Pricing

How pricing works

Kyma charges per token. Every request has:

input tokens for your prompt and context
output tokens for the model’s response

No monthly fee, no seat fee, no contract. Prices below are per 1 million tokens. The $0.50 signup credit spends on the free tier, not on the whole catalogue. Everything else that generates media needs a top-up first. For the live canonical table, use GET /v1/models or GET /v1/credits/pricing.

Very long prompts

A few models cost more per token once a single prompt passes a threshold, because the infrastructure serving them charges more there. The rates below are the standard ones, which is what almost every request pays.

Model	Input	Output	Above a longer prompt
`qwen3.7-flash`	$0.0504	$0.2187	$0.14 /$ 0.54 above 32K · $0.27 /$ 1.08 above 256K
`qwen-3.6-plus`	$0.454	$2.724	$1.76 /$ 5.27 above 256K
`qwen-3.7-plus`	$0.4431	$1.773	$1.30 /$ 5.18 above 256K
`grok-4.3`	$1.92	$3.838	$3.38 /$ 6.75 above 200K
`grok-build`	$1.389	$2.777	$2.70 /$ 5.40 above 200K
`gpt-5.6-terra`	$2.70	$16.20	$5.40 /$ 24.30 above 272K

The threshold is on the prompt, not the conversation, so you know which rate applies before you send. Every other model on Kyma is a single flat rate at any length, and usage.cost on the response is always the exact amount charged.

Pricing per model

Cheapest useful models

Use these for bulk automation, extraction, and simple workloads.

Model	Input / 1M	Output / 1M	Best for
`deepseek-v4-flash-vision-exp`	$0.2514	$0.7542	Cheap multimodal reasoning on V4 Flash (experimental)
`qwen3.8-27b`	$0.3576	$2.554	Open-weight multimodal reasoning at 1M context
`gemini-3.8-flash`	$2.026	$10.126	High-volume agentic and multimodal work
`claude-fable-5.1`	$13.50	$67.50	Hardest reasoning and long-horizon agents
`claude-fable-5`	$13.50	$67.50	Highest-end reasoning flagship
`grok-4.7`	$2.16	$6.48	Long-running coding and agentic work
`grok-4.6`	$2.758	$8.274	Frontier reasoning + agentic work
`llama-4-maverick`	$0.27	$1.08	Open weights, two suppliers
`muse-spark-1.1`	$1.542	$5.243	Long-context reasoning
`grok-4.20-multi-agent`	$1.767	$3.534	Reasoning at long context
`grok-4.20`	$1.763	$3.528	Reasoning at long context
`grok-4.5`	$2.73	$8.189	Reasoning at long context
`gpt-5.6-sol-pro`	$5.40	$27.00	Sol with the Pro serving profile, 1.05M context
`gpt-5.6-sol`	$5.40	$27.00	OpenAI’s top 5.6 tier, 1.05M context, files + images
`gpt-5.6-terra-pro`	$2.70	$16.20	Terra’s Pro tier, 1.05M context, files + images
`gpt-5.6-luna-pro`	$0.1548	$0.9288	Luna with the Pro serving profile, 1.05M context
`gpt-5.6-luna`	$0.3127	$1.877	OpenAI’s cheapest 5.6 tier, 1.05M context, files + images
`claude-sonnet-5`	$2.70	$13.50	Anthropic’s balanced tier, 1M context, extended reasoning
`claude-opus-5`	$6.75	$33.75	Anthropic’s flagship tier, 1M context, extended reasoning
`gpt-5.6-terra`	$2.70	$16.20	OpenAI balanced tier — 1M context, files + images
`qwen3.7-flash`	$0.0504	$0.2187	Cheapest in the catalogue — multimodal, 1M context
`gemini-3.5-flash-lite`	$0.405	$3.375	Cheapest 1M-context tier — high-volume extraction
`gemini-3.6-flash`	$1.013	$5.063	Flash tier — 1M context, multimodal input
`gemini-3.7-flash`	$2.026	$10.126	Newest Flash — 1M context, 64K output, multimodal input
`glm-4.7-flash`	$0.081	$0.54	Cheap long-context throughput
`gemma-4-31b`	$0.0763	$0.218	Cheap multimodal / vision
`deepseek-v4-flash`	$0.1264	$0.2526	Best value V4 — 1M context, native reasoning
`gpt-oss-120b`	$0.0494	$0.2406	Cheap writing + general tasks
`glm-4.5-air`	$0.1945	$1.272	Cheap agentic bulk work
`step-3.7-flash`	$0.27	$1.552	Cheap multimodal flash — text+image+video in, tool calling

