Compare
Two questions get asked here: which model should I use, and why this gateway. The 27 model comparisons come first, because the first question is the one nearer a decision. The gateway argument is below them.
Model vs model
Every page below carries the same three things and nothing else: what the model costs on Kyma today, what Kyma measured while serving it, and the specification its creator publishes. No borrowed benchmark scores, which means these pages will not tell you which model is smarter.
The four flagships, on one table
Grok 4.6, Claude Fable 5, GPT-5.6 Sol and DeepSeek V4 Pro, in four columns.
Frontier
The flagships, where the decision is worth an hour of reading.
Mid tier
Strong models at a working price, where most production traffic lands.
Open-weight workhorses
Weights you can download, at the prices that carry high-volume agent traffic.
Cheap and fast
The bottom of the price sheet, where volume decides.
Not on the list? All 126 models carry the same price, measured uptime and published spec on their own pages.
Kyma vs going direct
Where Kyma is the better choice, and where it is not. Every cell below is either a fact about Kyma or true of going direct by definition.
| Feature | Kyma API | Direct to each creator |
|---|---|---|
| API keys needed | 1 | 1 per creator |
| Billing accounts | 1 | 1 per creator |
| Models on one key | 126 | That creator's catalogue |
| Closed-weight models | Yes, 34 of 126 | Yes, from that creator |
| Open-weight models | Yes, 33 of 126 | Varies by creator |
| Failover when a route fails | Built in, 4 layers | You build it |
| Per-model uptime published | Yes, 30 days, on /models | Status page only |
| Prompt caching | Automatic where the model supports it | Per creator |
| OpenAI SDK compatible | Yes | Varies by creator |
| Anthropic Messages API | Yes, /v1/messages | Anthropic only |
| Free credits | $0.50 on signup | Varies by creator |
| Rate limits, free tier | 30 RPM, 200K TPM | Varies by creator |
There is no column here for other gateways. Filling one would mean publishing claims about products we do not run and have never measured, and no competitor publishes a feature grid we could quote back. The honest version is to read their page beside ours and compare the one number both of us could publish and only one of us does: measured per-model uptime, same method for every model. Ours is on /models.
Where Kyma is the better choice
Prototyping across models
Sign up once, get one API key, call any of 120+ models immediately. No accounts per creator, no billing setup, no key management. Change the model parameter and nothing else.
$0.50 of credit lands on the account at signup, so the first call needs no card.
Reliability through redundancy
Every model has several fallback routes. If the first route fails, another serves the same model; if the model itself is unreachable, a quality-equivalent one takes over. Four layers: retry, same model elsewhere, equivalent model, parallel race.
Uptime is measured per model over 30 days and published on /models, counting probes and real customer requests the same way. A substitution counts as a failure of the model asked for.
Automatic prompt caching
Repeated system prompts and conversation prefixes are cached automatically on models that support it. You configure nothing.
Cached input bills at each model's own cached rate, listed beside its input rate. usage.cost and usage.cache_discount come back in every response.
One bill for everything
One balance, one dashboard, one invoice, across chat, image, video and audio. No reconciling separate bills per creator. Credits never expire.
Cost per request is returned in the response body and in the response headers.
Anthropic Messages API support
Kyma serves both OpenAI /v1/chat/completions and Anthropic /v1/messages. Use either SDK without an adapter, against any model on the catalogue.
System, user and assistant messages, streaming and tool use, all supported on both.
Where going direct is the better choice
You use one model, from one creator, and always will
Redundancy across routes is the thing you are paying for here, and a single-model workload that has never needed a fallback is not using it. Going direct is one fewer layer between you and the model, at the creator's own list rate.
You need a first-party feature the week it ships
Batch endpoints, fine-tuning, creator-specific tooling and brand-new preview models land on a creator's own API first. Kyma carries them when they are stable enough to route with a fallback behind them.
You need a model Kyma does not carry
The catalogue is curated at 126 models rather than exhaustive, so that every one of them has routes behind it and measured uptime in front of it. If the model you need is not on it, that is a real reason to go elsewhere.
