Get started

gpt-oss-20b

Sold by 12 sellers, timed at 249 tok/s, scored 20.7 on the Coding Index, on 2 Design Arena boards, and with results on 1 openly licensed board.

retrieved sha256 · 0bb53a77713f

Models/OpenAI/gpt-oss-20bUSD per 1M tokens · as OpenRouter routes it
openai/gpt-oss-20bOpenAI · base listing · GPT tokenizer
OpenRouter list price · the routed rate, as the catalog carried itretrieved 17 Nov 16:29 UTC from openrouter.ai ↗
Input$0.018per 1M tokens
Output$0.09per 1M tokens
Cached input$0.009per 1M tokens
Context131k131,072 tokens
What Vitna read · 3 providers, each from its own pricing page, kept as dated recordsin / out per 1M tokens
Googleopen weights$0.07 / $0.25confirmed 22 Nov 2026
AWSopen weights$0.07 / $0.30confirmed 22 Nov 2026
Parasailopen weights$0.04 / $0.20confirmed 22 Nov 2026

The same weights served by 3 providers, each at its own list price. OpenRouter's rate is its routed rate across whichever of them it uses, and it need not match any one of them.

What the list says · the fields OpenRouter publishes for this id11 fields
AuthorOpenAI
Listed
Context131,072 tokens
Max output117,964 tokens
Modalitiestext → text
TokenizerGPT
Reasoningalways on · effort high, medium, low, medium by default
Moderationnot moderated by OpenRouter
Knowledge cutoff2024-06-30
Other listingbatch variant
What the weights say · read from the model's own repository on Hugging Faceretrieved from huggingface.co ↗
Licenceapache-2.0
Parameters20.9B · 20,914,757,184 declared
Model classGptOssForCausalLM
Downloadnot gated
Published

What sellers charge

12 sellers · 11.76x from cheapest to dearest

One tick per seller, on a log scale of input price per million from $0.017 at Darkbloom to $0.20 at PrimeIntellect. 1 seller undercuts the $0.018 listed; 10 are dearer. A faint tick declares no number format. Every seller is named on the endpoint ledger below.

The same sellers grouped by the number format each declares, on the same scale. The fp4 group alone spans 98% of the range, so the cheap end is not cheap for its number format. A declaration is checked by nobody: this groups the rows and certifies none.

Measured speed

93,041 requests · 12 endpoints · 30-minute window
Throughput249 tok/sMedian output rate, weighted across the 12 endpoints serving it by the requests each answered.the speed board
First token428msMedian wait before the first token, over the same window and weighted the same way.how Vitna reads a figure
Endpoint spread31 to 491The slowest and the fastest route serving this model, in tokens per second. Speed is a property of the route, not of the weights.the board, fastest first

OpenRouter's own timing of its routed traffic, weighted by requests, over one 30-minute window read . Every route is on the endpoint ledger below.

Every endpoint

12 priced · 12 timed · cheapest first
  1. Darkbloomdarkbloom/fp8$0.017$0.085$0.0085fp8118k82 tok/s52808ms17,244 req99.93%
  2. AkashMLakashml/fp4 · listed$0.018$0.09$0.009fp4118k74 tok/s42711ms19,245 req100%
  3. DeepInfradeepinfra/bf16/flex$0.024$0.112$0.012bf16118knot matchednonenonenot matched90.47%
  4. DekaLLMdekallm/bf16$0.029$0.14$0.029bf16118k31 tok/s21.6556ms86 req43.44%
  5. CoreWeavecoreweave/fp4$0.03$0.13$0.03fp4118k237 tok/s159108ms12,176 req100%
  6. DeepInfradeepinfra/bf16$0.03$0.14$0.015bf16118knot matchednonenonenot matched99.79%
  7. Parasailparasail/fp4$0.03$0.15$0.02fp4118k68.5 tok/s134.32s258 req99.80%
  8. SiliconFlowsiliconflow/fp8$0.04$0.18not listedfp88k57 tok/s231.09s41 req97.07%
  9. Amazon Bedrockamazon-bedrock$0.07$0.15not listednot declared118k410 tok/s321394ms472 req100%
  10. Amazon Bedrockamazon-bedrock/eu-west-1$0.07$0.15not listednot declared118k177 tok/s94634ms27 req100%
  11. Groqgroq$0.075$0.30$0.0375not declared66k491 tok/s195202ms32,478 req100%
  12. PrimeIntellectprimeintellect/fp4$0.20$0.30$0.18fp433k136 tok/s62769ms363 req95.41%
  13. DeepInfranot matchednot matchednot matchednot matchedno offer59 tok/s261.62s6 reqno offer
  14. DeepInfranot matchednot matchednot matchednot matchedno offer114 tok/s80247ms10,645 reqno offer

