Skip to content

OpenAI · Chat

GPT OSS 20b

OpenAI 🇺🇸ChatGPT familyOpen weights. One of its sources says the weights are publicly available.The maker lists this model as deprecated (15 Sep 2026).
From Cloudflare Workers AI's listing:
OpenAI’s open-weight models designed for powerful reasoning, agentic tasks, and versatile developer use cases – gpt-oss-20b is for lower latency, and local or specialized use-cases.
llminfo IQ
111.3
Max context
131K
Listings
133

Evidence and sources

🤗 openai/gpt-oss-20b Model architecture, as its sources state it llminfo IQ 111.3, ranked among chat models. From 251 benchmarks. Plenty of data behind it.How llminfo IQ is computedCoding IQ 88.2, 95% interval 76–101. From 12 benchmark families across 9 sources. Plenty of data behind it. Uses the model's best published results. It has its own scale, separate from llminfo IQ.How Coding IQ is computedFair value (48.9). Costs per task about what its benchmark results predict. From 22 benchmarks with measured task costs. Plenty of data behind it.How Value for money worksYap score: tokens used per coding task vs. models with the same resultsTerse (40 ±9). Generates fewer tokens than models with similar coding results. Measured at default effort. Little data behind it, so treat it as a rough estimate.Median 2,906 output tokens per task at default effort.How the Yap score worksReleased 5 Aug 2025 Training-data cutoff, as its sources state it Maker HQ: US (according to OpenAI's own website)
All rankings: #77 on OpenRouter this month #78 on OpenRouter this week #80 on OpenRouter today #127 on NanoGPT recentlySources: LMArena data, DeepInfra data, Cloudflare Workers AI, flex.ai API data, LiteLLM data, Models.dev data, AI/ML API data, OpenRouter data docs, llminfo researchPrimary data from LMArena from DeepInfra: name from Cloudflare Workers AI: architecture, description from flex.ai API: hugging face id from LiteLLM: deprecation date from Models.dev: weights availability from AI/ML API: primary mode from OpenRouter: knowledge cutoff from llminfo research: release date
Tags from its sources:instructreasoningtoolsopen-sourcecodingfunction-calling

Advantages and trade-offs

Compared with 10 chat models, picked mainly by popularity and IQ

Trade-offs

Deprecated by its maker

The maker has already deprecated this model (15 Sep 2026).

We only make a claim when there's enough data on models of the same type. Updated 22 Nov, 04:34 UTC.Prices come from the model's maker where they sell it, otherwise their official partner, otherwise the market median. This comparison uses 8 maker's prices; 2 market median prices.

Input & output

as stated by the model's providers
Text
Text

Capabilities

Supported operations and provider-stated features

ReasoningFunction callingStructured outputPrompt cachingWeb searchBatch API

Listings

 OpenAIgpt-oss-20bSource: GenAI Prices dataNot published by the provider$0.029$0.14——
 Cometapigpt-oss-20b-free · freeSource: CometAPI Pricing dataNot published by the provider$0.00$0.00——
 Fastrouteropenai/gpt-oss-20b:free · freeSource: FastRouter data131K$0.00$0.00——
 Flowdotopenai/gpt-oss-20b:free · freeSource: FlowDot data131K$0.00$0.00——
 OpenRouteropenai/gpt-oss-20b:free · freeSource: Helicone Costs dataNot published by the provider$0.00$0.00——
 Eden AIInference by Greenferencegreenference/gpt-oss-20bSource: Models.dev data131K$0.009cache read $0.0045$0.045——
 NanoGPTInference by GreenferenceRegions: 🇪🇺 EUopenai/gpt-oss-20b · fp8Source: NanoGPTNot published by the provider$0.0095cache read $0.0047$0.0473237.1307ms
 Sokkanopenai/gpt-oss-20bSource: Models.dev data66K$0.015cache read $0.008$0.075——
 Kiloopenai/gpt-oss-20bSources: Kilo AI Gateway data, Models.dev data131K$0.017cache read $0.0085$0.085——
 Nous Portalopenai/gpt-oss-20bSource: Nous Portal Inference Gateway data131K$0.017cache read $0.0085$0.085——

16 listings without a price are hidden.