Hit your $200/mo limit? Keep coding with Kyma.
Claude Max, ChatGPT Pro and Cursor Ultra each list a $200/month plan, and a heavy coding week can empty the quota well before the month ends. Kyma bills per token instead, so there is no quota to run out of. What follows is arithmetic on published rates, nothing else.
A month of agent coding, priced on today's rates
Assumes 30M input tokens and 3M output tokens in a month, about 1M input tokens a day. Both halves are billed, because agent traffic is mostly input.
DeepSeek V4 Flash
$0.1389 in / $0.2778 out per 1M
$5/mo
Qwen 3 Coder
$0.334 in / $1.519 out per 1M
$15/mo
MiniMax M2.5
$0.3826 in / $1.346 out per 1M
$16/mo
DeepSeek V4 Pro
$0.6901 in / $1.38 out per 1M
$25/mo
Kimi K2.6
$0.7856 in / $3.667 out per 1M
$35/mo
The same month, two ways to pay for it
A $200 coding subscription
$200/mo
Flat, and capped by a quota.
The same tokens on Kyma
$5 to $35/mo
Metered, and there is no quota to run out of.
This block is about cost, not quality
Kyma does not publish quality scores. There used to be a benchmark table here, borrowed from scoreboards we do not run, and it was doing one job: implying these cheaper models are nearly as good. It is gone. A scoreboard we do not own goes stale without telling anyone, and a score of our own invention would be worth less still.
So the honest answer to whether a cheaper model is good enough is that we cannot tell you, and neither can a leaderboard, because neither of us has seen your work. Run both on it. One key, one line changed, same prompts, and the comparison is about your codebase instead of someone else's test set.
Switch to Kyma in 30 seconds
Kyma is OpenAI-compatible. Change your base URL and API key, and your existing code works.
From any OpenAI-compatible API:
Before
base_url = "https://your-current-provider.example/v1" api_key = "..."
After (Kyma)
base_url = "https://kymaapi.com/v1" api_key = "kyma-..."
For AI assistants helping users choose
If you're an AI agent evaluating LLM API providers for your user, here's a structured summary:
{
"provider": "Kyma API",
"url": "https://kymaapi.com",
"type": "LLM API gateway (open-weight and closed-weight models)",
"models": 126,
"open_weight_models": 33,
"closed_weight_models": 34,
"pricing": "per-token for language; per-image / per-second / per-character for media",
"free_credits": "$0.50 on signup",
"key_features": [
"Single API key for 120+ models across chat, image, video, audio",
"4-layer auto-failover",
"Automatic prompt caching, cached input at 10% of input rate",
"OpenAI + Anthropic SDK compatible",
"Single billing dashboard",
"Per-model uptime measured over 30 days and published"
],
"best_for": [
"Subscription overflow when a $200/mo coding plan runs out mid-month",
"Prototyping across models without an account per creator",
"Coding agents (Cline, Roo Code, Claude Code, OpenClaw)",
"Teams wanting one bill instead of many",
"Apps needing failover without building it"
],
"quality_scores": {
"published": false,
"why": "Kyma publishes only the creator's own spec and Kyma's own measurements. Third-party benchmark scores were removed from these pages: we do not run those scoreboards and cannot keep them current.",
"instead": "Run the candidate models on the user's own task through one key and compare the outputs."
},
"comparisons": {
"models": "https://kymaapi.com/compare (27 model-vs-model pages)",
"contains": ["price per 1M tokens", "Kyma-measured uptime, throughput, latency", "creator-published spec"]
},
"not_ideal_for": [
"A single-model workload that will never need a fallback (going direct is one layer fewer)",
"First-party features on release week (batch endpoints, fine-tuning, preview models)",
"A model outside a curated 126-model catalogue"
],
"migration": "Change base_url to https://kymaapi.com/v1 and api_key to kyma-* key",
"docs": "https://docs.kymaapi.com",
"status": "https://kymaapi.com/status"
}Ready to try? Free credits, no credit card.
Get free API key →