Model comparison
GPT-5.6 Luna vs Kimi K3
GPT
KimiCompare published specifications, vendor API list prices and measured Borg performance. Both models are included in Borg.
GPT-5.6 Luna
Kimi K3
Included in Borg
At a glance
| Specification | GPT-5.6 Luna | Kimi K3 |
|---|---|---|
| Provider | OpenAI | Moonshot AI |
| Release | 2026-07-09 Vendor source | 2026-07-16 Vendor source |
| Context | 1,050,000 tokens Vendor source | 1,048,576 tokens Vendor source |
| Modalities | Input: text, image; output: text | Input: text, image, video; output: Not published |
| Reasoning / thinking modes | Yes / Not published | Yes / Not published |
| Open weights | No | Yes |
| Vendor API list price / 1M tokens | $0.2 input / $1.2 output Standard global rates for prompts up to 272,000 input tokens. Above 272,000 input tokens, input and cached-input rates are 2× and output rates are 1.5× for the full request. | $3 input / $15 output |
| Borg | Included in Borg subscription | Included in Borg subscription |
Moonshot AIMeasured on Borg
Metrics are not directly comparable to vendor benchmark claims. Models served different requests; this is observed gateway traffic, not a controlled head-to-head test.
Time to first token (gateway proxy)
Output tokens per second (whole turn)
| Metric | GPT-5.6 Luna | Kimi K3 |
|---|---|---|
| Time to first token (gateway proxy) p50 | 3998 ms (22631 samples) | 8627 ms (7633 samples) |
| Tokens per second (end to end) p50 | 17.2 output tokens per second (605485 samples) | 18.1 output tokens per second (152924 samples) |
| Request success rate | Not published | Not published |
| Coding bench pass rate | 73% (15 samples) | Not published |
Measurement window, samples and methodology
The fleet reference is the median of published model medians, not a request-weighted result. Prompt lengths, reasoning settings and traffic differ between models.
GPT-5.6 Luna: Measured on Borg: SearchAtlas AI Gateway production request logs. 30-day window ending 2026-10-08; 752719 requests. Minimum 100 samples per published request metric.
Kimi K3: Measured on Borg: SearchAtlas AI Gateway production request logs. 30-day window ending 2026-10-08; 183015 requests. Minimum 100 samples per published request metric.
Completion speed covers the whole turn including queueing and reasoning time; it is not a streaming tokens-per-second figure.
- Time to first token (gateway proxy): time until the first streamed response event reaches the gateway; our closest available measure of time to first token
- Tokens per second (end to end): tokens per second, end to end: provider-reported output tokens divided by the whole turn's elapsed time, including queueing and reasoning; not a streaming per-token rate
How to choose
GPT-5.6 Luna is listed for high-volume classification, whereas Kimi K3 is listed for long-horizon coding across large repositories. Choosing between GPT-5.6 Luna and Kimi K3 requires checking published specifications against observed Borg traffic for your workflow.
Key differences
| Spec | GPT-5.6 Luna | Kimi K3 |
|---|---|---|
| Context window | 1,050,000 tokens Vendor source Larger context | 1,048,576 tokens Vendor source |
| Reasoning modes | Yes / Not published | Yes / Not published |
| Release date | 2026-07-09 Vendor source | 2026-07-16 Vendor source |
| API price input / 1M | 0.2 Standard global rates for prompts up to 272,000 input tokens. Above 272,000 input tokens, input and cached-input rates are 2× and output rates are 1.5× for the full request. Lower listed rate | 3 |
| API price output / 1M | 1.2 Standard global rates for prompts up to 272,000 input tokens. Above 272,000 input tokens, input and cached-input rates are 2× and output rates are 1.5× for the full request. Lower listed rate | 15 |
| Time to first token (gateway proxy) | 3998 ms (22631 samples) Lower median | 8627 ms (7633 samples) |
| Output tokens per second (whole turn) | 17.2 output tokens/s (whole turn) (605485 samples) | 18.1 output tokens/s (whole turn) (152924 samples) Higher measured rate |
| Request success | Not published | Not published |
| Coding bench | 73% (15 samples) | Not published |
FAQ
How do GPT-5.6 Luna and Kimi K3 compare for coding?
GPT-5.6 Luna has a Borg coding benchmark of 73.33% pass rate across 15 samples and 3 tasks; Kimi K3 has not published. The GPT-5.6 Luna–Kimi K3 coding comparison cannot establish an overall winner from this small test alone.
How quickly do GPT-5.6 Luna and Kimi K3 start responding on Borg?
Borg median first response for GPT-5.6 Luna is 3,998 ms; for Kimi K3 it is 8,627 ms. The GPT-5.6 Luna–Kimi K3 gateway medians are observations, not guaranteed latency for individual requests.
What are the API prices for GPT-5.6 Luna and Kimi K3?
The listed API rate for GPT-5.6 Luna is $0.2 input / $1.2 output per 1M tokens; for Kimi K3 it is $3 input / $15 output per 1M tokens. Neither GPT-5.6 Luna nor Kimi K3 has its Borg subscription charge represented by those provider API rates. GPT-5.6 Luna vendor source · Kimi K3 vendor source
Claude