GPT-OSS 20B, built for the million-row job
OpenAI's 20B open-weight model is what you run when the task is clear and the volume is huge: classification, sentiment, moderation, short extraction. Fast, obedient to output formats, and priced so exhaustive processing beats sampling.
- OpenAI open-weight
- 128K context
- Best value in the catalog
- Batch at 50% off
Per token, nothing else.
Input / output per 1M tokens
$0.05 / $0.20
Batch jobs: $0.025 / $0.10 (50% off, 24h window)
openrelay/gpt-oss-20bFull 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-20b",
messages=[
{"role": "system", "content":
"Classify into exactly one of: billing, bug, feature_request. JSON: {label, confidence}."},
{"role": "user", "content": "I was charged twice this month."},
],
)
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-20b
- Parameters
- 20B (open weights)
- Context
- 128K tokens
- Endpoint
- POST /v1/chat/completions (streaming + non-streaming)
- Capabilities
- Tool calling, JSON mode, cached input rates
- Sweet spot
- High-volume labeled outputs: classify, score, flag, extract
Where GPT-OSS 20B earns its place.
What this model is actually for, versus the rest of the catalog.
The economics of analyzing everything
At $0.05/$0.20 per 1M realtime and half that in batch, a 150-token record with a short structured verdict costs about a thousandth of a cent. Sampling stops being a cost decision.
Small model, disciplined output
For label-from-a-list, score-on-a-scale, JSON-with-these-fields work, 20B tracks its bigger sibling closely while running faster and cheaper. The gap only opens on genuine reasoning.
Pairs with 120B as a two-tier pipeline
Screen everything with 20B, escalate its low-confidence or flagged rows to 120B in a second pass. Most content is easy; pay the big model only for the gray zone.
Run it in batch at half price.
This is the model the classification, sentiment, and moderation batch recipes are built around: $0.025 in / $0.10 out per 1M tokens on batch jobs.
Batch Inference API overviewGPT-OSS 20B, answered.
What is GPT-OSS 20B good for?
High-volume tasks with a clear instruction and a short structured answer: ticket routing, sentiment with aspects, policy screening, entity extraction from short text, paraphrase generation. For deep reasoning or gnarly schemas over messy documents, step up to 120B.
What does the GPT-OSS 20B API cost?
$0.05 per 1M input tokens and $0.20 per 1M output, cached input at $0.005, and half all of that on batch jobs. A 100,000-review sentiment run costs about $1.20 in batch.
Is GPT-OSS 20B good enough versus a bigger model?
Measure it: run a few hundred rows of your real task on both sizes and compare against a hand-labeled sample. On closed-set classification and scoring, 20B usually lands within a point or two of 120B at a third the price. That benchmark is one small batch job per model.
Can I self-host GPT-OSS 20B instead?
The weights are open and 20B fits on a single high-memory GPU, so yes. The per-token API wins on burst capacity, zero ops, and batch pricing; self-hosting wins at sustained saturation. Start on the API, measure, then decide.
First request in five lines.
Point the OpenAI SDK at inference.openrelay.inc/v1 and run GPT-OSS 20B per token. No contract, no minimums.