What is gpt-oss-120b?
gpt-oss-120b is OpenAI’s open weights under apache 2.0 model for self-hosted reasoning, tool use and production workloads that fit one 80GB GPU. It was released August 5, 2025. This guide separates documented specifications from vendor benchmark claims and gives teams a practical way to decide whether the model belongs in a real evaluation.
QUICK VERDICT: gpt-oss-120b is OpenAI’s practical open-weight option. Its value depends on whether a team can operate the serving, safety and evaluation stack responsibly.
Field | Verified value |
|---|---|
Provider | OpenAI |
Release date | August 5, 2025 |
Availability | Available; downloadable weights |
License | Open weights under Apache 2.0 |
Context window | 131,072 tokens |
Maximum output | Shares the model context budget |
Modalities | Text input; text output |
API pricing | No OpenAI token fee for self-hosting; infrastructure and hosted-provider rates vary |
Access | Hugging Face, compatible inference stacks and third-party hosts |
Best fit | self-hosted reasoning, tool use and production workloads that fit one 80GB GPU |
Last verified | September 2, 2026 |
What changed with gpt-oss-120b
gpt-oss-120b should not be evaluated as a name change. The important differences are the model’s reasoning controls, tool behavior, context policy, access route and economic profile. Those details determine whether a benchmark result transfers to production.
Area | What changed or matters |
|---|---|
License | Apache 2.0 open weights |
Architecture | 117B total parameters with 5.1B active |
Hardware | Designed to fit on one 80GB GPU |
Reasoning | Adjustable reasoning and tool use |
The correct comparison baseline is Earlier open model alternatives. Teams planning a new deployment should also include Hosted GPT-5 models for managed access where access allows. Testing only the newest model or only the incumbent hides the migration cost, output-style changes and tool-use regressions that often matter more than a small benchmark gap.
gpt-oss-120b capabilities
The clearest fit is self-hosted reasoning, tool use and production workloads that fit one 80GB GPU. That does not mean every task in those categories should use the model. A production system combines the base model with prompts, retrieval, tools, permissions, memory, retry logic and human review. The model card describes only one layer of that system.
Capability | Practical implication |
|---|---|
Long context | 131,072 tokens. Test retrieval accuracy at realistic lengths rather than assuming every token receives equal attention. |
Output capacity | Shares the model context budget. Long output is useful only when the verification process can keep up. |
Modalities | Text input; text output. Confirm format-specific accuracy with your own files. |
Agent use | Use explicit tool schemas, narrow permissions, approval gates and recoverable operations. |
Reasoning | Record the exact effort level because cost, latency and quality can change materially. |
For coding, evaluate repository navigation, test creation, regression rate, review burden and recovery after a failed tool call. For research, measure citation correctness, source coverage and whether the model distinguishes evidence from inference. For computer or browser use, record every unsafe click, wrong field, lost state and unapproved action.
Benchmarks and evidence
The evidence below comes from OpenAI or named launch partners. It is useful for identifying intended strengths, but it is not equivalent to an independent head-to-head test. Prompts, tools, reasoning budgets, sampling, infrastructure and scoring rules can differ.
Evidence | Reported result | How to interpret it |
|---|---|---|
Efficiency | OpenAI says the model fits a single H100 or MI300X | Deployment specification |
Reasoning | Reported to outperform similarly sized open models | Vendor benchmark comparison |
Tools | Supports structured tool-use patterns | Model-card capability |
Treat benchmark results as a shortlist signal. Before purchasing or migrating, reproduce representative work under one frozen protocol. Keep model snapshots, reasoning effort, tool access and token budgets constant. Report repeated-run pass rates and total cost per accepted result, not a single best attempt.
Pricing, access and deployment
Hugging Face, compatible inference stacks and third-party hosts. The current pricing reference is No OpenAI token fee for self-hosting; infrastructure and hosted-provider rates vary. Provider pricing changes frequently and can include cache rates, batch discounts, regional premiums, long-context multipliers, priority processing and tool-call fees. Recheck the official pricing page before budgeting.
