Reference
Changelog
New features, improvements and fixes in Kyara Intelligence.
Follow along as Kyara Intelligence evolves. Kyara was in active development for a while before this changelog started, so earlier work isn't captured here.
0.7
October 28, 2026 BillingPlan changes and renewals now handle your weekly allowance more precisely.
- Changing plans moves your allowance by the difference. Moving up mid-week adds the difference between the two weekly caps; moving down removes it. Switching plans no longer refills the full allowance. See Plans and credits.
- Renewals refill once the payment goes through. The full allowance now arrives with the paid renewal invoice.
0.6
August 23, 2026 BillingTop-up credits now work without an active subscription.
- Your credits keep working if your subscription lapses. If your plan ends or a payment fails,
requests keep running on the top-up credits you already bought.
403 no_active_planis only returned when you have no active plan and no top-up credits. Buying credit packs still takes an active subscription. See Plans and credits.
0.5
August 20, 2026 APIBilling fairnessNew request parameters, cleaner responses, and a few billing fixes.
- More sampling controls.
repetition_penaltyandmin_pare now accepted and passed to the model. - Sessions for prompt caching. Pass a
session_idto keep a conversation on the same prompt cache across turns. See Sessions and prompt caching. - New models.
deepseek/deepseek-v4-flash-0731anddeepseek/deepseek-v4-pro-0813are in the catalog. - Cleaner responses and consistent errors. Responses carry only public fields,
modelalways matches the catalog ID, and errors use a fixed set of codes documented in Errors. - Failed generations are free. A request that fails, including a stream that fails partway through, is not billed.
- Slow generations are no longer cut off. Long-running requests that went quiet for more than a few seconds could be dropped mid-generation.
0.4
August 2, 2026 BillingOne weekly allowance replaces the separate base and weekly buckets.
- One weekly allowance per plan. Spark includes $4.00 per week and Ember $8.00 per week. The base bucket and its eight-hour refresh are gone.
- Everyone resets Friday at 06:00 UTC. One global reset replaces per-user rolling windows. Unused allowance does not roll over.
- Going over carries into the next reset. Usage beyond your balance is deducted from Friday's allowance. Renewing or changing plans restores the full allowance. Top-up credits are untouched and never go negative.
GET /v1/balancehas a new shape.balances.weekly,caps.weeklyand a staticrefreshobject replace the base and overage balances, caps and window objects.balances.totalno longer reports a percentage, and reading your balance no longer triggers refills.- One weekly meter and a complete ledger in the dashboard. The meter shows when you are over the allowance.
0.3
July 31, 2026 Billing fairnessReasoningA billing fairness release: you no longer pay for empty generations, and a refill race that could overfill balances is fixed.
- Empty generations are free. A generation with no visible output costs nothing.
- Every request has a sane output cap. Requests without an output limit default to 8,192 output tokens, and explicit limits are clamped to the model's maximum, so the cost estimate and the generation agree on the same bound.
- Finer reasoning control. The
reasoningobject acceptsenabledandexcludealongsideeffortandmax_tokens, so you can turn reasoning on or off and keep reasoning text out of the response. - Credit refills no longer overfill. Concurrent requests arriving as a credit window elapsed could each apply a full refill, pushing a balance above its plan cap. Balances over-granted by this bug have been corrected.
0.2
July 24, 2026 StabilityBilling accuracyA large rewrite of the API: the whole stack is more stable, usage tracking is more precise, and caching hits far more often.
- Reasoning controls reach the model, with more levels.
reasoningandreasoning_effortare now passed on in the form the model expects instead of being dropped, and the effort levels gainedmaxandxhigh. - Legacy
max_tokensis honored, including in the cost estimate, so older OpenAI-style clients get the output limits they ask for. - No more concurrency cap. The limit of three concurrent generations per user and its
429error are gone. - Far fewer false insufficient balance rejections for users with a small balance left.
- Cleaner, more standard responses. Internal fields are stripped from responses and every streaming chunk, and
modelis the public model ID. - More accurate billing, based on the usage each response reports.
- Better error messages. An unknown model is a
400withcode: "model_not_found", and accounts without a plan get a403withcode: "no_active_plan". - Higher cache hit rate and slightly faster requests.
0.1
July 24, 2026 First tracked releaseThe starting point for the public changelog.
- OpenAI-compatible API live at
https://api.kyara-intelligence.com/v1. - A curated catalog of capable models behind one familiar API.
- A dashboard for creating API keys and managing your account.
- Plans and credits, with a bucket-based balance system.
- Documentation covering the quickstart, authentication, integrations and the API reference.