Each track runs from a route's p90, the rate nine in ten of its requests met or beat, up to its median. Weighted by requests, the figure above reports where the traffic went rather than what one route does: 10 of the 12 timed routes are slower than it, and the middle route runs at 98 tok/s. The widest median-to-p90 gap is Parasail's, 5.27x.

DeepInfra publishes 2 offers and had 2 routes timed. The two documents share no route id, so those offers and routes are printed apart rather than paired.

Uptime is OpenRouter's own figure for each seller over the trailing day, read once a day, and the column prints the latest reading, taken . 11 of the 13 sellers with a reading fell under 99% on at least one of the 23 days held, 23 Oct 2026 to 21 Nov 2026: Google at 0.00% on 23 Oct 2026, DekaLLM at 43.44% on 21 Nov 2026, DeepInfra (deepinfra/bf16/flex) at 66.66% on 15 Nov 2026, SiliconFlow at 79.53% on 18 Nov 2026, and 7 more. Source: OpenRouter's public endpoints API, one reading a day for each seller: 1,211 sellers carried one on the latest of the 23 days held.

Source: OpenRouter's public endpoints API. 1257 sellers across 359 models when the mirror ran, 186 of them served by more than one. List prices, not what anybody paid; a discount is the seller's own. Read . Speed is OpenRouter's own timing over one 30-minute window, read .

What each seller accepts

20 parameters · 12 sellers
  1. include_reasoningevery seller
  2. max_tokensevery seller
  3. reasoningevery seller
  4. reasoning_effortevery seller
  5. temperatureevery seller
  6. top_pevery seller
  7. stop11 of 12not SiliconFlow
  8. tool_choice11 of 12not SiliconFlow
  9. tools11 of 12not SiliconFlow
  10. response_format10 of 12not Amazon Bedrock (amazon-bedrock) and Amazon Bedrock (amazon-bedrock/eu-west-1)
  11. structured_outputs10 of 12not Amazon Bedrock (amazon-bedrock) and Amazon Bedrock (amazon-bedrock/eu-west-1)
  12. top_k10 of 12not DekaLLM and Groq
  13. frequency_penalty9 of 12not Amazon Bedrock (amazon-bedrock), Amazon Bedrock (amazon-bedrock/eu-west-1), and Groq
  14. seed9 of 12not Amazon Bedrock (amazon-bedrock), Amazon Bedrock (amazon-bedrock/eu-west-1), and SiliconFlow
  15. presence_penalty8 of 12
  16. repetition_penalty7 of 12
  17. logit_bias5 of 12
  18. logprobs5 of 12
  19. top_logprobs5 of 12
  20. min_p3 of 12only DeepInfra (deepinfra/bf16), DeepInfra (deepinfra/bf16/flex), and PrimeIntellect

6 of these 20 are accepted by every one of the 12 sellers that publish a list; the other 14 are declared by some sellers and not others, so what a request sending one of them gets depends on which seller serves it. Each list is the seller's own declaration in OpenRouter's endpoints API, read , and nothing here tested one.

