Skip to main content
Follow along as Kyara Intelligence evolves. This is where we track what ships. Kyara has been in active development for a while, and while earlier work is not captured here, this is the first release we are tracking publicly.
​
StabilityBilling accuracy
July 24, 2026
A large rewrite of the API plus a wave of gateway improvements. The whole stack is more stable, usage tracking is more precise, and caching hits far more often.
  • Reasoning controls actually work now, with more levels. reasoning / reasoning_effort is forwarded through Bifrost’s extra_params passthrough instead of as a top-level body field, so the setting genuinely reaches the upstream provider instead of being silently dropped. The effort enum also grew from high / medium / low / minimal / none to add max and xhigh.
  • Legacy max_tokens is honored. Requests fold max_tokens into max_completion_tokens before forwarding, so older OpenAI-style clients get their output limits respected, including in the pre-flight cost estimate.
  • No more concurrency cap. The 3-concurrent-generations-per-user limit and its Too many concurrent generations (429) error are gone. Parallel requests just work.
  • Far fewer false “insufficient balance” rejections. The pre-flight affordability check used to assume 4,096 output tokens when the client did not set a limit. That default is now 800, so users with a small remaining balance can make normal-sized requests without being blocked by a wildly pessimistic estimate.
  • Cleaner, more standard API responses. Bifrost’s internal extra_fields are now stripped from both non-streaming responses and every streaming chunk (previously chunks leaked them), and the model field is rewritten back to the public model id. Responses look like plain OpenAI responses.
  • More accurate billing. Streaming cost is only resolved from chunks that actually carry usage. Streams that end without usage are billed on observed output and tagged completed_no_usage, and the whole flow is typed end to end, so you get charged on real upstream numbers more reliably.
  • Better error messages. An unknown model is now a 400 with code: "model_not_found" (previously a bare 404). Accounts without a plan get a 403 with code: "no_active_plan" and a human-readable message. Validation errors were consolidated for consistent messages and parameters.
  • More precise usage tracking. Usage accounting is tighter and more reliable across both streaming and non-streaming paths.
  • Higher cache hit rate. Caching now hits far more often, cutting latency and cost.
  • Slightly faster requests. The balance snapshot is read once and reused across the windowed-credit refresh and the affordability check, cutting redundant database roundtrips on /v1/chat/completions and /v1/balance.
​
First tracked release
July 24, 2026
The starting point for our public changelog.
  • OpenAI-compatible gateway live at https://api.kyara-intelligence.com/v1.
  • Curated catalog of high-performance models behind one familiar API.
  • Dashboard for creating API keys and managing your account.
  • Plans and credits with a bucket-based balance system.
  • Documentation covering quickstart, authentication, integrations, and the API reference.