Benchmarks

Named effort levels run from lowest to highest. Select a result to see its source details.

Safety

HELM66 related benchmarks · 66 results
Source: Stanford HELM specialized suites data docs
BenchmarkScore
Safety: BBQ - BBQ accuracy
AIR-Bench: AIRBench 2024 - #1: Confidentiality
AIR-Bench: AIRBench 2024 - #2: Integrity
AIR-Bench: AIRBench 2024 - #3: Availability
Show 62 more benchmarksHide 62 benchmarks
BenchmarkScore
AIR-Bench: AIRBench 2024 - #4: Automated Decision-Making
AIR-Bench: AIRBench 2024 - #5: Autonomous Unsafe Operation of Systems
AIR-Bench: AIRBench 2024 - #6: Advice in Heavily Regulated Industries
AIR-Bench: AIRBench 2024 - #7: Supporting Malicious Organized Groups
AIR-Bench: AIRBench 2024 - #8: Celebrating Suffering
AIR-Bench: AIRBench 2024 - #9: Violent acts
AIR-Bench: AIRBench 2024 - #10: Depicting Violence
AIR-Bench: AIRBench 2024 - #11: Weapon Usage & Development
AIR-Bench: AIRBench 2024 - #12: Military and Warfare
AIR-Bench: AIRBench 2024 - #13: Harassment
AIR-Bench: AIRBench 2024 - #15: Perpetuating Harmful Beliefs
AIR-Bench: AIRBench 2024 - #16: Offensive Language
AIR-Bench: AIRBench 2024 - #17: Adult content
AIR-Bench: AIRBench 2024 - #18: Erotic
AIR-Bench: AIRBench 2024 - #19: Non-Consensual Nudity
AIR-Bench: AIRBench 2024 - #20: Monetized
AIR-Bench: AIRBench 2024 - #21: Endangerment, Harm, or Abuse of Children
AIR-Bench: AIRBench 2024 - #22: Child Sexual Abuse
AIR-Bench: AIRBench 2024 - #23: Suicidal and non-suicidal self Injury
AIR-Bench: AIRBench 2024 - #24: Political Persuasion
AIR-Bench: AIRBench 2024 - #25: Influencing Politics
AIR-Bench: AIRBench 2024 - #26: Deterring Democratic Participation
AIR-Bench: AIRBench 2024 - #27: Disrupting Social Order (China-unique)
AIR-Bench: AIRBench 2024 - #28: High-Risk Financial Activities
AIR-Bench: AIRBench 2024 - #29: Unfair Market Practices
AIR-Bench: AIRBench 2024 - #30: Disempowering Workers
AIR-Bench: AIRBench 2024 - #31: Fraudulent Schemes
AIR-Bench: AIRBench 2024 - #32: Fraud
AIR-Bench: AIRBench 2024 - #33: Academic dishonesty
AIR-Bench: AIRBench 2024 - #34: Mis/disinformation
AIR-Bench: AIRBench 2024 - #35: Sowing Division
AIR-Bench: AIRBench 2024 - #36: Misrepresentation
AIR-Bench: AIRBench 2024 - #37: Types of Defamation
AIR-Bench: AIRBench 2024 - #38: Specific Types of Rights
AIR-Bench: AIRBench 2024 - #39-40: Discrimination/Protected Characteristics Combinations
AIR-Bench: AIRBench 2024 - #41-42: Privacy Violations/Sensitive Data Combinations
AIR-Bench: AIRBench 2024 - #43: Illegal/Regulated substances/goods
AIR-Bench: AIRBench 2024 - #44: Illegal services/exploitation
AIR-Bench: AIRBench 2024 - #45: Other Illegal/Unlawful/Criminal Activities
AIR-Bench: AIRBench 2024 - Child Harm
AIR-Bench: AIRBench 2024 - Criminal Activities
AIR-Bench: AIRBench 2024 - Deception
AIR-Bench: AIRBench 2024 - Defamation
AIR-Bench: AIRBench 2024 - Discrimination/Bias
AIR-Bench: AIRBench 2024 - Economic Harm
AIR-Bench: AIRBench 2024 - Fundamental Rights
AIR-Bench: AIRBench 2024 - Hate/Toxicity
AIR-Bench: AIRBench 2024 - Manipulation
AIR-Bench: AIRBench 2024 - Operational Misuses
AIR-Bench: AIRBench 2024 - Political Usage
AIR-Bench: AIRBench 2024 - Privacy
AIR-Bench: AIRBench 2024 - Refusal Rate
AIR-Bench: AIRBench 2024 - Security Risks
AIR-Bench: AIRBench 2024 - Self-harm
AIR-Bench: AIRBench 2024 - Sexual Content
AIR-Bench: AIRBench 2024 - Violence & Extremism
Safety: Anthropic Red Team - LM Evaluated Safety score
Safety: BBQ - BBQ (ambiguous)
Safety: BBQ - BBQ (unambiguous)
Safety: HarmBench - LM Evaluated Safety score
Safety: SimpleSafetyTests - LM Evaluated Safety score
Safety: XSTest - LM Evaluated Safety score

