2 of 59 · every text model Kyma prices per token · 59 measured
Gemini 3.7 Flash vs GPT-5.6 Luna
Two sources on this page and no others: what Kyma measured while serving these two models, and the specification Google and OpenAI publish for their own. Both run on one key, so switching is a one-line change.
Cheapest input
GPT-5.6 Luna
$0.14per 1M input tokens, against $0.5063 for Gemini 3.7 Flash.
Fastest response
GPT-5.6 Luna
1.14 smedian, on one identical prompt sent to every model every six hours.
Availability
Too close to call
99.8 to 100.0%the gap between them is smaller than these samples can resolve.
Where these models sit in the catalogue Kyma measures
Two numbers side by side tell you which is bigger and nothing about whether either is any good, so every panel puts these models against every text model Kyma prices per token, all 59 of them. That set is a stated rule and not a list, so there is no line-up to accuse anyone of choosing. Price is what Kyma charges today. Availability, throughput and latency are Kyma's own readings over 30 days, on the same method as /models. No benchmark scores, borrowed or invented.
Input price
lower is better
GPT-5.6 Luna
Gemini 3.7 FlashTicks above the rail are the models on this page; below it is every other model Kyma measures on this, 59 in all, on the same scale. The dashed rule is the field median, $1.009; the solid rule is its best, $0.0494, and 6 models sit beyond the axis.
Output price
lower is betterGPT-5.6 LunaGemini 3.7 FlashTicks above the rail are the models on this page; below it is every other model Kyma measures on this, 59 in all, on the same scale. The dashed rule is the field median, $3.528; the solid rule is its best, $0.2164, and 6 models sit beyond the axis.
Failed requests
lower is betterToo close to call · 99.8 to 100.0%
Gemini 3.7 FlashGPT-5.6 LunaTicks above the rail are the models on this page; below it is every other model Kyma measures on this, 59 in all, on the same scale. The dashed rule is the field median, 1.49; the solid rule is its best, 0.00, and 6 models sit beyond the axis.
The gap between them is under 5.2 points, which is one failed request in the thinnest sample here. So these rows are in page order and nothing is marked: the ordering would be the claim, and it is the claim this sample cannot support.
Throughput
higher is betterGemini 3.7 FlashGPT-5.6 LunaTicks above the rail are the models on this page; below it is every other model Kyma measures on this, 59 in all, on the same scale. The dashed rule is the field median, 41; the solid rule is its best, 321, and 4 models sit beyond the axis.
Latency
lower is betterGPT-5.6 LunaGemini 3.7 FlashTicks above the rail are the models on this page; below it is every other model Kyma measures on this, 59 in all, on the same scale. The dashed rule is the field median, 2.25s; the solid rule is its best, 0.81s, and 6 models sit beyond the axis.
Rank in the field
Out of every text model Kyma prices per token, 59 of which Kyma has measured. Where these models lose is stated in the same words as where they win.
Gemini 3.7 Flash21st of 59 on cheapest input · 1st of 59 on most available · 5th of 59 on fastest tokens
GPT-5.6 Luna9th of 59 on cheapest input · 9th of 59 on most available · 30th of 59 on fastest tokens
The facts a chart cannot draw
A conditional rate, a date, a claim with three answers and two different refusals. Still grouped by who said them.
| Fact | Gemini 3.7 Flash | GPT-5.6 Luna OpenAI |
|---|---|---|
| Price on Kyma What a request costs today. The rates themselves are in the panels above, against the whole price sheet. These two are the parts of a price a bar cannot carry. | ||
Cached input per 1M tokens, on repeated prefixes | $0.05061 90% off input | $0.014 Best90% off input |
On offer a temporary rate, and the day it ends | 50% off until Aug 21, then $1.013 in / $5.063 out | 48% off until Aug 21, then $0.27 in / $1.62 out |
| Measured by Kyma Kyma's own numbers, on Kyma's own traffic. Availability, throughput and latency are drawn against the field above; this is the one measured fact with no scale to sit on. Same method, same figures as /models. | ||
Below release precision narrower than the creator shipped | Not established | Not established |
| Published by the creator The specification each creator publishes for its own model, read straight from the catalogue Kyma serves from. Nothing in this block is a Kyma opinion. | ||
Creator | Google | OpenAI |
Context window tokens in one request | 1.05M | 1.05M |
Max output hard ceiling on one response | 66K | 128KBest |
Accepts | Text, image, video, file, audio | Text, image, file |
Tool calling | Yes | Yes |
Reasoning | Yes | Yes |
Vision input | Yes | Yes |
Weights | Closed | Closed |
“Best” marks the winning cell on exact published facts only. Measured rows carry sampling error, so this page does not rank them by their last digit.
Two sources, and nothing borrowed
On this page
Measured by Kyma
Uptime, throughput, latency and the precision the routes behind each model report, from Kyma's own traffic. Ours to publish, and it refreshes itself. Every one of them is drawn against the whole catalogue, because a figure with nothing to compare it to is not evidence.
Published by the creator
Context window, maximum output, accepted inputs, and tool, reasoning and vision support. Each creator's own specification for its own model, as is the precision each creator released its model at, which is the only thing that makes a serving format count as below it.
Taken off it
Third-party benchmark scores
We do not run those scoreboards. One can go stale without telling anyone, and it takes the page's credibility with it.
A Kyma quality score
We do not have one. Inventing a number to fill the gap would be worth less than the scores that were removed.
So this page will not tell you which model is smarter. It tells you which is cheaper, which is faster, which stays up, where each one sits against every other model we sell, and whether it is served below the precision its creator released it at. For the quality question, run them both on your own prompts. One key, one line changed, and the answer is about your work rather than someone else's test set.
Switch with one line
# One key, one endpoint. Only the model string changes.
client = OpenAI(base_url="https://kymaapi.com/v1", api_key="YOUR_API_KEY")
client.chat.completions.create(model="gemini-3.7-flash", messages=[...])
client.chat.completions.create(model="gpt-5.6-luna", messages=[...])Questions
Which of Gemini 3.7 Flash and GPT-5.6 Luna is cheapest?
On input tokens today, GPT-5.6 Luna at $0.14 per 1M. Gemini 3.7 Flash is $0.5063. The panels above put both of those against every model Kyma prices, so you can see whether either is actually cheap.
Can I switch between Gemini 3.7 Flash and GPT-5.6 Luna without changing my code?
Yes. All of them are served through the same OpenAI-compatible Kyma endpoint (https://kymaapi.com/v1) on one API key and one balance. Switching is a one-line change to the model field: `gemini-3.7-flash`, `gpt-5.6-luna`.
Which has the largest context window?
Neither is larger in any way you would notice: all of them take about 1.05M tokens in a single request, so context is not the deciding factor here.
Where do these numbers come from?
Two places and no others. Uptime, throughput, latency and the precision the routes report are measured by Kyma on its own traffic over the last 30 days, across all 59 models it prices per token, 59 of which are measured. Context window, maximum output, accepted inputs, tool, reasoning and vision support are the specification each creator publishes for its own model, cited on the model's own page, and so is the precision each creator released its model at. The "below release precision" row is the two put together: a serving format only counts as below the release once you know what the release was. This page carries no third-party benchmark scores, so it cannot rank these models on quality, and it does not pretend to.
Is there a discount running on any of these?
Gemini 3.7 Flash is 50% off until Aug 21, at $0.5063 per 1M input instead of $1.013. GPT-5.6 Luna is 48% off until Aug 21, at $0.14 per 1M input instead of $0.27. The rate is honoured through that date, then returns to list.