Requesty
Fireworks AI

gpt-oss-20b

GPT-OSS-20B is OpenAI's lightweight open-weight model with 21.5 billion total parameters (3.6 billion active, Mixture-of-Experts). Optimized for lower latency and local or specialized use cases including agentic tasks, chain-of-thought reasoning, and web browsing. Released under Apache 2.0 license.

ReasoningCaching
CompareDocs

Specifications

Context window131K tokens
Max output66K tokens
API typechat
AddedMay 27, 2026
Model IDfireworks/gpt-oss-20b
Data retentionNo
Used for trainingNo
Provider location🇺🇸 US
Privacy policyFireworks Privacy Policy→

Benchmarks

Released 2025-08-05
Coding Indexcoding
20.7%

Artificial Analysis Coding Index — a composite of coding evaluations including LiveCodeBench, SciCode and Terminal-Bench.

GPQA Diamondreasoning
68.8%

Graduate-level physics, chemistry & biology questions designed to resist Googling.

Intelligence Indexreasoning
14.9%

Artificial Analysis Intelligence Index — a composite of multiple evaluations measuring overall model capability.

Scores are sourced from official model cards, Artificial Analysis, and public leaderboards. Benchmarks measure specific skills and do not capture every aspect of model quality. Always test on your own workload.

Pricing

Input / 1M
$0.07
Output / 1M
$0.30
Cache write
N/A
Cache read / 1M
$0.04
Estimated cost
100K input + 10K output$0.0100
1M input + 100K output$0.10
10M input + 1M output$1.00

Requesty charges exactly what the upstream provider charges, no markup, no per-request fees. Prompt caching and smart routing can reduce effective cost by 30-80%.

Quickstart

Drop-in compatible with the OpenAI SDK. Change the base URL, swap in your Requesty API key, and set the model to fireworks/gpt-oss-20b.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
from openai import OpenAI

client = OpenAI(
    api_key="YOUR_REQUESTY_API_KEY",
    base_url="https://router.requesty.ai/v1",
)

response = client.chat.completions.create(
    model="fireworks/gpt-oss-20b",
    messages=[
        {"role": "user", "content": "Explain quantum computing in one paragraph."},
    ],
)

print(response.choices[0].message.content)

Other Fireworks AI models

glm-5.2
1M ctx$1.40/1M
Coding 69
qwen3.7-plus
262K ctx$0.40/1M
Coding 56
nemotron-3-ultra-nvfp4
262K ctx$0.60/1M
minimax-m3
512K ctx$0.30/1M
Coding 59
kimi-k2.7-code
262K ctx$0.95/1M
Coding 61
gpt-oss-120b
131K ctx$0.15/1M
Coding 30

Frequently asked questions

How much does gpt-oss-20b cost?
gpt-oss-20b is priced at $0.07 per million input tokens and $0.30 per million output tokens when accessed via Requesty. Prompt caching is supported, which can cut effective input cost by up to 90% on repeated context. Requesty charges exactly what the upstream provider charges, we don't add markup.
What is the context window of gpt-oss-20b?
gpt-oss-20b has a context window of 131K tokens, with a maximum output of 66K tokens per response. That's roughly 175 words of input you can fit in a single prompt.
How does gpt-oss-20b perform on benchmarks?
gpt-oss-20b scores 89.3% on Math Index, 89.3% on AIME 2025, 77.7% on LiveCodeBench. See the full benchmark chart above for results across MMLU Pro, GPQA Diamond, SWE-Bench Verified, HumanEval, MATH, AIME, MMMU, and LiveBench.
What can gpt-oss-20b do?
gpt-oss-20b supports extended reasoning, prompt caching. You can call it through any OpenAI-compatible client by pointing base_url to Requesty.
How do I use gpt-oss-20b with the OpenAI SDK?
Install the OpenAI SDK, set base_url to "https://router.requesty.ai/v1", set your API key to your Requesty key, and set the model to "fireworks/gpt-oss-20b". The Quickstart above shows Python, JavaScript and cURL snippets.
Can I run gpt-oss-20b through Requesty?
Yes. gpt-oss-20b runs through Requesty's OpenAI-compatible API, served from Fireworks AI. You do not host the model yourself: point base_url at Requesty, set the model to "fireworks/gpt-oss-20b", and requests are routed to the upstream provider with automatic failover. The same key gives you 400+ other models too.

Access gpt-oss-20b through Requesty

One API key, 400+ models, OpenAI-compatible. No markup on provider prices, automatic failover, and smart caching built-in.

All Fireworks AI models