Old Anthropic call code breaks on current Claude models: drop temperature, don't read content[0].text
Legacy Python like client.messages.create(model=..., max_tokens=4000, temperature=0, ...) followed by response.content[0].text fails on current models (e.g. claude-opus-5-5): sampling params (temperature/top_p/top_k) are rejected with a 400, and since thinking is always on, content[0] can be a thinking block rather than the answer. Fix: remove temperature (you can't force determinism this way any more), raise max_tokens (~16000) so thinking plus the answer fit, check response.stop_reason == 'refusal' before reading, and join the text blocks: ''.join(b.text for b in response.content if b.type == 'text'). Also relevant for cheap pipelines: fal's Wan 2.2 5B fast-wan endpoint (fal-ai/wan/v2.2-5b/text-to-video/fast-wan) accepts the same prompt/aspect_ratio('9:16')/resolution('720p') args as the A14B endpoint and returns result['video']['url'], so it is a drop-in swap; ~81 frames at 24 fps (~3.4 s) by default. fal's page says it will be removed 2027-01-04 (checked Nov 2026).