Balanced value

Use these when you want strong quality without flagship pricing.

Model	Input / 1M	Output / 1M	Best for
`qwen-3-32b`	$0.108	$0.378	Fast coding loops
`minimax-m2.5`	$0.405	$1.62	Agentic coding
`minimax-m2.7`	$0.405	$1.62	Productivity + debugging
`minimax-m3`	$0.405	$1.62	Newest MiniMax — agentic coding, 1M context, multimodal
`deepseek-v3`	$0.518	$0.7571	Best value frontier-class model
`llama-3.3-70b`	$0.135	$0.432	Balanced open model

Flagship / premium

Use these when quality or capability matters more than cost. Rows showing · cached print the cached input price inline (10% of normal input — applies to repeated system prompts and tool definitions on caching-enabled models; see prompt caching).

Model	Input / 1M	Output / 1M	Best for
`qwen-3-coder`	$0.3586	$1.629	Code-specialized output
`deepseek-r1`	$0.7425	$2.957	Deep reasoning
`kimi-k3`	$3.768	$18.837	Moonshot’s newest — 1M context, vision, long-horizon agents
`kimi-k2.6`	$0.7856	$3.667	Agentic flagship
`kimi-k2.7-code`	$1.009	$4.774	Coding specialist — fewer reasoning tokens, 262K context
`qwen-3.6-plus`	$0.454	$2.724	Best default overall
`qwen-3.7-plus`	$0.4431	$1.773	Newest Qwen flagship — vision + 1M context
`claude-opus-4-7`	$6.75	$33.75	Anthropic flagship — top engineering quality
`claude-sonnet-4-6`	$4.05	$20.25	Anthropic balanced — long-running agentic
`claude-haiku-4-5`	$1.35	$6.75	Anthropic fast tier — 200K context
`glm-5.2`	$1.101	$3.76	Newest flagship — 1M context, #1 open-weight coding
`glm-5.1`	$1.89	$5.94	Long-running coding agents (previous-gen)
`nemotron-3-ultra-550b`	$0.675	$3.375	Largest US open-weight — 1M context, fast reasoning
`hermes-3-405b`	$1.35	$1.35	Steerable assistants and persona work — follows a system prompt further
`hermes-3-70b`	$0.945	$0.945	The same tune, 30% cheaper per million

Long-context & multimodal

Use these when context length or image input is the bottleneck.

Model	Input / 1M	Output / 1M	Best for
`gemini-3-flash`	$0.675	$4.05	Newer long-context preview
`gemma-4-31b`	$0.0763	$0.218	Vision + text workflows

Embeddings

Embedding models turn text into vectors for retrieval. Billed on input tokens only — the response is a vector, so there is no completion side to charge for.

