GPT-OSS 120B, hosted and metered per token
OpenAI's 120B open-weight model is the catalog's general-purpose workhorse: reasoning, tool calling, and reliable structured output, without the H100 cluster it takes to serve at home. Point the OpenAI SDK here and it runs.
- OpenAI open-weight
- 128K context
- Structured output
- Batch at 50% off
Per token, nothing else.
Input / output per 1M tokens
$0.15 / $0.60
Batch jobs: $0.075 / $0.30 (50% off, 24h window)
openrelay/gpt-oss-120bFull catalog on the inference pricing page. Deposit $5 to get $10.
from openai import OpenAI
client = OpenAI(
base_url="https://inference.openrelay.inc/v1",
api_key="vl_••••••••", # same SDK, new base URL
)
resp = client.chat.completions.create(
model="openrelay/gpt-oss-120b",
messages=[{"role": "user", "content":
"Extract the lease terms as JSON: {tenant, rent_usd, term_months}."}],
response_format={"type": "json_object"},
)
print(resp.choices[0].message.content)The spec sheet.
What the model is, what it takes in, and the surface it serves on.
- Model id
- openrelay/gpt-oss-120b
- Parameters
- 120B (sparse MoE, open weights)
- Context
- 128K tokens
- Endpoint
- POST /v1/chat/completions (streaming + non-streaming)
- Capabilities
- Reasoning, tool calling, JSON mode, cached input rates
- Cached input
- Prompt prefixes served from cache bill at 10x below fresh input
Where GPT-OSS 120B earns its place.
What this model is actually for, versus the rest of the catalog.
The default for agents and pipelines
Strong tool calling plus dependable JSON output is the combination agent frameworks and extraction pipelines actually need. This is the most-run model on the platform for a reason.
Serving it yourself is the expensive part
120B weights want multi-GPU nodes, tensor parallelism, and someone on call. Per-token hosting turns that into $0.15/$0.60 per 1M with zero infrastructure, and cached-input pricing rewards long shared prompts.
Open weights, no lock-in
The model is downloadable and self-hostable the day you outgrow an API, so building on it is not a bet on any one vendor, including us.
Run it in batch at half price.
GPT-OSS 120B is the recommended generator and extractor across the batch workload recipes: $0.075 in / $0.30 out per 1M tokens on batch jobs, JSONL in and out.
Batch Inference API overviewGPT-OSS 120B, answered.
What is GPT-OSS 120B?
OpenAI's 120B-parameter open-weight model: a sparse mixture-of-experts released for anyone to run, with reasoning, tool calling, and structured output. Here it is served behind the standard OpenAI-compatible API, so the SDK you already use works unchanged.
What does the GPT-OSS 120B API cost?
$0.15 per 1M input tokens, $0.60 per 1M output, with cached input at $0.015. Batch jobs bill at half the realtime rates. Per-token, no subscription, no GPU rental.
GPT-OSS 120B vs 20B: which one?
20B for high-volume mechanical work (classification, sentiment, moderation) at a third of the price; 120B when the task needs reasoning depth, nuanced instructions, or strict schemas over messy input. Many pipelines screen with 20B and escalate the hard 10% to 120B.
Can I run GPT-OSS 120B myself instead?
Yes, the weights are open, and that portability is part of the point. Practically it wants a multi-GPU node and serving expertise; most teams prototype and run production on a per-token API and revisit self-hosting when volume justifies dedicated hardware.
First request in five lines.
Point the OpenAI SDK at inference.openrelay.inc/v1 and run GPT-OSS 120B per token. No contract, no minimums.