GPT-OSS launched August 5, 2025 - OpenAI's first open-weight model family since GPT-2 in 2019. Six years is a long gap. The question worth asking is not whether the release is significant (it is), but whether the architecture, the benchmarks, and the pricing math add up to something worth building on - or whether the hype is running ahead of the reality.
Here is what the primary sources say.
What GPT-OSS actually is, architecturally
Both models are mixture-of-experts reasoning models with a 128K context window and configurable reasoning effort. The larger gpt-oss-120b carries 117B total parameters but only 5.1B active per forward pass, landing near o4-mini on core reasoning benchmarks while fitting on a single 80GB GPU. The smaller gpt-oss-20b has 21B total parameters with 3.6B active and runs on edge devices with about 16GB of memory.
The MoE design is the thing that makes the economics work. During inference, the MoE architecture means GPT-OSS 120B computes roughly the same amount of work per token as a 20B dense model - the expert routing selects a subset of specialist layers for each token.
The MoE weights are responsible for over 90% of the total parameter count, and quantizing these to MXFP4 enables the larger model to fit on a single 80GB GPU and the smaller model to run on systems with as little as 16GB of memory.
Both models were released under the permissive Apache 2.0 license. That matters. Apache 2.0 terms mean a downstream organization can download, fine-tune, and redistribute the model commercially without seeking Mistral's - or in this case OpenAI's - permission, triggering a legal review, or running into usage caps based on user scale.
The benchmark picture: strong, not frontier
OpenAI's own model card finds gpt-oss-120b and gpt-oss-20b generally perform on par with o4-mini. Independent analysis is more nuanced. A peer-reviewed evaluation tested both variants against six contemporary open-source models ranging from 14.7B to 235B parameters across ten benchmarks covering general knowledge, mathematical reasoning, code generation, multilingual understanding, and conversational ability. GPT-OSS models showed middle-tier performance with particular strength in code generation but weakness in multilingual tasks.
The larger gpt-oss-120b does not come in above DeepSeek R1 0528's score of 59 or Qwen3 235B's score of 64 on the Artificial Analysis Intelligence Index - but it is notably smaller in both total and active parameters than both of those models. That is the honest framing: not the best open-weight model on raw benchmarks, but probably the most capable model that fits on a single H100 while running under a genuinely permissive Western license.
The cost math, computed from primary sources
Artificial Analysis tracks median API pricing of $0.15/$0.69 per million input/output tokens for the 120B and $0.08/$0.35 for the 20B. Those prices put the 120B close to 10x cheaper than OpenAI's proprietary APIs for o4-mini ($1.1/$4.4) and o3 ($2/$8).
Put that in workload terms: a team running 50 million output tokens per month - roughly 200,000 Slack-length responses - pays about $34,500/year at o4-mini's API rate. The same volume on gpt-oss-120b at median third-party pricing costs around $4,140/year. That is a $30K annual gap before touching self-hosting.
Prices across providers vary up to 7.8x - the most affordable is CoreWeave at $0.05 per million tokens, while Cloudflare charges $0.39 per million tokens.
For output speed, the top providers reach Cerebras at 1,677.5 tokens/second and SambaNova at 707.4 tokens/second.
Self-hosting changes the math further. Even at 10M tokens per month, an A100 deployment costs roughly one-fifth of the API equivalent. The calculation tilts further toward self-hosting as volume grows, because the GPU cost is fixed per hour regardless of how many tokens you generate.
| gpt-oss-120b | gpt-oss-20b | o4-mini (API) | DeepSeek R1 | |
|---|---|---|---|---|
| Total parameters | 117B | 21B | closed | ~671B |
| Active per token | 5.1B | 3.6B | closed | ~37B |
| Context window | 128K | 128K | 128K | 128K |
| License | Apache 2.0 | Apache 2.0 | proprietary | MIT |
| Median API input price | $0.15/M | $0.08/M | $1.10/M | ~$0.55/M |
| Single-GPU fit | H100 80GB | 16GB RAM | n/a | No |
| Multilingual | Weak | Weak | Strong | Strong |
Where it fits in a real workflow
These models are designed to run on your own infrastructure, offer full chain-of-thought visibility and tool use, and provide near-parity performance with OpenAI's proprietary models. In practice that means a few specific use cases where gpt-oss makes more sense than the alternatives:
Agentic pipelines that generate a lot of output tokens. The MoE cost structure rewards high-throughput workflows - document drafting, code review loops, summarization queues - where output volume is predictable.