Model	Input / 1M	Dimensions	Best for
`embeddinggemma-300m`	$0.0027	768	Bulk corpus indexing — the cheapest way to embed at volume
`qwen3-embedding-8b`	$0.0135	4096	Multilingual retrieval, 32K input window, high recall
`qwen3-embedding-4b`	$0.027	2560	Existing 2560-dimension indexes — the 8B is better and cheaper
`qwen3-embedding-0.6b`	$0.0135	1024	High-volume indexing, 32K window, smallest vector to store
`bge-m3`	$0.0135	1024	Multilingual RAG across 100+ languages, 8K window
`bge-large-en`	$0.0135	1024	English-only retrieval, and existing BGE indexes
`bge-base-en`	$0.00675	768	English retrieval at half the vector, and half the index
`gte-base`	$0.00675	768	Out-of-domain English retrieval — Alibaba’s GTE line
`multilingual-e5-large`	$0.0135	1024	Cross-language retrieval on short chunks, 94 languages
`all-minilm-l12`	$0.00675	384	Semantic search at the smallest index size
`all-minilm-l6`	$0.00675	384	Matching an existing MiniLM index, local or hosted

Reranking

Scores how well each document answers a query, so a retriever’s top 50 becomes the 5 worth sending to a model. Billed on input tokens — the query plus every document you send.

Model	$/1M input	Context	Best for
`qwen3-reranker-8b`	$0.135	40K	Cutting retrieval noise before it reaches the context window

For comparison, OpenAI’s text-embedding-3-small lists at $0.02 / 1M. See the embeddings API reference.

Web search

Perplexity’s Sonar models run a live web search on every request and return cited answers. They bill two components on the same request: normal token pricing plus a flat web-search fee for the live search.

Model	Input / 1M	Output / 1M	Web-search fee	Best for
`sonar`	$1.35	$1.35	$0.00675 / request	Quick cited answers, current events
`sonar-pro`	$4.05	$20.25	$0.00675 / request	Deep research, longer cited reports

The web-search fee is flat per request — it does not scale with tokens — so a short sonar question is dominated by the search fee, not the tokens. Every response’s usage.cost already includes it. (The GET /v1/pricing catalog lists the per-token prices above; the flat search fee is applied per request at billing time, not as a catalog field.) Standard models have no per-request fee; use Sonar only when you actually need fresh web data.

Image generation

Image models bill per image, not per token. The cost shown is exactly what you pay per image. See the Image Generation guide and the API reference.

Model	Cost / image	Best for
`gpt-image-2`	$0.009 low ·$ 0.072 medium · $0.285 high	OpenAI flagship — text-in-image, multilingual typography
`gpt-image-2.5-flare`	$0.009 low ·$ 0.019 medium · $0.285 high, +$ 0.012 per reference image	GPT Image 2.5 — consistent characters from references, exact text. Sizes up to 1536 px a side
`nano-banana-3-flash`	$0.061	Google Gemini 3.1 image-gen — native edit mode (image-in + prompt → image-out)
`flux-2-pro`	$0.041 (1MP) ·$ 0.101 (4MP)	Photoreal, multi-reference blend (up to 10 sources)
`recraft-v4-pro`	$0.338	4MP print-ready design
`recraft-v4-vector`	$0.108	Native SVG output — editable paths and layers
`ideogram-v3`	$0.081	Typography, logos, packaging
`recraft-v4`	$0.054	Design-quality default — #1 HF Arena
`flux-kontext-pro`	$0.054	Image edit, inpaint (requires `image_url`)
`minimax-image-01`	$0.005	Sub-cent budget tier, bulk

A request with n: 4 is billed as 4 images. gpt-image-2 and gpt-image-2.5-flare requests can opt into quality: low/medium/high per call; the hold books the right tier amount up front so high-quality requests reserve $0.285 (no refund-and-rebill drift). Holds refund in full if generation fails. recraft-v3, recraft-v4-vector-pro and flux-1.1-ultra are paused as of 12 September 2026. Use recraft-v4, recraft-v4-vector and flux-2-pro instead.

Use Which model should I use? to pick by task rather than by price.

What does a typical request cost?

