- 🧠 OpenAI open-weight 117B Mixture-of-Experts, activating 5.1B parameters per token
- 🔒 Runs in a Trusted Execution Environment with hardware attestation evidence
- 🔧 Configurable reasoning effort (low, medium, high) and native tool use
- 📏 128K-token context window for long inputs
- 🆕 Permissive Apache 2.0 license for commercial deployment
- 👁️ Full chain-of-thought access for debugging and transparency
- ⚡ Designed to run efficiently on a single 80GB GPU
- 🏢 Released March 2026 as a privacy-focused confidential-compute deployment
OpenAI is an American artificial intelligence research organization headquartered in San Francisco, structured as both a for-profit public benefit corporation and a nonprofit foundation. The lab developed the GPT family of large language models, the DALL-E image generation…
Explore 22 more models by OpenAI →GPT OSS 120B is the larger member of OpenAI's open-weight gpt-oss series, a Transformer using mixture-of-experts to keep only 5.1B of its 117B total parameters active per token. This catalog entry packages that model inside a Trusted Execution Environment (TEE), adding hardware attestation evidence so users can independently verify that inference runs in a confidential, tamper-resistant enclave. It carries forward the base model's permissive Apache 2.0 license, configurable reasoning depth, full chain-of-thought visibility, and native tool use including function calling, browsing, and structured output.
Within this confidential-compute family, the model is the higher-capacity counterpart to [[sibling:e2ee-gpt-oss-20b-p|GPT OSS 20B]]. The two share the same architecture and TEE wrapper, but the 120B variant activates 5.1B parameters per token versus the 20B model's 3.6B, and OpenAI positions the larger model for production, general-purpose, high-reasoning workloads while the smaller one targets lower-latency or memory-constrained deployment.
Compared to the standard non-enclave release, [[sibling:openai-gpt-oss-120b|OpenAI GPT OSS 120B]], the weights and capabilities are identical; the distinguishing feature here is end-to-end encryption and verifiable execution rather than any change to the model itself.
On capability, OpenAI reports that gpt-oss-120b reaches near-parity with its o4-mini model on core reasoning benchmarks while running efficiently on a single 80GB GPU. For broader chat, image, and embedding needs, the provider's same-period lineup also includes models such as [[sibling:openai-gpt-55|GPT-5.5]].
This About section is AI-generated from public sources via VeniceStats + Venice inference, with no human editing. It may contain inaccuracies.
| Seller | Reputation↓ | Routing | Input $/M | Cached $/M | Output $/M | Categories | API |
|---|---|---|---|---|---|---|---|
| surplusintelligence.ai 0x0e49…8927 | 68 | #1 | $0.039 | $0.039 | $0.195 | anon,chat,cheap,e2ee,frontier,privacy,reasoning,research,tasks,tee,web-search | openai-chat-completions |
| Fire Ant 🔥🐜 0xbe05…bc5d | 37 | gated | $0.1107 | $0.1107 | $0.5533 | chat,coding,free,json,math,tools | — |
| antseed-neon-puma-944e 0x6650…944e | 3 | gated | $0.0443 | $0.0443 | $0.2217 | chat,math | openai-chat-completions |
| antseed-opal-badger-2580 0xc85d…2580 | 2 | gated | $0.065 | $0.065 | $0.325 | chat | openai-chat-completions |
| Leftermute 0x388b…5389 | 0 | gated | $0.0118 | $0.0118 | $0.0591 | chat,coding,json,tools | openai-chat-completions |
"Best price" and the seller table are live AntSeed catalog data (advertised $/1M tokens — or $ per generated image for unit-billed image models — not settled amounts). Reputation = on-chain trust (0-100). "Routing" = the SDK's default buyer routing (what the VPR desktop app ships with): a trust ≥ 60 gate on the effective reputation, then cheapest-first among routable sellers; live failover state (per-peer cooldowns) is buyer-side runtime and not included. Model knowledge (TLDR, provider, About) via the VeniceStats enrichment layer. Advertised catalog, not the model used in any specific purchase. "Usage on AntSeed" counts only settlements whose buyers share the per-model split on-chain (metadata v2/v3, opt-in), so every usage figure is a lower bound.