OpenAI API Models for Reasoning, Coding and Multimodal AI
OpenAI's models cover more surface than any other family in this catalogue: reasoning and coding models across three capability tiers, image generation and editing, realtime speech, high-accuracy transcription, and text embeddings. The reasoning models share one property that shapes how they are used — deliberation is a per-request setting rather than a property of the model, so the same identifier can serve a real-time endpoint and a long analytical job. They accept text and images against context windows beyond a million tokens, support structured outputs and function calling, and are reachable on DEVUP AI through the standard OpenAI-compatible endpoint, billed in Algerian Dinar.
OpenAI's models cover more surface than any other family in this catalogue: reasoning and coding models across three capability tiers, image generation and editing, realtime speech, high-accuracy transcription, and text embeddings. The reasoning models share one property that shapes how they are used — deliberation is a per-request setting rather than a property of the model, so the same identifier can serve a real-time endpoint and a long analytical job. They accept text and images against context windows beyond a million tokens, support structured outputs and function calling, and are reachable on DEVUP AI through the standard OpenAI-compatible endpoint, billed in Algerian Dinar.
Available OpenAI Models
Deploy and access all published models in this family with unified DZD pricing and zero foreign card requirement.
| Model | Context | DZD in / 1M tokens | DZD out / 1M tokens | Actions |
|---|---|---|---|---|
| gpt-5.6-solopenai/gpt-5.6-sol | 1M Tokens | 1400 | 7000 | View model |
| gpt-5.6-terraopenai/gpt-5.6-terra | 1M Tokens | 700 | 4200 | View model |
| gpt-5.6-lunaopenai/gpt-5.6-luna | 1M Tokens | 70 | 420 | View model |
| gpt-6-astraopenai/gpt-6-astra | 1M Tokens | 3500 | 17500 | View model |
Call any OpenAI model
Every model on this page uses the same API key and standard OpenAI-compatible endpoint. The example below uses the featured model openai/gpt-5.6-terra.
from openai import OpenAI
import os
client = OpenAI(
api_key=os.environ["DEVUP_API_KEY"],
base_url="https://api.devupai.com/v1"
)
response = client.chat.completions.create(
model="openai/gpt-5.6-terra",
messages=[
{"role": "user", "content": "Hello world!"}
],
max_tokens=1024
)
print(response.choices[0].message.content)import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.DEVUP_API_KEY,
baseURL: "https://api.devupai.com/v1",
});
async function main() {
const response = await client.chat.completions.create({
model: "openai/gpt-5.6-terra",
messages: [
{ role: "user", content: "Hello world!" },
],
max_tokens: 1024,
});
console.log(response.choices[0]?.message?.content);
}
main();curl -X POST "https://api.devupai.com/v1/chat/completions" \
-H "Authorization: Bearer $DEVUP_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-5.6-terra",
"messages": [
{ "role": "user", "content": "Hello world!" }
],
"max_tokens": 1024
}'Frequently Asked Questions
Because the accepted levels differ between generations. The current flagship generation rejects the zero-reasoning setting entirely, while the generation below it accepts it. A value that is documented in some places is also rejected across every model in these families. Validate the effort level per model in your own code — the error returned does not name which values are accepted, so a wrong value produces an unhelpful message rather than a clear one.
For text and reasoning, start with a mid-tier model at the middle effort setting and adjust from there. It carries most production work well, and because tiers within a generation share an interface, moving up or down later is a change to one field. Move up when a specific input has already failed; move down when the task is procedural rather than deductive. For images, speech, or embeddings, the family has dedicated models — check the individual model pages.
No, and this is measured rather than theoretical. On a multi-step reasoning task the lowest reasoning setting was both more expensive than switching reasoning off and less accurate than it — the model deliberated for hundreds of tokens and reached the wrong answer every time. Accuracy became reliable at the middle setting. Either remove deliberation or give it a real budget.
The current reasoning models accept images alongside text in the same request, which covers document understanding, screenshot and interface reading, and chart analysis. They do not accept audio or video, and they output text only. Image generation and editing, speech, and transcription are handled by separate dedicated models in the same family.
For most models in this family, no — no parameter counts, no layer structure, no training details, and no downloadable weights. What is documented is behaviour: input types, context and output limits, effort levels, and supported features. OpenAI has released open-weight models under a permissive licence as an exception; where those appear in this catalogue, their model pages state it explicitly.
The current reasoning models specify context windows beyond a million tokens with a 128,000-token output ceiling. Treat a specification as a property of the model rather than a guarantee from every path to it — client software, account tier, and rollout state all sit in between. Send a request near your intended ceiling and read what comes back before designing a pipeline around it.
Not for the core call. Text and vision models share the chat completions endpoint, embeddings use the embeddings endpoint, and audio models use their own. Within a generation, switching between reasoning tiers is a change to the model field. What varies is the accepted effort levels and the recommended settings, which the individual model pages document.