Data-sensitive environments. Open models create new opportunities for institutions with strict data-residency or security requirements. Where data cannot leave a country and using a third-party cloud service is not possible, open-weight models can provide a secure and flexible way to harness advanced AI while keeping sensitive information under local control.
Fine-tuning for a specific domain. Multiple fine-tuning tutorials and frameworks now exist - from Unsloth's consumer-friendly LoRA guides to AWS's enterprise-scale SageMaker workflows - and researchers have noted that domain-specific fine-tuned models will become increasingly common as the tools to create them are freely available.
A teammate like Beagle, running inside Slack, could route classification-grade requests to gpt-oss-20b at $0.08/M and escalate longer reasoning tasks to the 120B variant - letting the draft-and-approve model stay cost-efficient even on high-volume channels.
Beagle in action#legal-review, 2:07pmWhat is genuinely new versus incremental
The architecture is not new - MoE has been the dominant open-weight design since Mixtral in late 2023. MXFP4 quantization is well-established. What is new is that OpenAI has released its first open-weight large language models since 2019's GPT-2. The significance is more institutional than technical.
Three things are worth tracking honestly:
What is real: A 117B-parameter reasoning model that fits on one H100, costs a tenth of o4-mini via API, ships under Apache 2.0, and benchmarks near o4-mini on English reasoning and code. That is a legitimate step change for teams that previously had to choose between performance and self-hosting cost.
What is incremental: The benchmark position. Competitors continue to push the frontier: Alibaba's Qwen3-Coder-Next and DeepSeek V3.2 demonstrate that the open-model ecosystem is accelerating rapidly, ensuring GPT-OSS will face ongoing competitive pressure. At release, gpt-oss-120b already sat behind Qwen3-235B and DeepSeek R1 on the intelligence index.
What to watch: Following GPT-OSS, OpenAI has already expanded the open-weight family with gpt-oss-safeguard models released in October 2025, and the community has rapidly built on the base models. The companion safety models - which use chain-of-thought reasoning to interpret developer-defined safety policies at inference time - are the more interesting long-term development for enterprise deployments. Over the past 90 days, the cheapest input price has fallen 26%, from $0.050 to $0.037 per million tokens
- the commodity pressure is already visible.
GPT-OSS open weight model: common questions
What is gpt-oss and how is it different from GPT-4?
GPT-OSS is OpenAI's family of open-weight models - gpt-oss-120b and gpt-oss-20b - released August 5, 2025 under the Apache 2.0 license. They are OpenAI's first open-weight large language models since GPT-2 in 2019. Unlike the proprietary GPT-5 line, anyone can download the weights, run them locally, and use them commercially without restriction. GPT-4 remains closed and accessible only via API.
Can gpt-oss-120b run on a single GPU?
OpenAI released both models in MXFP4 precision: gpt-oss-120b comes in at 60.8GB and gpt-oss-20b at 12.8GB. The 120B can be run in its native precision on a single NVIDIA H100, and the 20B can be run easily on a consumer GPU or laptop with more than 16GB of RAM.
How does gpt-oss pricing compare to OpenAI's own API?
Median pricing across API providers sits at $0.15/$0.69 per million input/output tokens for gpt-oss-120b and $0.08/$0.35 for the 20B. That puts the 120B close to 10x cheaper than OpenAI's o4-mini ($1.1/$4.4) and o3 ($2/$8) via the proprietary API. The cheapest providers push even lower - around $0.04/M input on DeepInfra.
Is gpt-oss the best open-weight model available?
No. On the Artificial Analysis Intelligence Index, gpt-oss-120b scores 58, coming behind DeepSeek R1 and Qwen3-235B. It is the most capable model that fits on a single H100 under an Apache 2.0 license from a US lab - a narrower but practically useful distinction.
What is gpt-oss weak at?
Independent evaluation found GPT-OSS models show middle-tier performance, with particular strength in code generation but weakness in multilingual tasks. The model is also text-only - no image or audio input - which matters for workflows that previously relied on GPT-4V or similar multimodal models.