Multilingual

Bangla LLM BenchSource: Bangla LLM Bench data docs
MMMLU45 results
Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
official model card Table 2 · reasoning high · translated-answer parser
official model card Table 2 · reasoning low · translated-answer parser
official model card Table 2 · reasoning medium · translated-answer parser
Reasoning effort · official model card Table 3 · 14-language average · pass@1

Science

GPQASource: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
GPQA Diamond24 results
Source: Epoch AI Benchmarking Hub data
Reasoning effort
Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reasoning effort · official model card Table 3 · no tools · pass@1
Reasoning effort · official model card Table 3 · with tools; Table 3 does not name the exact tool · pass@1
Source: OpenEvals Leaderboard data · Reported by the model's developer. The source relaying it did not verify it. · Mar 2026
Setting not stated
Source: OpenRouter benchmark exports
gpt-oss-20b · 90-day results
Across 13 providers lowest 57.7% median 66.5% highest 68.7%OpenRouter's automatic routing 65.6%
Show results from 13 providers and OpenRouter's automatic routingHide results from 13 providers and OpenRouter's automatic routing
  • DekaLLM
  • AkashML
  • DeepInfra
  • CoreWeave
  • Parasail
  • Amazon Bedrock
  • Phala
  • OpenRouter's automatic routing
  • Darkbloom
  • SiliconFlow
  • Groq
  • PrimeIntellect
  • Novita
  • Google

Human preference

Coding

Aider PolyglotReasoning effort · official model card Table 3 · published setting; Table 3 does not further specify tool access · pass@1 · Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reasoning effort · official model card Table 3 · no tools · pass@1
Reasoning effort · official model card Table 3 · with tools; Table 3 does not name the exact tool · pass@1
Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reasoning effort · official model card Table 3 · published setting; Table 3 does not further specify tool access · pass@1
Source: Hugging Face Official Benchmarks data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated

Agentic

Source: OpenRouter benchmark exports
gpt-oss-20b · 90-day results
Across 8 providers lowest 1.5% median 42.5% highest 52.1%OpenRouter's automatic routing 53.2%
Show results from 8 providers and OpenRouter's automatic routingHide results from 8 providers and OpenRouter's automatic routing
  • OpenRouter's automatic routing
  • Darkbloom
  • DeepInfra
  • Groq
  • Parasail
  • AkashML
  • PrimeIntellect
  • Amazon Bedrock
  • CoreWeave
τ-bench AirlineReasoning effort · official model card Table 3 · developer functions · pass@1 · Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
τ-bench RetailReasoning effort · official model card Table 3 · developer functions · pass@1 · Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.

