Home → Guides → GPT-5.6 Sol in production
Measured over 34 real billed calls on this platform. The cost numbers were roughly what we expected. The latency numbers were not.
gpt-5.6-terra did the same work at
about a third of the cost per call in our measurements.Same platform, same billing system, real invoices. This is the comparison that is hard to find anywhere else, because it requires having actually run all of them:
| Model | Calls | Avg cost / call | Avg first token | Avg output |
|---|---|---|---|---|
gpt-6-astra | 14 | $0.1047 | 2,176 ms | 372 tok |
gpt-5.6-sol | 29 | $0.0796 | 11,873 ms | 975 tok |
gpt-5.6-terra | 13 | $0.0233 | 7,767 ms | 124 tok |
The most expensive model was the fastest to first token by a factor of five. If you assumed the bigger model would be slower, that assumption does not hold here.
Two caveats worth stating plainly. These are different workloads, not a controlled benchmark — average output length differs by a factor of eight across the three rows, and that alone moves both cost and duration. And the sample sizes are small. Treat this as a directional signal from real traffic, not a lab result.
| Per call | Cost at list price |
|---|---|
| Cheapest | $0.00000 |
| Median | $0.08331 |
| Mean | $0.08507 |
| Most expensive | $0.26128 |
34 calls · 328,905 input tokens · 35,159 output · 386,176 cache-read · $2.89 total at list price. Regenerated from billing data.
Median $0.08331 against mean $0.08507. Where those two diverge sharply — as they do on claude-sonnet-5 — it is cache behaviour talking, not the model: most calls read a cached prefix cheaply while a few rebuild it expensively, and the mean carries the rebuilds.
First token averaged 11,953 ms across these calls, and the slowest was 71,337 ms. Total durations ran to 135 seconds on long generations.
The three GPT-5.6 tiers differ enormously in price: $5.00 / $30.00 for sol, $2.00 / $12.00 for terra, $0.49 / $2.01 for luna. That is a ten-times spread from top to bottom, for the same API shape and one word different in your request.
Routing the routine steps down a tier usually moves blended cost closer to the cheap model than the expensive one, because routine calls dominate the count.
from openai import OpenAI
client = OpenAI(api_key="YOUR_KEY", base_url="https://apiclan.us/v1")
resp = client.chat.completions.create(
model="gpt-5.6-sol",
messages=[{"role": "user", "content": "Hello"}],
stream=True, # first token is what the user feels
)
The model string keeps the dot: gpt-5.6-sol. A misspelled model
name returns a 404 from the edge, which means the request never reaches the gateway
and never appears in your usage log — it looks exactly like a bad key. If a
call fails and nothing shows up in your usage history at all, check the spelling and
the base URL first.
| Per 1M tokens | List price | APICLAN | You save |
|---|---|---|---|
| Input | $5.00 | $1.00 | 80% |
| Output | $30.00 | $6.00 | 80% |
Calculator and the rest of the family on the price page.
We measured it, we did not cause it, and we cannot see inside the provider's routing. What we can say is that in our traffic the correlation between price and first-token time was negative. Measure it on your own workload before designing around either assumption.
Terra unless output quality is the product. It averaged about a third of the cost per call here. Run both on twenty real inputs from your own product and read the outputs — that settles it faster than any benchmark.
That is your client's default, not a failure. Long generations in this sample ran to 135 seconds. Raise the timeout or switch to streaming.
No subscription, no monthly minimum, no sales call. Top up with USDT and spend what you use — 1 USDT gives you 2 credits of API balance.
Read the 30-second quickstartPrices quoted on this page are regenerated automatically from live billing data. Third-party terms are quoted from that party's own published documentation.