In short:
- OpenAI released four GPT-6 models to the API in September 2026: GPT-6 Astra, GPT-6 Sol, GPT-6 Luna and, at DevDay on September 29, GPT-6.1 Sol.
- GPT-6.1 Sol lists at $2 per million input tokens and $10 per million output, with cached input at $0.10. GPT-6 Luna is the low-cost option at $0.10 / $0.50.
- DevDay added a paid Ultrafast speed tier and a beta Agents API. On October 6 OpenAI released the Decisions API in beta and cut its usage tiers from five to three.
- Older models are on a clock. Many GPT-4-era and o-series models shut down on October 23. We cover that list in a separate post.
The four models
OpenAI’s API changelog records the releases in this order.
GPT-6 Astra (gpt-6-astra) arrived first, in early September. OpenAI calls it its “most capable model, built for the hardest end-to-end work”. It is available through the Responses API and Chat Completions. The same changelog entry added three Responses API features: async tool calling, steering a response mid-turn, and changing reasoning effort mid-conversation.
GPT-6 Sol (gpt-6-sol) and GPT-6 Luna (gpt-6-luna) followed on September 22. Both are reasoning models that accept text and images and return text. The changelog lists their prices per million tokens:
| Model | Input | Cached input | Output |
|---|---|---|---|
gpt-6-sol |
$2.00 | $0.20 | $10.00 |
gpt-6-luna |
$0.10 | $0.01 | $0.50 |
gpt-6.1-sol |
$2.00 | $0.10 | $10.00 |
Three days later OpenAI fixed an image-encoding bug that had degraded image understanding in both Sol and Luna. If you evaluated either model on screenshots or diagrams before September 25, run that evaluation again.
GPT-6.1 Sol (gpt-6.1-sol) shipped at DevDay on September 29. OpenAI describes it as a step up for coding and computer use. The changelog lists $2 input, $0.10 cached input, $2.50 per million for cache writes and $10 output. Those prices apply to prompts up to 272K tokens, so check the pricing page before sending very long prompts. It is available through both the Responses API and Chat Completions.
DevDay: speed tiers and agents
Most of the DevDay news was about speed and hosted agents rather than new models:
- Ultrafast is a paid service tier. You request it with
service_tier: "ultrafast"in the Responses API. At launch it covered GPT-6 Astra, with Sol to follow. OpenAI says it generates tokens up to 6× faster in the API and up to 8× faster in Codex. These are OpenAI’s own figures, not independent measurements. - OpenAI also said that time to first token in the API had fallen by 45% and that tool calls and workflows ran over 30% faster. Again, these are vendor claims. Measure your own latency before you change timeouts.
- The Agents API went to public beta on September 10 with durable sessions that continue work across turns. At DevDay it gained computer use in an OpenAI-hosted browser.
October: the Decisions API and simpler usage tiers
On October 6 OpenAI released the Decisions API in beta. It runs on GPT-6 Luna and turns text and images into typed answers from a set of options you define. Typical uses are classification, routing a request and choosing an action. OpenAI says it answers 10× faster than the Responses API for this kind of task. The DevDay post described it as a limited preview, and the changelog now lists it as a beta, so check your account’s access before you plan around it.
The same day OpenAI simplified API usage tiers from five to three: Build, Launch and Grow. Accounts move up automatically based on credit purchases. If your code reads tier names or assumes old rate limits, review it.
On October 7 OpenAI updated the chat-latest alias and repeated its advice to use the GPT-6 family, not chat-latest, for production work. A floating alias can change behaviour under you, so pin an exact model ID.
Two smaller changes worth handling
Clearer overload errors (September 2). OpenAI now separates two failure modes. Sending traffic too fast returns 429 with the code slow_down. Model overload returns 503 with server_is_overloaded. Both are worth retrying with backoff, and neither is worth retrying instantly in a tight loop. Here is a retry wrapper that works against any OpenAI-compatible endpoint, including ours:
import random, time
from openai import OpenAI, APIStatusError
client = OpenAI(base_url="https://api.oneaiapikey.com/v1", api_key="sk-your-key", max_retries=0)
def chat_with_backoff(messages, model="gpt-6.1-sol", attempts=5):
for attempt in range(attempts):
try:
return client.chat.completions.create(model=model, messages=messages)
except APIStatusError as e:
# 429 = you are sending too fast; 503 = the model is overloaded.
if e.status_code not in (429, 503) or attempt == attempts - 1:
raise
retry_after = e.response.headers.get("retry-after")
delay = float(retry_after) if retry_after else min(2 ** attempt, 30)
time.sleep(delay + random.uniform(0, 0.5)) # jitter avoids retry storms
reply = chat_with_backoff([{"role": "user", "content": "Say ok"}])
print(reply.choices[0].message.content)A temporary GPT-5.6 Sol price cut (August 21). OpenAI lowered gpt-5.6-sol to $4 per million input tokens and $20 per million output. The changelog says the reduction runs through November 21, 2026. If you budgeted at the new price, set a reminder for that date.
What this means if you use Relaykit
The GPT models in our catalogue today:
- GPT-6.1 Sol (
gpt-6.1-sol): $2.20 input / $11 output per million tokens at our published rate. - GPT-5.6 Sol (
gpt-5.6-sol): $5.50 input / $33 output per million tokens at our published rate. - GPT-5.6 Terra (
gpt-5.6-terra): $2.20 input / $13.20 output per million tokens at our published rate.
GPT-6.1 Sol is available under its standard model ID through our OpenAI-compatible Chat Completions endpoint. GPT-6 Astra, GPT-6 Sol and GPT-6 Luna are not in our catalogue at the time of writing.
Some of what OpenAI announced is tied to OpenAI’s own platform and is not available through Relaykit. That includes the Ultrafast service tier, the Agents API, the Decisions API and OpenAI’s usage tiers. We document and test Chat Completions and the Anthropic Messages format. Check the API reference before you rely on any other endpoint.
Our rates are our own published prices. They are not OpenAI’s list prices and do not follow OpenAI’s temporary discounts automatically. The models page shows the current rate, and the usage log in the console shows what each request cost.
Relaykit is an independent service and is not affiliated with OpenAI.