ASI:One Models
ASI:One offers three models that share the same API surface, the same API key, and the same capabilities. They differ in how deeply they think, how fast they answer, and how much they cost. Switching between them is a one-line change.
The model field is the only thing that changes between them:
asi1
asi1-ultra
asi1-mini
Choosing a model
Start with asi1. Move off it only when you have a reason:
- Is latency or cost your binding constraint? Use
asi1-mini. If you find yourself wishing the answers were longer or more thorough, you have outgrown it for that workload. - Is the task hard enough that quality beats speed? Use
asi1-ultra, and expect noticeably higher latency. - Otherwise use
asi1.
Most applications end up using more than one: asi1-mini for the high-volume path,
asi1 or asi1-ultra for the requests that need care.
Comparison
Everything not in that table is the same across all three:
streaming, tool
calling, reasoning
(the controls differ, support does not), image
input (same
formats, same limits and same errors), structured
data, and OpenAI
compatibility on both
/v1/chat/completions and
/v1/responses.
Planner mode is the exception: it
runs on /v1/chat/completions, on all three models.
Latency and cost above are relative to each other rather than absolute. Your account’s usage and balance are in your ASI:One account.
Pricing
Prices are per million tokens, in US dollars. Every model bills input tokens, output tokens, and reasoning tokens separately; see Reasoning for how reasoning billing works.
Cached input tokens are billed at 10% of the input rate, and reasoning tokens are billed at the output rate for the model.
Context windows
The context window is the combined budget for your input and the model’s
output. A 250,000-token prompt to asi1-mini leaves only about 12,000 tokens of
headroom for the response, so leave room for the answer you want.
Over-length requests are rejected rather than truncated. ASI:One does not drop your earliest messages to make an oversized prompt fit, so trim your own history before you approach the limit.
There is no separate cap on output length. max_tokens bounds a single response
if you set it, but the real ceiling is whatever the context window leaves after
your input.
One exception, on asi1 and asi1-ultra: reasoning counts against the same
budget, and when a response spends its entire max_tokens on reasoning without
producing an answer, the platform retries with a larger allowance - never more
than 32,768 tokens - so the response can still contain one. A response
recovered this way can therefore exceed the max_tokens you asked for, and its
usage reports (and bills) every token that was generated, across all attempts.
Tools that compact the conversation themselves need a window to compact against, and the full window is the wrong number to give them. Because input and output share it, a tool told it has the whole window only compacts once there is no room left to answer in.
So ASI:One publishes a second figure per model, a working window, sized to
leave that room: asi1 300,000 tokens, asi1-ultra 800,000 and asi1-mini
160,000. Treat it as the budget for everything you send, with the remainder
covering the reply and any tool output.
Codex asks ASI:One for these and configures itself. OpenCode, OpenWork and Zed cannot, so their setups paste the same values in.
Sampling defaults
Each model is tuned for the values it was evaluated at. Send no sampling fields and you get these:
temperature and top_p are the two you can set per request. Each replaces the
matching default and nothing else, so a request that sets only temperature
keeps every other tuned value. The other values in the table are fixed per
model.
On asi1 and asi1-ultra, prefer to leave both alone. If you do need to move
one, change temperature or top_p, not both - these models are evaluated
at 1.0 and 0.95, and moving both together degrades them.
asi1-mini carries a second set for when it is thinking. Sending
enable_thinking: true switches temperature to 0.6, top_p to 0.95 and
presence_penalty to 0, and keeps the rest. This is worth knowing before you
pin a value: an explicit temperature wins over both sets, so a request that
hard-codes 0.7 stays at 0.7 even once thinking is on, where the default would
have moved.
When to deviate at all:
- Output reads flat or repetitive. Raise
temperatureonasi1-mini, which tunes well between 0.7 and 1.0. Onasi1andasi1-ultrastay at 1.0 and steer with the prompt instead. - You want repeatable output for extraction, classification or tests. On
asi1-mini, turn thinking off and lowertemperaturetoward 0. Onasi1andasi1-ultralowertop_pinstead, and change only that one.
List the models from the API
GET /v1/models returns the models in OpenAI’s model list format, so OpenAI SDKs
and tools that discover models can find them:
GET /v1/models/{model} returns one model, and a 404 with the code
model_not_found for an id that is not in the list. See the API
Reference for both.
What each model is for
asi1
General chat and assistants, tool calling, and mixed workloads where you cannot tell in advance how hard any individual request will be.
asi1-ultra
Work where quality matters more than speed: deep research and synthesis, code and contract review, long multi-step runs, and strategic planning.
asi1-mini
Real-time chat and voice, classification and routing, autocomplete, and high-volume paths where each request is straightforward on its own. What you trade for the speed is depth and response length.
Switching models
Change one field. Everything else about the request stays the same.
The same API key works for all three, so you can route per request rather than per application. Two patterns worth knowing:
- Cascade. Send everything to
asi1-minifirst and retry onasi1orasi1-ultrawhen the answer is not good enough. - Route by request. Classify the incoming request and pick the model up front.
When you change models, A/B the workloads you care about and measure quality rather than assuming the more capable model is better for every task. On simple requests it is mostly slower.
Next steps
- Chat Completions API - The default endpoint in full, including streaming
- Reasoning - Control how deeply the model thinks, and what that costs
- Tool Calling - Let the model call your own functions
- Planner Mode - Multi-step work against Agentverse agents
- Tuning - Which dial to reach for when output is not what you wanted