Skip to content
All models
llm
OpenAI

gpt-oss-20b (self-hosted)

OpenAI gpt-oss-20b · gpt-oss

OpenAI's open-weight 20B reasoning model under Apache-2.0, served on our own hardware - strong general reasoning, coding and agentic tool use without data leaving the deployment.

Decision and plan target (no provider attached) On Prem boundary PII allowed Tool calling
Price per 1M tokens
$0.40
$0.0004 per 1k
Requests · 30d
0
no traffic
Latency
1800 ms
declared
Success rate
n/a
quality prior 0.78

Underlying model

Published by the vendor; the platform bills at the target's declared price.

OpenAI

gpt-oss-20b

reasoning open weights tools

OpenAI openai/gpt-oss-20b

gpt-oss-20b is an open-weight 21B parameter model released by OpenAI under the Apache 2.0 license. It uses a Mixture-of-Experts (MoE) architecture with 3.6B active parameters per forward pass, optimized for...

Input $/1M
$0.03
Output $/1M
$0.13
Context
131K
Max output
118K
Released
Aug 5, 2025
Modalities
text
Intelligence index-
Coding index21
Agentic index1

Cached input reads $0.03 per 1M · weights on Hugging Face as openai/gpt-oss-20b

Traffic over 30 days

Requests per day this target answered on this deployment, with the cost line.

No requests yet. Route something from the playground to see usage here.
Tokens
0
Spend
$0
Previous window
0 req

Routed for

Domains the signal layer detected on requests that ended here.

No routed requests yet. Declared domains: General, Marketing, Finance.

Capabilities

DomainsGeneral, Marketing, Finance
ActionsReasoning, Generation, Summarize, Qa
Complexity band0.35 - 0.70
Languages*
Context window128K tokens
Streamingyes

Policy constraints

Hard stops enforced before scoring.

Data boundaryOn Prem
RegionsUS
PIIallowed
Max input tokensunlimited
Tenantsall

Representative prompts

Examples the similarity strategy matches against.

  • Explain the trade-offs between speculative decoding and batching for a 20B model on one GPU.
  • Draft a technical blog post comparing three open-weight models for on-prem agents.
  • Plan the steps an agent needs to migrate a Flask app to FastAPI.

Route to it

This llm is not executed by the platform. Ask the router for a decision restricted to kinds=["llm"] and dispatch the winner in your own stack; send feedback afterwards so the learners improve.

curlbash
curl https://api.opensmartroute.ai/api/v1/route \
  -H "Authorization: Bearer $OSR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text": "Explain the trade-offs between speculative decoding and batching for a 20B model on one GPU.", "kinds": ["llm"], "top_k": 3}'

# Returns the chosen target, confidence, the ranked alternatives and the
# per-strategy trace. Execution of llm targets happens in your stack.
httpxpython
import httpx

r = httpx.post(
    "https://api.opensmartroute.ai/api/v1/route",
    headers={"Authorization": "Bearer osr_live_..."},
    json={"text": "Explain the trade-offs between speculative decoding and batching for a 20B model on one GPU.", "kinds": ["llm"]},
)
d = r.json()
print(d["target"]["id"], d["confidence"])
print([t["id"] for t in d["ranked"]])  # e.g. gpt-oss first when it fits