gpt-oss-20b
OpenAIReleased Aug 5, 2025Open weights · Apache 2.0gpt-oss-20b
- Intelligence (Epoch)
- 138 · 87th of 111
- Reasoning
- yes, with effort levels
- Knowledge cutoff
- June 2024
- Context
- 131K tokens
- Max output
- 64k tokens
- Input
- text
- API features
- tools, structured outputs, JSON mode, seed, logprobs
Read Oct 23, 06:00 UTC from the hosts' published offers and OpenRouter, spec sheet from OpenAI · cutoff: platform.openai.com
Intelligence for the price
On the best-deal line: no model scores higher for less than $0.12 per M.
Epoch Capabilities Index against the lowest input + output price per M, log scale · prices Oct 23. Score: Epoch AI, CC BY 4.0, 138 with an interval of 132 to 140.
Against its rivals
The best of its price range and the top score
- gpt-oss-20bthis model$0.12138
- Claude Opus 5.5top score$24.00169
- DeepSeek V4 Flash 0731same price range$0.21155
- GLM-5.3-Flashsame price range$0.19153
Bars: Epoch Capabilities Index (Epoch AI, CC BY 4.0), drawn from 100, below the lowest score in the catalogue · price: lowest input + output per M.
Speed by provider
Output speed of gpt-oss-20b at each host, fastest first
- Groq389 tok/s0.57 s to first token
- CoreWeave195 tok/s0.18 s to first token
- DeepInfra110 tok/s0.27 s to first token
- AkashML104 tok/s0.45 s to first token
- Parasail101 tok/s0.50 s to first token
- SiliconFlow79 tok/s1.21 s to first token
- DekaLLM44 tok/s0.43 s to first token
Speed: OpenRouter, median of the last 30 min, read Oct 23, 06:00 UTC · a snapshot of real traffic before the daily read, not a daily average · bars from zero.
Run it yourself
open weights, Apache 2.0 · fits 6 of the 6 workstations in our guideSelf-host guide ›- Parameters
- 20.9B
- 3.6B active per token (MoE)
- Memory needed
- 13 GB
- in INT4 · 25 GB in INT8
- Smallest cloud setup
- RTX 3090
- 24 GB · $0.17/h, about $124 a month
- Worth it above
- 1B tokens
- a month, against the cheapest API at $0.12 per M
| Quantization | Memory | Smallest cloud GPU | Quality |
|---|---|---|---|
| INT4 (Q4) | 13 GB | RTX 3090 · $0.17/h | smallest footprint, minor quality loss |
| INT8 (Q8) | 25 GB | A100 · $1.19/h | near-lossless |
| FP16 (full) | 50 GB | A100 · $1.19/h | reference quality |
At home: in INT4 it fits a RTX 5090 (32 GB) + host PC (plus a host PC), about $3,500 once, or $0.1332/h over 3 years, electricity aside; it pays for itself after about 29B tokens against the cheapest API.
Memory ≈ parameters × bytes per parameter × 1.2 · cloud price: on-demand rate, running 24/7 · builds: GGUF, AWQ, MLX, GPTQ, Ollama · break-even calculator
| Provider | Input | Output | Cache ↑ | tok/s | First token | Uptime 24h | Checked |
|---|---|---|---|---|---|---|---|
| Parasailfp4 · 128K ctx | $0.03 | $0.15 | $0.02 | 101 | 0.50 s | 99.7% | Oct 23 |
| DekaLLMbf16 (mxfp4) · 128K ctx | $0.029 | $0.14 | $0.029 | 44 | 0.43 s | 99.9% | Oct 23 |
| CoreWeavefp4 · 128K ctx | $0.03 | $0.13 | $0.03 | 195 | 0.18 s | 100% | Oct 23 |
| Fireworks AI125K ctx | $0.07 | $0.30 | $0.035 | — | — | — | Sep 8 |
| Groqprecision not stated · 128K ctx · 64K out | $0.075 | $0.30 | $0.0375 | 389 | 0.57 s | 99.7% | Oct 23 |
| AkashMLfp4 · 128K ctx | $0.02 | $0.10 | — | 104 | 0.45 s | 100% | Oct 23 |
| DeepInfrabf16 (mxfp4) · 128K ctx | $0.03 | $0.14 | — | 110 | 0.27 s | 99.3% | Oct 23 |
| Novita AIprecision not stated · 125K ctx · 32K out | $0.04 | $0.15 | — | — | — | — | Oct 8 |
| Phalaprecision not read: host terms · 128K ctx | $0.04 | $0.15 | — | — | — | 99.7% | Sep 16 |
| SiliconFlowfp8 (mxfp4) · 128K ctx · 8K out | $0.04 | $0.18 | — | 79 | 1.21 s | 86.8% | Oct 23 |
| Amazon Bedrockprecision not read: host terms · 128K ctx | $0.07 | $0.15 | — | — | — | 100% | Oct 23 |
| Together AIprecision not stated · 125K ctx | $0.05 | $0.20 | — | — | — | — | Sep 22 |
| Googleprecision not stated · 128K ctx · 32K out | $0.07 | $0.25 | — | — | — | 31.6% | Oct 22 |
