gpt-oss-20b is a compact open-weight reasoning model with official documentation and repository support for local deployment, coding, tool use, and developer experimentation. Its 131k context window, reported 200.6 output tokens/s, and $0.06/$0.20 per-million-token hosted pricing in Artificial Analysis support unusually strong cost and speed scores. The supplied evidence does not establish vision or broad multimodal input, so multimodal capability remains at the calibration floor. Coding evidence is mixed. OpenAI reports SWE-bench Verified results rising from 37.4% at low to 60.7% at high reasoning, plus 2,230 Codeforces Elo without tools. However, Terminal-Bench 2.0 reports 3% ±1 across 74 tasks with Mini-SWE-Agent and Terminus 2, and DeepSWE and LiveCodeBench publish no result. This supports capability above Gemma 3's coding anchor (42), but below DeepSeek-V3.2 (84) and Kimi K2.5 (86), whose catalog scores reflect stronger agentic evidence. Reasoning and public-preference evidence is modest: Arena lists rank 231 overall and the cited third-party Arena-Hard V2 page reports 48.6%, well below frontier anchors such as Claude Opus 4.8 and GPT-5.5. Open weights, official setup materials, and low serving cost make developer experience strong. Pricing is relatively clear through the cited external rate card, but provider-controlled terms, deployment availability, and license details should be verified for a chosen host.
