gpt-oss-20b APIOPEN WEIGHTS131K context

Pricing & specs mirror our live pricing as of July 2026 (pay-as-you-go).

What is 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...

gpt-oss-20b on the release timeline

gpt-oss-20b pricing

Pay-as-you-go on DataLLM Lab — these are our live list prices (identical to the pricing page):

On output price, gpt-oss-20b is cheaper than 94% of the 309 models in the catalog. It is served by 13 providers; the best combined rate at our last snapshot was ≈ $0.03 / $0.13 per 1M via WandB, with upstream quantizations bf16, fp4, fp8.

What gpt-oss-20b costs per month

Cost = input price × input volume + output price × output volume. The same five workloads run on every model page, so any two compare directly.

gpt-oss-20b vs alternatives

Specs

How to call gpt-oss-20b

Copyfrom openai import OpenAI
client = OpenAI(base_url="https://api.datallmlab.com/v1", api_key="YOUR_DATALLM_LAB_KEY")
resp = client.chat.completions.create(
    model="openai/gpt-oss-20b",
    messages=[{"role": "user", "content": "Hello"}],
    # supports tools=[...] and tool_choice
)
print(resp.choices[0].message.content)

Copycurl https://api.datallmlab.com/v1/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"openai/gpt-oss-20b","messages":[{"role":"user","content":"Hello"}]}'

Copyimport OpenAI from "openai";
const client = new OpenAI({ baseURL: "https://api.datallmlab.com/v1", apiKey: process.env.DATALLM_KEY });
const r = await client.chat.completions.create({
  model: "openai/gpt-oss-20b",
  messages: [{ role: "user", content: "Hello" }],
});
console.log(r.choices[0].message.content);

The same key routes gpt-oss-20b and 300+ other models — switch models by changing one string. How OpenAI-compatible APIs work.

Prompting tips for gpt-oss-20b

• State the goal, not every step. gpt-oss-20b reasons internally — give it the objective, constraints and success criteria and let it plan the approach; over-scripting each step tends to lower quality. Turn effort up for hard problems.
• Give it real tools. Pass a tools=[...] schema instead of asking it to "pretend" — let gpt-oss-20b emit tool calls, execute them, and feed results back for the next turn.
• Constrain JSON with a schema. Use response_format / structured outputs rather than "reply in JSON" to get valid, parseable objects every time.
• Cheap enough to batch. Great for high-volume classification, extraction and drafting — template the prompt, batch requests, and validate outputs programmatically.
• Use the system role for rules. Put persistent instructions in the system message and keep user turns task-specific; pair with structured outputs for reliable automation.

Tips are derived from gpt-oss-20b's actual capabilities — context window, tool & JSON support, modality and price tier — not generic advice.

Open-source tools for gpt-oss-20b

Popular open-source projects for running and building with gpt-oss-20b — star counts pulled from GitHub (July 2026).

Listed by GitHub stars; inclusion is by ecosystem relevance (inference engines, agent frameworks and SDKs), not affiliation. Stars change — see each repo for current numbers.

Migrating to gpt-oss-20b

Coming from OpenAI or another gateway? On an OpenAI-compatible setup the only changes are base_url → https://api.datallmlab.com/v1 and model → openai/gpt-oss-20b. Messages, streaming, tool calls and the rest of your code stay the same. Routing & failover guide.

Self-host or use the API?

gpt-oss-20b ships open weights (openai/gpt-oss-20b, ~7.0M downloads and 4.8K likes in the last 30 days), with community quantizations (bf16, fp4, fp8) for smaller GPUs, so you can run it yourself. Most teams still use the API: no GPU to provision or keep warm, no inference ops, and instant access across 13 providers with automatic failover. Self-host when data residency or fixed per-token economics matter most.

Rate limits & reliability

Rate limits here are the DataLLM Lab gateway's, not the upstream vendor's. On 429 (rate limited) or 503 (provider busy), retry with exponential backoff; because gpt-oss-20b is served by 13 providers, requests can fail over to a healthy one automatically. See the error-code guide and failover setup.

Related reading

Frequently asked questions