gpt-oss-20b

# Fix list: gpt-oss-20b

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/gpt-oss-20b, the October 2026 research run, assessed 21 October 2026. Grade B, 69.9 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on gpt-oss-20b: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Capability, 49 out of 100, up to 17.9 more on the total

Why it scored 49: From 18 independent benchmark results in 5 areas, each scored as its share of the way from the benchmark's floor to the best result any model has on the same leaderboard. By area, Instruction following 79, Coding 77, Reasoning and knowledge 55, Long context 39, Tool calling and agent tasks 19. Publisher-submitted results count at half weight, and numbers that appear only on the publisher's card don't count.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-capability):

How well the model does on the benchmarks that matter to an agent. We don't average a publisher's own numbers into a grade. A result counts in full when we ran it on our reference machine, or when an independent party ran it and published the method (the Berkeley Function Calling Leaderboard, Artificial Analysis, Epoch AI's Benchmarking Hub, the SWE-bench and Terminal-Bench leaderboards, the Open ASR Leaderboard and others that run models themselves). A result the publisher submitted to such a leaderboard counts at half weight. A number that appears only on the publisher's card is shown on the listing as the publisher's claim and doesn't count.

Each result is scored as its share of the way from the benchmark's floor (chance for multiple choice, zero for a pass rate) to the best result any model, open or closed, has on the same leaderboard and setting on the day we read it. A model that matches the best model anywhere scores 100 on that benchmark. For an error rate such as word error rate, lower is better and the scale runs the other way, and where the leaderboard gives no floor the rate is read against the best model's, so twice the best model's error rate scores 50. Speed as a multiple of real time (RTFx) runs from about 1 to over 20,000, so it's read on a log scale from real time to the fastest model. On the MTEB leaderboard, where most models score between 0.5 and 0.75, the floor is the median of every model the same tab lists, and its results count at half weight because MTEB doesn't say who ran them and most are submitted with the model. An arena rating (Elo) isn't a ratio, so its floor is 400 points below the best model on that arena, the gap at which the better model wins about ten games in eleven. Where a model has results at several settings (reasoning effort, say), its best one counts, because that's the setting an agent would run.

The results group into areas, each weighted within the category.

- Tool calling and agent tasks, 35 (BFCL, τ²-bench, Terminal-Bench and agent suites like them)
- Coding, 20 (SWE-bench Verified, LiveCodeBench, Aider Polyglot)
- Reasoning and knowledge, 20 (GPQA Diamond, MMLU-Pro, Humanity's Last Exam, composite indexes that publish their components)
- Instruction following, 15 (IFEval, IFBench)
- Long context, 10 (RULER, LongBench and long-context retrieval tests)

A speech model is read on its own areas, recognition accuracy (word error rate on the Open ASR Leaderboard's sets, 60) and speed (real-time factor, 40) for speech to text, and intelligibility and speaker similarity (Seed-TTS-eval, the Open TTS Leaderboard, 70) and speed (30) for text to speech. A publisher's own converted copy of the same weights (Qwen's -hf builds, say) counts as the same checkpoint. An embedding model is read on retrieval (60) and the other MTEB task types, classification, clustering, reranking and similarity among them (40), from the MTEB leaderboard. A vision model adds an image area (MMMU, ChartQA, DocVQA or OCRBench) at 20, and the others shrink to fit.

Capability is the weighted mean of the areas with results, and needs results in at least two areas (one for a speech or embedding model). With less than that it's pending for that listing, its weight is shared across the other four, and the total is capped at 61, so a model nobody has benchmarked independently can't reach B, and every model with results ranked B or better stays above it. The listing says so. Until 21 October 2026 the cap was 69, which put models nobody had benchmarked above ones that had been.

## 2. Documentation and upkeep, 42 out of 100, up to 5.8 more on the total

Why it scored 42: The model card earns 17, covering six of the seven points and missing a statement of supported languages. Training data earns 10, a general description with a June 2024 cut-off and no sources or mix. Last change earns 5, because the files haven't changed since August 2025 and the reference code and harmony parser last changed in early November 2025, while this year's commits to the gpt-oss repository only touch the browser tool, CI and a version number. Responsiveness earns 0, with no OpenAI reply to any of the 40 most recent discussions and none at all since August 2025. Version policy earns 5 for a first release kept in its own repository. File health earns 5, since template reports and an install report wait without a reply.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-upkeep):