Math

AIME 20246 results
Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reasoning effort · official model card Table 3 · no tools · pass@1
Reasoning effort · official model card Table 3 · with tools; Table 3 does not name the exact tool · pass@1
AIME 20256 results
Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reasoning effort · official model card Table 3 · no tools · pass@1
Reasoning effort · official model card Table 3 · with tools; Table 3 does not name the exact tool · pass@1
ValsSource: Vals AI

Reasoning

Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reasoning effort · official model card Table 3 · no tools · pass@1
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated
Humanity's Last Exam (with tools)Reasoning effort · official model card Table 3 · with tools; Table 3 does not name the exact tool · pass@1 · Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.

Vision

Knowledge

MMLU4 results
Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reasoning effort · official model card Table 3 · published setting; Table 3 does not further specify tool access · pass@1
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated

General

Instruction following

Speech

Other

HealthBench3 related benchmarks · 12 results
HealthBench
Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reasoning effort · official model card Table 3 · published setting; Table 3 does not further specify tool access · pass@1
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated
Consensus
Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reasoning effort · official model card Table 3 · published setting; Table 3 does not further specify tool access · pass@1
Hard
Reported by OpenAI · via BenchBoard docs · Reported by the model's developer. The source relaying it did not verify it.
Reasoning effort · official model card Table 3 · published setting; Table 3 does not further specify tool access · pass@1
Source: HealthBench Hard (Arcophos) data docs · Reported by the model's developer. The source relaying it did not verify it. · Aug 2025
raw score (%), reasoning level high, gpt-oss model card Table 3. Not length-adjusted, unlike the GPT-5.x rows on this board.
Source: LLM Stats data · Reported by the model's developer. The source relaying it did not verify it. claim
Setting not stated
Sayf16 related benchmarks · 16 results
Show 10 more resultsHide 10 results
ValsSource: Vals AI

Data credits: BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links. · SEA-HELM leaderboard, AI Singapore (AI Products team), National University of Singapore. Published aggregate scores; results licence not stated (SEA-HELM code is MIT). · Epoch AI Benchmarking Hub, CC BY 4.0 (licence) · LMArena, CC BY 4.0 (licence) · Bangla LLM Bench, Md Deluair Hossen. Source tasks: Hendrycks et al. 2021 (MMLU), OpenAI MMMLU, Bandarkar et al. 2024 (Belebele), Shi et al. 2023 (MGSM), Nahin et al. 2025 (TituLLMs). Public numeric community results; results licence not separately stated. · BrowserGym Leaderboard (ServiceNow), MIT (licence) · ChemBench leaderboard, Jablonka group / lamalab-org. Public verified numeric result summaries; results licence not separately stated. · Data by LLM Stats · EQ-Bench (Samuel J. Paech) — cite paech2023eqbench · OpenEvals/leaderboard-data, MIT (licence) · OpenRouter public benchmark leaderboard and search-lane exports, CC BY 4.0. Fractional source scores converted to percentages; immutable dated snapshots. (licence) · Arcophos. HealthBench Hard Leaderboard. https://healthbenchhard.ai. CC BY 4.0. Source document linked on each result. (licence) · Stanford HELM (Stanford CRFM); scenario scores from Capabilities, Safety, Long Context, MedHELM, AIR-Bench, MMLU and Finance · Hugging Face official benchmark leaderboards; licence varies by benchmark · JMLE2026-Bench and IgakuQA119 by Naoto Iwase; results licence not stated (code MIT, question data CC BY 4.0). · LLM Chess by Maxim Saplin, Apache-2.0 (Copyright [2024] [Maxim Saplin]) (licence) · Sayf-eval. Copyright (c) 2026 Qatar Computing Research Institute (QCRI). MIT License. (licence) · Shaberi v2.1, Shisa.AI, Apache-2.0. GPT-5.1-judge scores only; judges are not comparable. (licence) · Swallow LLM Leaderboard v2, Swallow LLM team (Institute of Science Tokyo). Contents CC BY 4.0; Japanese post-trained task scores on the source's 0..1 scale. (licence) · Terminal-Bench (Laude Institute), Apache-2.0 (licence) · UzLiB, Tahrirchi (Abror Shopulatov). Public LEADERBOARD.md numbers; repository MIT covers software and documentation, no separate results licence stated and none asserted.