10 of 13 providers · prices per million tokens, read Oct 23, 06:00 UTC · speed: OpenRouter, median of the last 30 min, read Oct 23, 06:00 UTC · uptime: OpenRouter, last 24 h · under each host: weights precision, context and max output as served, the maker's value in brackets when the host's differs
| Provider | Weights | Context | Max output | Tools | Structured outputs | Reasoning |
|---|---|---|---|---|---|---|
| AkashML | fp4 | 128K | not stated by the host | yes | yes | yes |
| CoreWeave | fp4 | 128K | not checked yet | yes | yes | yes |
| DekaLLM | bf16 | 128K | not checked yet | yes | yes | yes |
| DeepInfra | bf16 | 128K | not stated by the host | yes | yes | yes |
| Parasail | fp4 | 128K | not read: host terms | yes | yes | yes |
| Novita AI | not stated by the host | 125K | 32K | — | — | — |
| Phala | not read: host terms | 128K | not read: host terms | — | — | — |
| SiliconFlow | fp8 | 128K | 8K | no | yes | yes |
| Amazon Bedrock | not read: host terms | 128K | not read: host terms | yes | no | yes |
| Together AI | not stated by the host | 125K | not stated by the host | — | — | — |
| not stated by the host | 128K | 32K | no | yes | yes | |
| Fireworks AI | not checked yet | 125K | not checked yet | — | — | — |
| Groq | not stated by the host | 128K | 64K | yes | yes | yes |
Weights, context and max output: read at each host's own catalogue or docs where its terms allow, otherwise as the host declares to OpenRouter; hover a cell for its source and date · tools, structured outputs, reasoning: as declared to OpenRouter · not verified by us
gpt-oss-20b spec sheet, field by field, with sources
gpt-oss-20b spec sheetmaker-declared
What OpenAI declares about the model itself — read from its own weights repository and documentation, not from any host. What each endpoint actually serves is in the pricing above: the two can differ, and both are true.
- Context window
- 131,072 tokens
- Parameters
- 20,914,757,184≈ 20.9B
- Active experts
- 4 of 32 active per token
- Weight precision
- mxfp4
- License
- Apache 2.0 — open weights, permissivecuratedScope: the licence on the published weights, and nothing else. It is not the maker's acceptable-use policy, which is a separate document, and it is not the contract you sign with whichever host you call — hosts set their own terms, which this index does not read.
- Knowledge cutoff
- June 2024maker's docs
- Modalities
- text
- Max output (official)
- to verify
- Release date
- to verify≈ August 5, 2025 third-party catalogue — not the maker's figure
- Maker lifecycle status
- to verify
Also declared in the weights
- Architecture class
- GptOssForCausalLM
- Layers
- 24
- Vocabulary
- 201,088 tokens
- Weights first published
- August 4, 2025
- Weights last updated
- August 26, 2025
- Paper
- arXiv:2508.10925
openai/gpt-oss-20b · config.json @ 6cee5e8 · read 2026-10-08 · 7 of 10 fields published by the maker
| Provider | Chatbot 1,000 in / 500 out | RAG / long context 8,000 in / 500 out | Batch summarize 4,000 in / 1,000 out |
|---|---|---|---|
| AkashML | $0.07 | $0.21 | $0.18 |
| CoreWeave | $0.10 | $0.30 | $0.25 |
| DekaLLM | $0.10 | $0.30 | $0.26 |
| DeepInfra | $0.10 | $0.31 | $0.26 |
| Parasail | $0.11 | $0.32 | $0.27 |
| Novita AI | $0.12 | $0.40 | $0.31 |
| Phala | $0.12 | $0.40 | $0.31 |
| SiliconFlow | $0.13 | $0.41 | $0.34 |
| Amazon Bedrock | $0.14 | $0.64 | $0.43 |
| Together AI | $0.15 | $0.50 | $0.40 |
| $0.20 | $0.68 | $0.53 | |
| Fireworks AI | $0.22 | $0.71 | $0.58 |
| Groq | $0.23 | $0.75 | $0.60 |
What each host bills for 1,000 requests of a representative job
Your own workload
Other OpenAI models · all its models, hosts and history ›
- gpt-oss-120b19 providers · from $0.18
- GPT-5.6 Luna3 providers · from $1.40
- GPT-5.6 Sol3 providers · from $24.00
- GPT-5.6 Terra3 providers · from $14.00
- GPT-4.11 provider · from $10.00
- GPT-4.1 mini1 provider · from $2.00