GPT-OSS 120B in QueryWise: strong reasoning, but no leader status

GPT-OSS 120B’s main strength is reasoning

GPT-OSS 120B is an OpenAI model with a context window of 131K tokens. On paper, that is substantial capacity for long documents, conversations with many conditions, and multi-step tasks. But the number alone does not make every answer good. The model should be judged by how well it maintains logic, notices contradictions, and avoids jumping to the first plausible conclusion.

In the QueryWise leaderboard, its overall result is 5.2 out of 10, placing it 22. In reasoning, however, it scored 9.9 out of 10. That is its strongest argument: GPT-OSS 120B looks considerably more interesting on tasks that involve comparing conditions, tracing cause and effect, or checking someone else’s solution than as a general-purpose conversational partner.

We added it on release day and initially tested it with short logic scenarios. The impression was mixed: the model often sees a task’s structure faster than expected, but its overall ranking does not make it an undisputed favorite. Qwen3 Max is currently above it in the table, while Claude Haiku 4.5 is below. For reference, Claude Fable 5 leads with a result of 9.8.

What the leaderboard and our measurements show

QueryWise does not evaluate GPT-OSS 120B based on one impressive answer. Its overall score is 5.2 out of 10, with 5.1 for coding and 9.9 for reasoning. The gap between the two areas is telling: the model handles logical tasks much better than programming. That is not a failure, but it is not the profile of a top developer assistant either.

Median response speed is currently recorded as — seconds, while our telemetry does not yet contain enough data to assess reliability. We therefore will not present an early observation as a statistical fact. Some questions are solved quickly, but a few sessions are not enough to honestly judge service stability under load.

In the current snapshot, the model ranks 22. The top part of the leaderboard includes Claude Fable 5, GPT-5.6 Sol, Kimi K3, Grok 4.5; Claude Fable 5 is first with a score of 9.8. The gap to the top is 4.6 points. That distance matters more than the ordinal number itself: GPT-OSS 120B is still in the competition, but it is not yet a model you can use for every request without checking.

My telemetry-based conclusion is simple: this is a candidate for reasoning tasks, not a universal “do everything” button. Reliability data is still limited, and that constraint matters.

Which questions does it actually handle well?

GPT-OSS 120B performs best when a question can be broken down into conditions. For example: “Compare two mortgages in tenge with the same down payment and show which details need to be clarified.” This is not financial advice, but the model can carefully work through the assumptions and identify missing information.

Another strong scenario is everyday logic: “Our train from Almaty to Astana leaves in the evening, one passenger is late, and luggage cannot be checked in early. Which options do not violate the conditions?” Here, the useful quality is not encyclopedic knowledge but the ability to keep track of constraints. Another example: “Review my physics exam preparation plan if I only have an hour and a half on weekdays.” GPT-OSS 120B can analyze the schedule, find conflicts, and suggest an order of action.

The large 131K-token window is useful when a prompt contains a long conversation, contract, or set of notes. But the model does not eliminate the need to check facts. For medical, legal, and financial decisions, it is a tool for preliminary analysis, not a specialist.

Its coding result is 5.1 out of 10. So “fix a small Python script and explain the error” is a sensible request, while “design a production service with security and tests” requires human review. For such tasks, I would compare its answer with another model directly in QueryWise.

Where GPT-OSS 120B starts to fall short

The main weakness is visible in the balance of scores. With 9.9 for reasoning, the model received only 5.1 for coding. This means that a sound logical outline does not guarantee clean code, correct imports, or working edge cases. An answer may sound convincing while still requiring line-by-line execution and checking.

The model is also weaker when users expect an exact, up-to-date fact. The question “What are today’s rules for carrying a power bank on a flight from Almaty?” requires current information and a link to an official source. GPT-OSS 120B can help create a checklist of what to verify, but it should not be the final authority. The same applies to taxes, visas, medical treatment, and bank product terms.

There is also a practical caveat regarding stability. Median speed in our system is currently listed as — seconds, while reliability has not yet been assessed using a sufficient number of observations. It is therefore too early to promise consistent quality in every conversation. We have seen strong reasoning, but the 5.2 out of 10 rating shows that the average user result is noticeably more modest than the most impressive individual answers.

Finally, it should not be compared with the leader on just one parameter. Claude Fable 5 currently scores 9.8, while GPT-OSS 120B trails first place by 4.6 points. Qwen3 Max is higher in the table and Claude Haiku 4.5 is lower. That is a good reason to run several answers side by side rather than blindly arguing with a single model.

Pricing and availability in Kazakhstan

In QueryWise, one average question to GPT-OSS 120B costs approximately ≈10 ₸. That seems reasonable for testing ideas and comparing answers, especially when one prompt can be sent to several leading AI models at once. Still, costs should be monitored during long conversations or bulk generation.

The service is available in Kazakhstan: you can pay in tenge, and the interface works in Russian and Kazakh. After registration, three free questions are available. That is enough to test the model on your own material: give it a logic problem, a code fragment, and a long text with several conditions.

My testing method is straightforward: first, send the same prompt to GPT-OSS 120B and several competitors; then compare not the polished wording, but missed conditions, factual errors, and the usefulness of the next step. In QueryWise, the answers appear side by side, so differences become visible faster than when switching tabs manually. The leaderboard currently lists 25 models, with Claude Fable 5, GPT-5.6 Sol, Kimi K3, Grok 4.5 in the top four.

The verdict without the marketing haze: GPT-OSS 120B is good at unpacking complex conditions and explaining educational topics, but it does not look like the best choice for coding or tasks that require verified freshness. Trying it for ≈10 ₸ is reasonable. Leaving its answers unchecked is not.