OpenAI hasn’t launched an open-weight language mannequin since GPT-2 again in 2019. Six years later, they shocked everybody with two: gpt-oss-120b and the smaller gpt-oss-20b.
Naturally, we wished to know — how do they really carry out?
To seek out out, we ran each fashions by way of our open-source workflow optimization framework, syftr. It evaluates fashions throughout completely different configurations — quick vs. low-cost, excessive vs. low accuracy — and contains assist for OpenAI’s new “considering effort” setting.
In concept, extra considering ought to imply higher solutions. In apply? Not at all times.
We additionally use syftr to discover questions like “is LLM-as-a-Decide truly working?” and “what workflows carry out properly throughout many datasets?”.
Our first outcomes with GPT-OSS may shock you: the perfect performer wasn’t the most important mannequin or the deepest thinker.
As an alternative, the 20b mannequin with low considering effort constantly landed on the Pareto frontier, even rivaling the 120b medium configuration on benchmarks like FinanceBench, HotpotQA, and MultihopRAG. In the meantime, excessive considering effort not often mattered in any respect.
How we arrange our experiments
We didn’t simply pit GPT-OSS in opposition to itself. As an alternative, we wished to see the way it stacked up in opposition to different robust open-weight fashions. So we in contrast gpt-oss-20b and gpt-oss-120b with:
- qwen3-235b-a22b
- glm-4.5-air
- nemotron-super-49b
- qwen3-30b-a3b
- gemma3-27b-it
- phi-4-multimodal-instruct
To check OpenAI’s new “considering effort” function, we ran every GPT-OSS mannequin in three modes: low, medium, and excessive considering effort. That gave us six configurations in complete:
- gpt-oss-120b-low / -medium / -high
- gpt-oss-20b-low / -medium / -high
For analysis, we forged a large internet: 5 RAG and agent modes, 16 embedding fashions, and a spread of movement configuration choices. To guage mannequin responses, we used GPT-4o-mini and in contrast solutions in opposition to identified floor reality.
Lastly, we examined throughout 4 datasets:
- FinanceBench (monetary reasoning)
- HotpotQA (multi-hop QA)
- MultihopRAG (retrieval-augmented reasoning)
- PhantomWiki (artificial Q&A pairs)
We optimized workflows twice: as soon as for accuracy + latency, and as soon as for accuracy + price—capturing the tradeoffs that matter most in real-world deployments.
Optimizing for latency, price, and accuracy
Once we optimized the GPT-OSS fashions, we checked out two tradeoffs: accuracy vs. latency and accuracy vs. price. The outcomes had been extra stunning than we anticipated:
- GPT-OSS 20b (low considering effort):
Quick, cheap, and constantly correct. This setup appeared on the Pareto frontier repeatedly, making it the perfect default alternative for many non-scientific duties. In apply, meaning faster responses and decrease payments in comparison with greater considering efforts. - GPT-OSS 120b (medium considering effort):
Greatest suited to duties that demand deeper reasoning, like monetary benchmarks. Use this when accuracy on advanced issues issues greater than price. - GPT-OSS 120b (excessive considering effort):
Costly and normally pointless. Preserve it in your again pocket for edge instances the place different fashions fall brief. For our benchmarks, it didn’t add worth.
Studying the outcomes extra fastidiously
At first look, the outcomes look easy. However there’s an necessary nuance: an LLM’s high accuracy rating relies upon not simply on the mannequin itself, however on how the optimizer weighs it in opposition to different fashions within the combine. As an instance, let’s have a look at FinanceBench.
When optimizing for latency, all GPT-OSS fashions (besides excessive considering effort) landed with related Pareto-frontiers. On this case, the optimizer had little cause to focus on the 20b low considering configuration—its high accuracy was solely 51%.
When optimizing for price, the image shifts dramatically. The identical 20b low considering configuration jumps to 57% accuracy, whereas the 120b medium configuration truly drops 22%. Why? As a result of the 20b mannequin is much cheaper, so the optimizer shifts extra weight towards it.
The takeaway: Efficiency will depend on context. Optimizers will favor completely different fashions relying on whether or not you’re prioritizing pace, price, or accuracy. And given the massive search house of potential configurations, there could also be even higher setups past those we examined.
Discovering agentic workflows that work properly in your setup
The brand new GPT-OSS fashions carried out strongly in our exams — particularly the 20b with low considering effort, which regularly outpaced costlier rivals. The larger lesson? Extra mannequin and extra effort doesn’t at all times imply extra accuracy. Generally, paying extra simply will get you much less.
That is precisely why we constructed syftr and made it open-source. Each use case is completely different, and the perfect workflow for you will depend on the tradeoffs you care about most. Need decrease prices? Sooner responses? Most accuracy?
Run your individual experiments and discover the Pareto candy spot that balances these priorities on your setup.