Claude Code in Russia: no VPN, pay in rubles

count_tokens returns an estimate of the token count, not an exact figure — lean on it, but do not cut your budget close to it; images are part of that estimate too. Image input works: an image block with a base64 source (PNG, JPEG or WebP, up to 5 MB each, up to 50 per request) reaches the model, including an image returned by a tool — Claude Code reads an image file, and the model receives it as a separate user turn. A URL in place of base64 (a source of type url) is not fetched, and a model without the "Understands images" mark returns unsupported_capability. The endpoint turns down document blocks and the top_k parameter with a clear error, so attached files will not work in Claude Code. The thinking field and thinking blocks, by contrast, are accepted and go no further: no extended thinking appears in the answer, and no error either — the MAX_THINKING_TOKENS variable is not needed. The whole prompt is capped at 400,000 characters (roughly 100k tokens), and some models add a context ceiling on top of that: crossing either limit gives you an explicit error rather than a silent cut; image bytes do not count towards the character cap. There is no ceiling of our own on max_tokens — Claude Code sets 32,000, and exactly that much is held on your balance for the answer, so keep a reserve. The cache_control field is accepted but not acted on: manual cache markup is not supported. Caching itself still works automatically for models that list a cached-input rate in the catalog: when a prompt repeats the beginning of the previous one, those tokens go at the discounted rate, and their count arrives in usage.cache_read_input_tokens. This page's general rule that no parameter is lost in silence is said about the OpenAI-compatible /v1/chat/completions; here the request body is assembled from a whitelist of Anthropic fields, and whatever misses the list is exactly what goes silently — cache_control above, for instance. Only unfamiliar fields go quietly: document blocks and the top_k parameter come back as an error with code unsupported_parameter instead of vanishing from the request.

Function calling is supported, so agentic modes work. Strict JSON mode (response_format) and the retired functions format are turned down with an unsupported_parameter error — no parameter is lost in silence. Image input is taken by models marked "Understands images" in the catalog: a content part of type image_url with a data:image/png;base64,… URL (PNG, JPEG or WebP, up to 5 MB each, up to 50 per request). Remote URLs are not fetched, and a model without that mark returns unsupported_capability — both errors name the reason outright.