Compatibility
What works through the gateway, from the last full run of our compatibility test (version R1): 15 scenarios x 8 combinations of line, model alias and streaming / non-streaming, 87 of 120 cells passed. Providers differ in what they implement; where a result depends on which provider answered, we say so.
Short answers. "Served" means the answer was delivered; it does not mean the answer was clean. Providers of the decentralized network require a minimum of 64 generated tokens on every request, a known behaviour of the network (gonka-ai/gonka issue 1955), so on very short requests (a one-word answer, a short JSON) they append repeated text. In our test of 2026-10-27, 17% to 23% of a mixed set of 60 requests were affected when the decentralized-network providers were called directly (15% and 18% through our two lines), up to 70% of one-word requests and up to 50% of short JSON requests. In the R1 24-hour test, 3.8% to 5.4% of the short answers served by the decentralized-network providers ended at the output limit. The provider outside the decentralized network did not show this (0 of 243 in R1, 0 of 60 on 2026-10-27). The network has published a fix (gonka-ai/gonka pull request 1972); when it will be switched on is not known. Super Deal (economical tariff, depends on the decentralized network, no insurance outside that network) has no provider outside the decentralized network, and on Standard that provider is used only when the earlier ones fail. Since 2026-10-30 the gateway cuts the repeated tail before the answer is delivered, on both lines, and bills only the delivered text. In a live check on that day (one request at a time, a unique mark in every prompt), 2.0% (0.8% to 5.0%) of 200 answers on Super Deal (economical tariff, depends on the decentralized network, no insurance outside that network) and 2.9% (1.5% to 5.5%) of 280 answers on Standard still contained a repeat or text where the model thinks aloud; 96.5% and 95.5% of the requests of a labelled set of short questions were answered right, against 91.5% (410 requests) for the provider outside the decentralized network. These are short single checks, not a promise. Details: benchmarks.
| Feature | Status | Notes |
|---|---|---|
stream (server-sent events) | supported | usage in the last chunk with stream_options.include_usage |
tools, tool calls | supported | one call and parallel calls pass in every combination |
tool_choice: "required" | depends on the provider | sometimes the model answers in text instead of calling a tool |
tool_choice naming a function | not honored | the model does not follow it (it answers in text or calls another tool); the same in every earlier test and straight on every provider |
| tool loops of several turns | depends on the provider | mostly work; sometimes no final answer within four turns |
response_format json_object and json_schema | supported | pass in every combination |
stop sequences | supported | |
system message, temperature, top_p | supported | |
max_tokens, max_completion_tokens | depends on the provider | accepted and used as an upper bound; on decentralized-network providers a very small value (below 64 tokens) can still return up to 64 tokens, and a very short answer can come back lengthened or repeated (see the note above) |
seed | depends on the provider | accepted; identical output for the same seed can differ |
n above 1 | not honored | rejected with an explicit 400 |
logprobs, top_logprobs | depends on the provider | returned only when the answer comes from Provider A or C; Provider B and the insurance provider ignore the parameter and return none |
reasoning_effort | depends on the mode | none turns thinking off; low, high, max turn it on (no gradation); with thinking on use stream: true, because non-streaming requests often end with 503 (see API docs) |
| long context (400,000-token window) | supported | very long requests are slower; on Super Deal (economical tariff) they time out more often (see benchmarks) |
The test is repeatable (bench/c18/compat_matrix.py); individual cells change from run to run with the provider that answered.