Use case	Model	Tokens (in+out)	Cost
Quick extraction	`glm-4.7-flash`	~500	~$0.00016
Screenshot understanding	`gemma-4-31b`	~1,000	~$0.00014
Code review	`qwen-3-32b`	~3,000	~$0.00073
Long document summary	`gemini-3-flash`	~10,000	~$0.0236
Deep reasoning task	`deepseek-r1`	~5,000	~$0.00958
Repo-scale agent step	`glm-5.1`	~8,000	~$0.03132

With the $0.50 signup credit, you can make hundreds or thousands of requests depending on model choice.

Free tier

The signup credit is a free tier, not cash for the whole catalogue. Until you buy credits, it spends on text models at or under

3/M output and

0.50/M input, plus media at or under $0.01 per billing unit. A billing unit is whatever that model is quoted in — a minute of transcription, 1,000 characters of speech, one image, one video, one second, one song, one call. At a cent a unit, $0.50 always buys at least 50 of them. The media models in the free tier today: gemini-3.5-transcribe, gpt-4o-mini-transcribe-2025-12-15, gpt-transcribe, minimax-image-01, whisper-v3-turbo Everything else that generates media — image, video, music, sound effects, the premium voices, and the realtime translation and live sessions — needs a top-up. Ask for one on a fresh account and you get:

HTTP 402
{
  "topup": "https://api.kymaapi.com/v1/topup",
  "balance": 0.5,
  "required": 0,
  "topup_url": "https://kymaapi.com/billing",
  "error": {
    "message": "gpt-image-2 is not in the free tier; add credits to use it. Free-tier media: ...",
    "type": "billing_error",
    "code": "model_requires_paid"
  },
  "details": {
    "balance": 0.5,
    "required": 0,
    "topup_url": "https://kymaapi.com/billing",
    "hint": "Top up at https://kymaapi.com/billing to continue, or POST https://api.kymaapi.com/v1/topup to pay without one",
    "endpoint_id": "gpt-image-2"
  }
}

message lists the free-tier media ids in full; they are trimmed here because the list moves with prices. required is 0 because no top-up makes a premium model free — any purchase unlocks the whole catalogue. When the refusal IS about an amount (code: "insufficient_credits"), required, details.required and need_usd all carry the figure you are short of. Any purchase — the $5 package included — unlocks the full catalogue permanently. The list above is computed from published prices, so it moves when a price does. Updated 2026-09-07.

How billing works

Before each request, Kyma checks that your balance covers it; if it does not, you get a 402 and nothing is charged
After the response completes, Kyma charges the real token cost; usage.cost in the response is the price of that call
Image and video jobs hold an estimate before they run; when the job ends the real cost is charged and the rest of the hold is returned

You pay for actual usage, not an estimate.

Credits

Action	Amount
Signup bonus	+$0.50
Referral reward	+$0.50
Purchase packages	$5 /$ 20 / $100 /$ 500

The signup bonus spends on the free tier; purchased credits spend on anything. Credits never expire. Auto top-up is supported in the dashboard.

# Check your balance
curl https://kymaapi.com/v1/credits/balance \
  -H "Authorization: Bearer kyma-your-api-key"

# Full pricing catalog (text + image + video + audio)
curl https://kymaapi.com/v1/pricing

Video generation

Billed per second of generated footage. A failed job is refunded in full.

Model	Cost / sec	Durations	Audio
`kling-avatar-v2-pro`	$0.15525	as long as your audio, 5 to 60s	your own track
`veo-3.1-fast`	$0.135	4, 6 or 8s	native

Other video ids are paused and return 404 model_not_found. See Video Generation.

