Skip to main content

​
Pricing & access

Yes to start. You get $0.50 of free credit once you verify your email (Google sign-in gets it at once). No credit card required.Signup credit spends on the free tier — text models at or under 3/Moutputand3/M output and 0.50/M input, plus media at or under $0.01 per billing unit. Most image models and every text-to-speech model need a top-up first; minimax-image-01 is the free-tier image model; the video model needs a top-up. Music and realtime sessions are paused. See Free tier for the current list.For heavier usage, buy credits starting at $5.Updated 2026-09-22.
You’ll get a 429 response with a Retry-After header. Wait that long and retry. Limits are tier-based: Tier 0 = 30 RPM, up to Tier 4 = 4,000 RPM. The live numbers for every tier are on GET /v1/limits/tiers.
Yes. Rate limits increase automatically as you purchase credits. Tier 1 unlocks at $5 purchased (600 RPM). See GET /v1/limits/tiers for all tiers.

​
Models & capabilities

Start with qwen-3.6-plus if you want the safest default. Use kimi-k2.6 for tool-heavy agents, deepseek-r1 for hard reasoning, and gemini-3-flash for long-context tasks. See Which model should I use? for the full decision guide.
Some are, and the catalogue says which. Kyma serves both: open-weight models whose creators publish a downloadable checkpoint (DeepSeek, Kimi, GLM, Llama, Gemma, gpt-oss, Nemotron and others), and closed-weight models reachable only through an API (Claude, Gemini, GPT-5.6, Grok, and most of Alibaba’s Max and Plus models). The weights field on /v1/models and the Weights column on /models answer it per model, with the source we read it from — and where we could not establish it from the creator’s own material we say “not established” rather than guess. Kyma itself is a managed service, not open source.
Yes. gemma-4-31b is the cheapest strong multimodal option, and kimi-k2.6 is the better pick if you also need tool-heavy agent behavior.
Yes. Kyma serves image generation through gpt-image-2 (text-in-image), gpt-image-2.5-flare (consistent characters, exact text), flux-2-pro (photoreal), flux-kontext-pro (image edit), ideogram-v3 (typography), recraft-v4 and recraft-v4-vector (design and SVG), and minimax-image-01 (budget). Video generation runs on veo-3.1-fast; music and sound-effect generation are paused as of 12 September 2026.
Up to 2M tokens (grok-4.20), and more than 40 models take 1M or more. Many other active models support 128K-262K. Check GET /v1/models for the live catalog and context windows.

​
Technical limits

No fixed cap. The max_tokens parameter from your request is forwarded to the model, and the max_output_tokens value shown in /v1/models reflects the model creator’s published specification, not a Kyma restriction. You can verify this via GET /v1/capabilities.One exception: when your balance cannot pay for the full max_tokens, Kyma lowers max_tokens to what the balance can buy and says so in the x-kyma-max-tokens-clamped response header (with your balance in x-kyma-credits-have). Top up, or ask for fewer tokens, to avoid it.
Kyma does not fine-tune or modify any model: the weights are the ones the creator published (Alibaba, Google, DeepSeek, Meta, etc.). Some routes serve a model at reduced precision (FP8 or FP4). The reduced_precision field on each model in GET /v1/models says which ones.
Four common causes:
  1. You didn’t set max_tokens, so the model’s own default applied. Set it explicitly for longer outputs (e.g., max_tokens: 8192).
  2. Model’s output limit — each model has a maximum output capacity (check max_output_tokens in /v1/models). For long generation, use models like glm-5.1 (65K) or deepseek-r1 (32K).
  3. finish_reason: length — the model hit its limit naturally. Increase max_tokens or use a model with higher output capacity.
  4. Your balance was low. When the balance cannot pay for the full max_tokens, Kyma lowers it and sets the x-kyma-max-tokens-clamped response header. Otherwise Kyma does not truncate output: if finish_reason is length, the model itself stopped.
The model’s own default. Kyma does not add one: your request goes to the model without a max_tokens. For the balance check, Kyma reserves the cost of 4,096 output tokens when you send none. Always set max_tokens explicitly if you need longer outputs.
Call GET /v1/models — each model includes max_output_tokens, context_window, gateway_output_limit (always null), and max_tokens_passthrough (always true). For a gateway-level overview, call GET /v1/capabilities.

​
Compatibility

Yes. Just change base_url to https://kymaapi.com/v1 and use your kyma- API key. All OpenAI SDK features work.
Yes. Kyma supports the Anthropic Messages API at /v1/messages. See our Anthropic guide.
Yes. Use the OpenAI provider in LangChain with Kyma’s base URL.
Yes. Configure Kyma as an OpenAI-compatible provider with your kyma- API key and https://kymaapi.com/v1 as the base URL.

​
Reliability

If the model has more than one route, Kyma retries your request on another route for the same model, and the outage is invisible to you. Many models have a single route. When every route for the model has failed, you get a 503 with Retry-After, or an answer from a substitute model of similar quality. X-Kyma-Fallback: true tells you a fallback happened; x-kyma-served-model names the model that answered when it is not the one you asked for.
When a fallback occurs, these headers are included in the response:
  • X-Kyma-Fallback: true: a fallback was used
  • x-kyma-served-model: present only when a different model answered, and names it
To know whether you got the model you asked for, compare x-kyma-served-model with your requested model. If the header is absent, the model you asked for answered.
It is measured per model, not promised for the whole gateway. A model with several routes rides out one route failing; a model with a single route is down when that route is. Status publishes measured uptime for every model.

​
API keys & account

Keys start with kyma-. You can create multiple keys on the Keys page. Each key shares the same account balance and rate limits.
Kyma logs API requests to run and bill them. The content of a request, your prompts and the model’s responses, is kept for 90 days and then deleted automatically. The counts and timing (model, tokens, latency) are kept for your usage history and billing. The full retention table is in the privacy policy.
Yes. Higher tiers unlock automatically as you purchase credits. Tier 1 and up get 600 to 4,000 RPM, credits never expire, and you can switch models without changing your integration.