Price position

342 priced models

Blended $0.036 per million tokens, 3 in to 1 out: cheaper than 335 of the 342 priced models in the catalog. Each bar is one stretch of the catalog on a logarithmic price scale; the mark is this listing.

Measured quality

145 models with an index · 132 on a design arena board

Each mark is one board this model stands on, placed at the share of that board it scores above, and led to its number in the list under the rule. The columns and the ledger below carry the figures.

Intelligence Indexnot scoredArtificial Analysis has not scored this model on its Intelligence Index.the intelligence leaderboard
Coding Index20.7Position 97 of 118, above 18% of the models scored.the coding leaderboard
Agentic Index1.2Position 77 of 94, above 17% of the models scored.the agentic leaderboard
  1. Data visualisation · Models933 Elowin rate 39.7% · source rank 127the board, 110 ranked
  2. Websites · Models843 Elowin rate 27.9% · source rank 149the board, 126 ranked
Epoch Capabilities Index · a fit by a third party over every benchmark this model has results onthe whole index, with intervals
Index134.9132.3 to 137.3
Position144 of 258above 44% of the models placed
Results behind it6benchmark results
As Epoch names itgpt-oss-20breleased 5 Aug 2025

Position 144 of the 258 models Epoch has placed, which is not the population of either board above, on a scale anchored at Claude 3.5 Sonnet 130 and GPT-5 150. Epoch states that absolute values are meaningless by themselves and that comparisons between models are what the index is for, that values are not linearly related to benchmark accuracy, and that a highly specialised model may score low while being capable in its own domain.

Benchmark results

1 of 6 boards · 6 results
Epoch AI Benchmarking Hub · publisher-run, CC-BY-4.02 of 22 tasks
TaskAll runsEffortScore
GPQA diamond342high0.46
medium0.61
low0.53
OTIS Mock AIME 2024-2025319high0.51
medium0.65
low0.40

Mean accuracy, 0 to 1, on the task the row names, as Epoch AI runs it and averaged over repeated runs. All eighteen of their tasks, including the held-out FrontierMath tiers, named as they name them. The model id carries the effort setting the run used. Epoch AI Benchmarking Hub names these runs gpt-oss-20b_high, gpt-oss-20b_medium and gpt-oss-20b_low, and the effort beside each score says which run it is. All runs is how many runs the publisher scored on the task, every model's. Read from epoch.ai ↗, which dates its own data . Epoch AI Benchmarking Hub data, by Epoch AI, under CC BY 4.0.

1 of the 6 openly licensed boards Vitna mirrors carries a result for gpt-oss-20b, each printed as its publisher published it and none added to another. MTEB, LiveBench, HELM, Vectara hallucination leaderboard and Aider polyglot have not published one under a name that matches its listing, and names are matched to the catalog exactly and never guessed. A task that is a link opens its column on /capability, with every run scored on it. Every board, with its licence and what its publisher measures, is on /benchmarks.

Sources

Source: OpenRouter (openrouter.ai/models), as of 17 Nov 16:29 UTC. sha256 0bb53a77713f over 853,270 bytes.
Catalog: OpenRouter's public model list.
Intelligence, coding and agentic indexes: Artificial Analysis, as published by OpenRouter.
Elo, win rate and rank: Design Arena, as published by OpenRouter.
Source: each model's own Hugging Face repository (huggingface.co), as of 20 Nov 2026. 162 repositories read, 4 did not answer.
Licence, parameter count, model class and gating: each model's own Hugging Face repository, the one OpenRouter names for the listing.
A licence is the tag the uploader put on the repository, read from nothing else and verified by nobody.
A parameter count is the total the repository's safetensors files declare; for a mixture of experts that is every expert, not the share one token runs through.
Source: the Epoch AI Capabilities Index (epoch.ai/eci), as of 20 Nov 2026. 258 models placed over 2917 benchmark results.
Epoch AI Benchmarking Hub data, by Epoch AI, under CC BY 4.0.
Vitna ran none of these evaluations and adjusted none of these figures.