Audio

Audio splits by capability — transcription, understanding, realtime, TTS, music, voice, SFX. Speech-to-text (per minute, 1-min minimum billable)

Model	Cost	Notes
`whisper-v3-turbo`	$0.0009/min	Whisper v3 Turbo — 228× realtime. Alias `transcribe`
`gpt-transcribe`	$0.006075/min	OpenAI premium STT — code-switching, conversational. Pin by SKU
`gpt-4o-mini-transcribe-2025-12-15`	$0.00405/min	Alias `transcribe-quality`. Retiring 2027-02-26 — move to `gpt-transcribe`

TTS (per 1k characters)

Model	Cost	Notes
`minimax-speech-turbo`	$0.040	Lowest-latency voice
`minimax-speech-hd`	$0.070	Production multilingual
`eleven-flash-v2-5`	$0.2025	ElevenLabs fast
`eleven-turbo-v2-5`	$0.2025	ElevenLabs turbo
`eleven-multilingual-v2`	$0.405	ElevenLabs quality

Paused as of 12 September 2026: realtime translation and live sessions, music, voice clone, voice design, sound effects, and the eleven-v3 voice. No model in those groups is being served, so they have no price here; a request returns 404 model_not_found. Use eleven-multilingual-v2 in place of eleven-v3. Live source: GET /v1/pricing (audio rows).

Prompt caching

Cached input tokens are charged at that model’s own cached-input rate, published beside its input rate above. There is no catalogue-wide discount: measured across the models that publish one, the cached rate runs from a tenth of the input rate up to the input rate itself. Models with no published cached rate send pricing.cached: null. That means:

repeated system prompts get much cheaper
tool definitions become cheaper over repeated runs
long agent sessions benefit the most

Learn more about prompt caching →

Rate limits

Kyma uses a tier-based system similar to xAI and Anthropic-style paid access: your tier increases based on total credits purchased, not lifetime usage.

Tier	Min Purchase	RPM	Per-model RPM	TPM
0 (Free)	$0	30	25	200K
1 (Starter)	$10	60	40	500K
2 (Builder)	$50	120	80	2M
3 (Pro)	$250	200	150	5M
4 (Enterprise)	$1,000	300	200	10M

RPM = requests per minute Per-model RPM = how many requests per minute a single model can take TPM = tokens per minute across your account There are no daily or monthly caps. Your balance and tier are the real limits.

# Check your current limits
curl https://kymaapi.com/v1/auth/limits \
  -H "Authorization: Bearer kyma-your-api-key"

What happens when you hit a limit?

Out of credits

{
  "error": {
    "message": "Insufficient credits. Add credits at https://kymaapi.com/dashboard.",
    "type": "insufficient_credits"
  }
}

Too many requests

{
  "error": {
    "message": "Rate limit exceeded (30 RPM for Tier 0). Try again in a few seconds.",
    "type": "rate_limit"
  }
}

The Retry-After header tells you when to retry.

How to spend less

Use cheaper models for extraction, routing, and repetitive automation
Use qwen-3-32b instead of a flagship if you mainly need fast coding help
Use gemini-3-flash only when long context is the real need
Use deepseek-r1 only when the task truly needs deeper reasoning
Purchase credits if you need higher rate limits, not just more balance

Getting Started

More

How pricing works

Very long prompts

Pricing per model

Cheapest useful models

Balanced value

Flagship / premium

Long-context & multimodal

Embeddings

Reranking

Web search

Image generation

What does a typical request cost?

Free tier

How billing works

Credits

Video generation

Audio

Prompt caching

Rate limits

What happens when you hit a limit?

How to spend less

Next steps

​How pricing works

​Very long prompts

​Pricing per model

​Cheapest useful models

​Balanced value

​Flagship / premium

​Long-context & multimodal

​Embeddings

​Reranking

​Web search

​Image generation

​What does a typical request cost?

​Free tier

​How billing works

​Credits

​Video generation

​Audio

​Prompt caching

​Rate limits

​What happens when you hit a limit?

​How to spend less

​Next steps

How pricing works

Very long prompts

Pricing per model

Cheapest useful models

Balanced value

Flagship / premium

Long-context & multimodal

Embeddings

Reranking

Web search

Image generation

What does a typical request cost?

Free tier

How billing works

Credits

Video generation

Audio

Prompt caching

Rate limits

What happens when you hit a limit?

How to spend less

Next steps