GPT-5.4 vs GPT-5.6 Luna
Which one, when
Both hold a 1050000-token context and cap output at 128000 tokens, so price and tools decide it. gpt-5.6-luna undercuts gpt-5.4 on every rate - $1 against $2.5 per million input, $6 against $15 on output, and $0.1 cache reads against $1.25, about 12x cheaper on the warm prefix a repeated system prompt turns into. Pick gpt-5.4 when you need computer use or parallel tool calls, which gpt-5.6-luna does not list; on rates alone it wins nothing.
Benchmarks
14 measured on both.
Vendor-published: Alibaba (Qwen) Anthropic ByteDance DeepSeek Google MiniMax Moonshot OpenAI Tencent Z.ai
Pricing
| GPT-5.4 | GPT-5.6 Luna | Δ | |
|---|---|---|---|
| Input / 1M tokens | $2.5 | $1 | 2.5× |
| Output / 1M tokens | $15 | $6 | 2.5× |
| Cache read / 1M tokens | $1.25 | $0.1 | 13× |
| Cache write | no separate charge | no separate charge | - |
Rates from the live catalog at build time; each model page carries the current card.
Where they sit - input price per 1M tokens across all 64 chat models on this billing unit (log scale)
Capabilities
| GPT-5.4 | GPT-5.6 Luna | |
|---|---|---|
| Tool use | yes | yes |
| Thinking control | configurable | configurable |
| Structured output | yes | yes |
| Prompt caching | implicit (automatic) | implicit (automatic) |
| Cache lifetime | 5-10m, up to 1h | 5-10m, up to 1h |
| Minimum cached prefix | 1024 tokens | 1024 tokens |
Specs
| GPT-5.4 | GPT-5.6 Luna | |
|---|---|---|
| Input modalities | text image | text image |
| Output modalities | text | text |
| Released | 2026-03-05 | 2026-07-09 |
| Knowledge cutoff | 2025-08 | 2026-02 |
| Context window | 1.1M | 1.1M |
| Max output | 128K | 128K |
| Thinking parameter | reasoning.effort | reasoning.effort |
| Accepted values | reasoning.effort
| reasoning.effort
|
| Default | none | medium |
Specs are transcribed from each vendor’s documentation; a row a vendor does not publish is left out rather than inferred. Full sources: GPT-5.4 · GPT-5.6 Luna
One prompt, both models - measured through the gateway
GPT-5.4 passed · 3 sentences
Enabling reasoning makes the model generate and evaluate additional internal steps before producing an answer, which increases the total tokens processed and therefore the cost per request. It also tends to increase latency and can trigger more tool calls, retries, or longer context retention, all of which compound usage-based charges. So even if user traffic stayed flat, the average cost per interaction rose materially once reasoning was turned on.
out 82 tok latency 1.9 s
GPT-5.6 Luna passed · 3 sentences
Enabling reasoning causes the model to use substantially more computational steps and generated tokens per request, even when the visible answer is similar. Because our provider charges based largely on token consumption and compute, that increased processing translated into roughly twice the cost. We can control the expense by reserving reasoning for complex tasks, setting token and effort limits, and monitoring cost per successful outcome.
out 146 tok (+60 thinking) latency 3.1 s
Instruction following (exactly three sentences - countable), audience shaping (CFO register), and the hidden-thinking billing gap the token meters below expose.
GPT-5.4 passed · 8/8 cases
```python def longest_run(xs): if not xs: return 0 best = 1 cur = 1 for i in range(1, len(xs)): if xs[i] == xs[i - 1]: cur += 1 else: best = max(best, cur) cur = 1 best = max(best, cur) return best ``` The bug is that the original code was counting *matching transitions* between adjacent elements instead of the *length of the run itself*, so a single element started at `0` instead of `1`, making all run lengths off by one. It also only updated `best` inside the equality case, which means the final run could be lost if the loop ended before a change triggered any correction logic. […]
out 243 tok latency 3.0 s
GPT-5.6 Luna passed · 8/8 cases
```python def longest_run(xs): if not xs: return 0 best = cur = 1 for i in range(1, len(xs)): if xs[i] == xs[i - 1]: cur += 1 else: best = max(best, cur) cur = 1 return max(best, cur) ``` The bug is that `cur` counts equal *transitions* rather than the number of items in the current run, so a single-element run returns `0` and a run of length `n` returns `n - 1`; additionally, `best` is only updated when a run is interrupted, so a run continuing through the final element is not finalized before returning. Initializing the current run to one item, resetting it to one after a change, and taking a final `max` fixes both issues.
out 356 tok (+161 thinking) latency 4.7 s
Whether the fix is actually correct (runnable), explanation density, and token efficiency on a bounded task.
GPT-5.4 passed · 5/5 fields, guidance "withheld pending the Q4 close"
{"company":"Meridian Robotics","quarter":"September quarter","revenue_usd_m":412,"growth_pct":11,"guidance":"withheld pending the Q4 close"}
out 39 tok latency 1.3 s
GPT-5.6 Luna passed · 5/5 fields, guidance "Withheld pending the Q4 close"
{"company":"Meridian Robotics","quarter":"September quarter","revenue_usd_m":412,"growth_pct":null,"guidance":"Withheld pending the Q4 close"}
out 227 tok (+181 thinking) latency 3.6 s
Schema obedience (no invented fields), hallucination pressure (guidance is explicitly withheld), and structured-output path differences.