Cost driver | What to measure |
|---|---|
Input tokens | Prompt, retrieved context, tool results and repeated history |
Output tokens | Visible answer plus any billable reasoning or generated artifacts |
Caching | Eligible repeated prefixes, cache-read price and expiration policy |
Tools | Search, computer use, code execution and third-party API fees |
Retries | Failed runs, verifier loops and human rework |
Success-adjusted cost | Total spend divided by deliverables that pass review |
A cheaper token price can lose to a more expensive model if it takes more steps, retries more often or produces work that needs heavy correction. Conversely, a frontier model can be wasteful when a smaller model already passes the task rubric. Route by measured task difficulty rather than brand prestige.
gpt-oss-120b limitations
- Text-only model.
- Self-hosting shifts safety and uptime to the operator.
- 131K context is smaller than current frontier models.
- Quantization can change quality.
- Hosted implementations may differ.
High-stakes medical, legal, financial, security and scientific work requires qualified review. Store the exact model identifier, prompt version, tools, source documents and approvals for each consequential run. Build a rollback path before granting write access to repositories, browsers, databases or cloud infrastructure.
Risk | Minimum control |
|---|---|
Hallucination | Require source checks or executable tests |
Prompt injection | Separate untrusted content from instructions and restrict tools |
Over-permission | Use least privilege and approval gates |
Silent model change | Pin snapshots where possible and run regression tests |
Data exposure | Review retention, regional processing and provider terms |
Runaway cost | Set token, time, tool-call and retry budgets |
How to evaluate gpt-oss-120b
Create 20 to 50 tasks from real work. Freeze the tasks and rubric before testing. Include easy tasks, normal tasks, edge cases and adversarial inputs. Give each model equivalent tools and enough budget to finish, but cap time and retries. Repeat non-deterministic runs so one lucky result does not decide the winner.
Test area | Record |
|---|---|
Task completion | Pass or fail plus rubric score |
Reliability | Repeated-run success and variance |
Quality | Factuality, instruction adherence and usefulness |
Tool use | Wrong calls, retries, recovery and permission errors |
Efficiency | Wall time, tokens, cache use, tool fees and human review |
Safety | Unsafe actions, injection response and sensitive-data handling |
Migration | Prompt changes, integration work and output-style regressions |
Compare gpt-oss-120b with Earlier open model alternatives and Hosted GPT-5 models for managed access. Choose the least expensive configuration that meets the acceptance threshold with an adequate safety margin. Re-run the suite after any model snapshot, system prompt, retrieval or tool change.
Frequently asked questions
Is gpt-oss-120b available now?
Available; downloadable weights. The documented access routes are Hugging Face, compatible inference stacks and third-party hosts. Availability can differ by region, plan and partner platform.
Is gpt-oss-120b open source?
gpt-oss-120b has an open-weight route. Open weights do not necessarily mean the training data, training code and full software stack are open source. Review the exact license before commercial deployment.
How much does gpt-oss-120b cost?
No OpenAI token fee for self-hosting; infrastructure and hosted-provider rates vary. Budget with measured end-to-end workloads because token rates alone omit retries, tools, caching, long-context multipliers and review time.
What is gpt-oss-120b best used for?
Its strongest documented fit is self-hosted reasoning, tool use and production workloads that fit one 80GB GPU. Start with a supervised pilot and keep human sign-off for consequential work.
Should I migrate from Earlier open model alternatives?
Only after a side-by-side evaluation. Measure pass rate, total cost, latency, output style, tool reliability and the engineering work required to migrate. A newer model is not automatically the better operational choice.
Related model guides
Start with the broader Best AI Models directory.
Compare with the GPT-5.6 Sol guide.
Official sources and update policy
Checked September 2, 2026. We update this guide when the provider changes the model specification, pricing, access, licensing or safety documentation. Vendor benchmarks are attributed and are not presented as independent testing. No vendor payment or affiliate relationship determined inclusion.