- 0 to 20, the model card. Intended use, limitations, architecture, parameters, context length, languages and knowledge cut-off. 20 for all seven, 3 off for each that's missing.
- 0 to 20, training data. 20 for the data published or listed with sources and sizes, 15 for a data card that names the sources or the mix, the cut-off and the filtering, 10 for a general description with a cut-off, 5 without one, 0 for none.
- 0 to 20, the last change to the files, the template or the reference code. 20 within 30 days, 15 within 90, 10 within 180, 5 within a year, 0 older.
- 0 to 20, responsiveness. The publisher's replies to the 40 most recent questions and reports in the repository's discussions or issues. 20 for most, 10 for some, 5 for a few, 0 for none.
- 0 to 10, version policy. 10 for earlier versions kept in their own repositories with a dated note when one is replaced, 5 for earlier versions kept without notes or a first release, 0 when files were replaced in place without a note.
- 0 to 10, file health. 10 when config, tokenizer, template and card agree and no report of a broken file waits without a reply, 5 when one does, 0 when the files contradict the card.

## 3. Risk, 85 out of 100, up to 3 more on the total

Why it scored 85: File safety earns 25, since the weights are safetensors and metal/model.bin is OpenAI's own raw Metal format rather than pickle, and nothing in the repository runs as code (a judgement call on the .bin name). Provenance earns 15, files in OpenAI's own organisation and a checksum on every file. Safety evaluation earns 10, because the published safety results for this size are OpenAI's own, and the outside review covered gpt-oss-120b. Agent guidance earns 15 for an instruction hierarchy the model was trained to follow, with prompt-injection results, though OpenAI reports it trails o4-mini there. Licence risk keeps 15, with nothing to deduct under Apache 2.0. The security programme earns 5 for OpenAI's security.txt with a Bugcrowd contact. Public advisories for the gpt-oss code couldn't be checked, so that half scores as absent.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-risk):

What could go wrong when you run it, and what the publisher has done about it. A higher score means less risk.

- 0 to 25, file safety. 25 for weights in safetensors or GGUF only, 15 when pickle files (.bin, .pt, .pth) sit beside them, 0 for pickle only. 10 off when the documented way to load it or to build its prompts runs code from the repository (`trust_remote_code`, or the publisher's scripts in place of a template).
- 0 to 15, provenance. 10 when the files come from the publisher's own organisation, the one its site or docs link to, and 5 for a checksum on every file.
- 0 to 20, safety evaluation. 20 when results are published and someone other than the publisher ran them or can rerun them, 10 for the publisher's own results, 0 for none.
- 0 to 15, guidance for agent use. 15 for documented prompt-injection guidance or an instruction hierarchy the model was trained to follow, 10 for a statement of what its trained refusals cover and the safeguards to add around it, 5 for general cautions, 0 for none.
- 0 to 15, licence risk. 15 to start, 5 off for each of these, a use policy or field-of-use limits beyond the law, terms the licensor can change or revoke, a revenue or user threshold, and a licence that changed without a dated note or whose link is broken. Not below 0.
- 0 to 10, a security programme. 5 for a way to report a flaw (security.txt, a disclosure policy or a security contact), 5 for advisories handled in public.

When our reference-machine runs start, measured robustness (prompt injection with AgentDojo, over-refusal with XSTest) joins Risk.

## 4. Reach, 90 out of 100, up to 2.5 more on the total

Why it scored 90: Memory earns 25, since OpenAI's own MXFP4 build is 13.76 GB and a fifth on top is 16.5 GB, just past the 16 GB band, and Ollama's 14 GB build lands in the same 24 GB band. Runtimes earn the full 25 from model code in released Transformers, vLLM and SGLang, the architecture in llama.cpp and the model in Ollama's library. Builds earn 10, because OpenAI ships one quantised precision, MXFP4, and the original and metal folders are the same weights for its reference code. Terms earn 15 for Apache 2.0, whose usage policy only asks users to follow the law. Download earns 10, since the repository isn't gated. Hosted earns 5, with five inference providers live for it on its Hugging Face page, while OpenAI's own API lists it with no usable rate limit and isn't counted.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reach):

Whether an agent can run it, and where.

- 0 to 30, memory. The smallest build from the publisher or a runtime's own library that keeps 4-bit precision or better, at its file size plus a fifth for the runtime and a working context. 30 for 16 GB or less, 25 for 24 GB, 20 for 32 GB, 15 for 64 GB, 10 for what fits our reference machine (the GPU share of 128 GB of unified memory), 5 for 256 GB, 0 above.
- 0 to 25, runtimes. 5 for each of llama.cpp, Ollama, vLLM, SGLang, LM Studio and Hugging Face Transformers that runs the architecture in a released version, from the runtime's own code, library or documentation. A publisher's fork or an unreleased branch doesn't count. 25 at most.
- 0 to 15, builds from the publisher. 15 for quantised builds at two precisions or more (FP8, 4-bit, GGUF), 10 for one, 0 when only other people's quantisations exist.
- 0 to 15, terms of use. 15 for an OSI-approved licence with nothing added, 10 for a custom licence that lets anyone use it commercially for free below a threshold or under a use policy, 5 for research or non-commercial use only, 0 for no licence or unclear terms.
- 0 to 10, download. 10 when the files download with no account and no terms to accept, 5 for a click-through that needs an account, 0 when a person approves each request.
- 5, hosted too. The publisher's own API serves it, or at least two inference providers do, so an agent without the hardware can still call it.

