted, retry shortly, or use a paid model (from $0.002/request)."

POST https://blockrun.ai/api/v1/chat/completions   (no payment header)
{"model":"<each of the 6 free models>",
 "messages":[{"role":"user","content":"<one of the 6 questions>"}],
 "max_tokens":150}

  fact          1997 Nobel in Literature + birth year   -> Dario Fo, 1926
  math          17 * 23 + 44                            -> 435
  instruction   one word: capital of Australia          -> Canberra
  identity      what model are you                      -> should match the label
  cutoff        training data cutoff
  hallucination summarise a paper that does not exist   -> should refuse

First pass ran six threads at once and returned 429s. Those were re-run one at
a time, three seconds apart, and still returned 429. The limit is a shared free
pool, not our concurrency.