Also known as: @cf/openai/gpt-oss-20b, accounts/fireworks/models/gpt-oss-20b, bedrock/ap-northeast-1/openai.gpt-oss-20b-1:0, bedrock/ap-south-1/openai.gpt-oss-20b-1:0, bedrock/ap-southeast-3/openai.gpt-oss-20b-1:0, bedrock/eu-central-1/openai.gpt-oss-20b-1:0, bedrock/eu-south-1/openai.gpt-oss-20b-1:0, bedrock/eu-west-1/openai.gpt-oss-20b-1:0, bedrock/eu-west-2/openai.gpt-oss-20b-1:0, bedrock/sa-east-1/openai.gpt-oss-20b-1:0, bedrock/us-gov-east-1/openai.gpt-oss-20b-1:0, bedrock/us-gov-west-1/openai.gpt-oss-20b-1:0, bedrock_mantle/ap-northeast-1/openai.gpt-oss-20b, bedrock_mantle/ap-south-1/openai.gpt-oss-20b, bedrock_mantle/ap-southeast-2/openai.gpt-oss-20b, bedrock_mantle/ap-southeast-3/openai.gpt-oss-20b, bedrock_mantle/eu-central-1/openai.gpt-oss-20b, bedrock_mantle/eu-south-1/openai.gpt-oss-20b, bedrock_mantle/eu-west-1/openai.gpt-oss-20b, bedrock_mantle/eu-west-2/openai.gpt-oss-20b, bedrock_mantle/openai.gpt-oss-20b, bedrock_mantle/sa-east-1/openai.gpt-oss-20b, bedrock_mantle/us-gov-east-1/openai.gpt-oss-20b, bedrock_mantle/us-gov-west-1/openai.gpt-oss-20b, cloudflare/@cf/openai/gpt-oss-20b, databricks-gpt-oss-20b, databricks/databricks-gpt-oss-20b, databricks/databricks-gpt-oss-20b@eu, deepinfra/openai/gpt-oss-20b, fireworks/gpt-oss-20b, fireworks_ai/accounts/fireworks/models/gpt-oss-20b, fireworks_ai/gpt-oss-20b, flexai/gpt-oss-20b, flexai/gpt-oss-20b@eu, gmi/openai/gpt-oss-20b, gpt oss 20b, gpt-oss-20b-free, gpt-oss:20b, greenference/gpt-oss-20b, groq/gpt-oss-20b, groq/openai/gpt-oss-20b, novita/openai/gpt-oss-20b, nvidia_nim/openai/gpt-oss-20b, oci/openai.gpt-oss-20b, ollama/gpt-oss:20b-cloud, openai-gpt-oss-20b, openai.gpt-oss-20b, openai.gpt-oss-20b-1:0, openai/gpt-oss-20b, openai/gpt-oss-20b-1:0, openai/gpt-oss-20b-maas, openai/gpt-oss-20b:batch, openai/gpt-oss-20b:free, ovhcloud/gpt-oss-20b, parasail-gpt-oss-20b, replicate/openai/gpt-oss-20b, tensormesh/openai/gpt-oss-20b, together_ai/openai/gpt-oss-20b, us-gov.openai.gpt-oss-20b-1:0, vertex_ai/openai/gpt-oss-20b-maas, wandb/openai/gpt-oss-20b

Like this one. Tap or hover any dotted value, label or icon to see where it comes from and how it is measured. open details when tapped or hovered. How this data works