When our reference-machine runs start, speed on that machine (tokens per second and time to first token at the build that fits) joins Reach.

## 5. Agent interface, 90 out of 100, up to 1 more on the total

Why it scored 90: The chat template earns 20, a Jinja file in the repository. The tool-calling format earns 25, since OpenAI's harmony guide says how functions are declared and how calls come back, and OpenAI ships the openai-harmony parser. Trained tool calling earns 10, because each call ends the message and waits for its result, so it makes one call at a time. Structured output earns 15 for a JSON Schema given in the developer message, documented in the harmony guide. Reasoning control earns 10, with low, medium or high effort per request and the reasoning on its own channel, though it can't be switched off. Context earns 10 for 131,072 tokens, which OpenAI describes as trained in even though the config reaches it with YaRN.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-interface):

What a model and a developer need to talk to it. Speech, image and other models with no chat interface read the tool-calling and structured-output lines for what the model takes and returns (voices, timestamps, streaming, typed outputs), and the note says how.

- 0 to 20, the chat template. 20 for a template in the repository (Jinja, in the tokenizer config or its own file), 10 when the prompt format is only given as code you run.
- 0 to 25, the tool-calling format. 25 when the publisher documents how tools are declared and how calls come back, with the parser to use in a named runtime, 10 when only the template or a runtime's own recipe says it.
- 0 to 20, tool calling the model was trained for. 20 with several calls in one turn, 10 for one call at a time.
- 0 to 15, structured output. 15 when JSON or JSON Schema output is documented for the model with how to ask for it, 5 when the card claims it without saying how.
- 10, reasoning control. Thinking that can be turned off or budgeted per request, with where the reasoning text comes back.
- 0 to 10, context. 10 for 128K tokens or more trained, 5 for 32K to 127K. Context reached only by scaling the position encoding counts one band down.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- unchecked, public security advisories for github.com/openai/gpt-oss and github.com/openai/harmony, since GitHub's site and API answered 403 to this session; scored as absent
- unchecked, GitHub issues on openai/gpt-oss and openai/harmony, so responsiveness rests on the Hugging Face discussions only
- OpenAI's own API lists gpt-oss-20b with a rate limit of 0 on every tier and no price, so it isn't counted as hosted by the publisher; hosted rests on five inference providers instead
- File safety is a judgement call on metal/model.bin, which has a .bin name but isn't pickle; read strictly by extension it would be 15
- Reasoning control counts low, medium and high effort as a budget, as the calibration did, though thinking can't be switched off
- Unchecked, whether the 20b card's Transformers route fetches kernels-community/gpt-oss-triton-kernels from the Hub for MXFP4, as gpt-oss-120b's does. It would be code from outside the model repository, so no deduction was taken, as for gpt-oss-120b. A reviewer may read it otherwise, which would take 10 off file safety for both

## Weaknesses

- No reply from OpenAI to any discussion since August 2025, and template reports wait unanswered
- One tool call per message
- Reasoning can be lowered but not switched off
- Only one quantised build from OpenAI, and its weights and template haven't changed since August 2025
- Safety results for this size are OpenAI's own, and it trails o4-mini on the instruction hierarchy

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Start it with `ollama run gpt-oss:20b` (the build that fits the reference machine), or `vllm serve openai/gpt-oss-20b` with a current vLLM release rather than the card's pinned vllm==0.10.1+gptoss wheel
- Use the harmony format through the chat template or openai-harmony; OpenAI says the model doesn't work correctly without it
- Parse tool calls and reasoning with openai-harmony; calls come on the commentary channel ending in `<|call|>`, reasoning on the analysis channel
- Set effort with `Reasoning: high` (or low, medium) in the system message, or `reasoning_effort` in the template; medium is the default
- Ask for JSON by putting a JSON Schema under `# Response Formats` in the developer message
- Drop earlier reasoning after a final answer but pass it back after tool calls, and don't show the analysis channel to users
- Expect one tool call per turn; run several in sequence rather than in parallel

## What the review panel asked for

- update the vLLM pin
- answer template reports
- a priced OpenAI endpoint
- a supported-language statement
- an outside safety assessment for the 20b

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.