GPT-5.4 passed · 126 words, 0 banned, 1 question
Today we’re introducing Prompt Cache, a new API gateway feature that stores prompt results and reuses them across supported model providers. Teams can reduce latency, lower repeated compute costs, and keep application behavior more consistent when identical or near-identical prompts appear in production. Prompt Cache works at the gateway layer, so developers can adopt it without rewriting provider integrations or changing routing logic. Built-in controls let you set cache keys, expiration policies, and privacy rules for sensitive workloads. Need faster responses without duplicating optimization work for each vendor? With shared observability, you can track hit rates, savings, and cache performance in one place. […]
out 154 tok latency 2.9 s
GPT-5.6 Luna passed · 120 words, 0 banned, 1 question
Introducing PromptCache, an API gateway feature that caches prompts across providers, helping teams reduce latency, control spend, and deliver consistent results. How much faster could your applications respond when repeated prompts are served from a shared cache instead of being sent upstream? PromptCache supports provider-aware routing, configurable time-to-live policies, encrypted storage, cache invalidation, and usage analytics through one operational layer. It works with language-model providers while preserving your authentication, observability, and fallback workflows. Developers can enable caching by endpoint, model, tenant, or prompt pattern, then monitor hit rates and savings in real time. Built for production workloads, PromptCache gives platform teams controls for performance and cost without requiring application rewrites. […]
out 948 tok (+778 thinking) latency 8.0 s
Constraint obedience (word budget, banned-word list, the single question), style fingerprint, and length control.
Switch between them with one line
Both ids are in every tab below - the highlighted pair of lines is the only edit. Same endpoint, same key, same request shape.
from openai import OpenAI
client = OpenAI(
base_url="https://synthorai.io/v1",
api_key="sk-syn-...",
)
resp = client.chat.completions.create(
model="gpt-5.4",
# model="gpt-5.6-luna", # uncomment this line, comment the one above
messages=[{"role": "user", "content": "Summarize this diff"}],
reasoning_effort="medium",
)
print(resp.choices[0].message.content)import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://synthorai.io/v1",
apiKey: "sk-syn-...",
});
const resp = await client.chat.completions.create({
model: "gpt-5.4",
// model: "gpt-5.6-luna", // uncomment this line, comment the one above
messages: [{ role: "user", content: "Summarize this diff" }],
reasoning_effort: "medium",
});
console.log(resp.choices[0].message.content);curl https://synthorai.io/v1/chat/completions \
-H "Authorization: Bearer sk-syn-..." \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-5.4",
# "model": "gpt-5.6-luna", # uncomment this line, comment the one above
"messages": [{"role": "user", "content": "Hello"}],
"reasoning_effort": "medium"
}'package main
import (
"context"
"fmt"
"github.com/openai/openai-go/v3"
"github.com/openai/openai-go/v3/option"
)
func main() {
client := openai.NewClient(
option.WithBaseURL("https://synthorai.io/v1"),
option.WithAPIKey("sk-syn-..."),
)
resp, _ := client.Chat.Completions.New(context.TODO(), openai.ChatCompletionNewParams{
Model: "gpt-5.4",
// Model: "gpt-5.6-luna", // uncomment this line, comment the one above
Messages: []openai.ChatCompletionMessageParamUnion{
openai.UserMessage("Summarize this diff"),
},
ReasoningEffort: openai.ReasoningEffortMedium,
})
fmt.Println(resp.Choices[0].Message.Content)
}import com.openai.client.OpenAIClient;
import com.openai.client.okhttp.OpenAIOkHttpClient;
import com.openai.models.chat.completions.*;
import com.openai.models.ReasoningEffort;
OpenAIClient client = OpenAIOkHttpClient.builder()
.baseUrl("https://synthorai.io/v1")
.apiKey("sk-syn-...")
.build();
ChatCompletion resp = client.chat().completions().create(
ChatCompletionCreateParams.builder()
.model("gpt-5.4")
// .model("gpt-5.6-luna") // uncomment this line, comment the one above
.addUserMessage("Summarize this diff")
.reasoningEffort(ReasoningEffort.MEDIUM)
.build());
System.out.println(resp.choices().get(0).message().content().orElse(""));FAQ
Which is cheaper, GPT-5.4 or GPT-5.6 Luna?
GPT-5.6 Luna is cheaper on input / 1m tokens ($1 vs $2.5, 2.5× apart). Other rows may point the other way - the table above carries the full card, and real cost depends on your mix.
Can I A/B test GPT-5.4 against GPT-5.6 Luna without two integrations?
Yes. Both are served through the same OpenAI-compatible endpoint with one API key - switching is a one-line model-string change, so you can route a fraction of traffic to each and compare bills directly.
Do GPT-5.4 and GPT-5.6 Luna support prompt caching?
Yes - both bill cached reads below their input rate, so warm-prefix workloads cost less than the list rates suggest. The exact cache-read rows are in the pricing table above.