model comparison

vs

GPT-OSS 120B vs GPT-OSS 20B

two OpenAI models, compared on listed price, real workload cost, context and capabilities.

the short answer

  • on SWE-Bench Verified (high reasoning), as each maker reports it, GPT-OSS 120B scores 62.4% and GPT-OSS 20B 60.7%.
  • GPT-OSS 20B lists lower input and output prices.

side-by-side specs

specGPT-OSS 120BGPT-OSS 20B
providerOpenAIOpenAI
input price$0.35/M$0.03/M
output price$0.75/M$0.1/M
blended price (3:1)$0.45/M$0.048/M
context window128k tokens128k tokens
inputtexttext
structured outputyesyes
weightsopen (Apache 2.0)open (Apache 2.0)
releasedAugust 5, 2025—
SWE-Bench Verified (high reasoning) (maker-reported)62.4%60.7%
GPQA Diamond (no tools, high reasoning) (maker-reported)80.1%71.5%
HLE (no tools, high reasoning) (maker-reported)14.9%10.9%
AIME 2025 (no tools, high reasoning) (maker-reported)92.5%91.7%
Tau-Bench Retail (high reasoning) (maker-reported)67.8%54.8%

same work, both models

workloadGPT-OSS 120BGPT-OSS 20B
one chat reply3,000 tokens in, 500 outGPT-OSS 120B costs 10× as much as GPT-OSS 20B$0.0014$0.00014
summarise a 100-page report50,000 tokens in, 1,500 out, one requestGPT-OSS 120B costs 11× as much as GPT-OSS 20B$0.02$0.0017
1,000 classification calls500 tokens in, 20 out, per call$0.19$0.02
coding-agent session2M tokens in across ~60 turns, 60,000 out, no cache$0.75$0.07

at listed rates. token counts are typical sizes, not measurements.

when to pick which

GPT-OSS 120B

GPT-OSS 120B is the larger of OpenAI's two Apache 2.0 open-weight models, with 117B total and 5.1B active parameters. it takes text only and sits above the cheaper GPT-OSS 20B.

pick it if

  • you want Apache 2.0 open weights you could also self-host
  • you need adjustable reasoning effort: low, medium or high
  • you need function calling and structured outputs from an open model

skip it if

  • you need image input; it takes text only
  • you are on the free plan, which includes only the 20B sibling

GPT-OSS 20B

GPT-OSS 20B is the smaller of OpenAI's two open-weight reasoning models: 21B parameters with 3.6B active, text input only. it is the low-cost sibling of GPT-OSS 120B and is the GPT-OSS model on the free plan.

pick it if

  • you want low-cost text work at under a tenth of GPT-OSS 120B input
  • you want Apache 2.0 weights that run within 16GB of memory
  • you want adjustable reasoning effort: low, medium or high

skip it if

  • you need image input; it reads text only
  • prompts repeat heavily: no cached-input discount applies here

faq

which is better, GPT-OSS 120B or GPT-OSS 20B?

on SWE-Bench Verified (high reasoning), as each maker reports it, GPT-OSS 120B scores 62.4% and GPT-OSS 20B 60.7%. these are each maker's own numbers, not a controlled head-to-head; the spec table lists every benchmark both publish.

which is cheaper, GPT-OSS 120B or GPT-OSS 20B?

GPT-OSS 20B lists lower input and output prices. on a coding-agent session (2M tokens in, 60,000 out) GPT-OSS 120B costs $0.75 and GPT-OSS 20B $0.07.

can i switch between GPT-OSS 120B and GPT-OSS 20B mid-task?

yes. in an octomind session, /model octohub:gpt-oss-120b and /model octohub:gpt-oss-20b switch models without losing the conversation, so you can draft on one and review on the other.

try either model

switch between models inside the same octomind session.

prices from the catalog of 2026-11-18