[{"id":"jev","name":"Jev 1.13","vendor":"TypeSafe AI","kind":"decision","released":"2026-09-15","question_types":["yes_no","choice","score"],"max_options":255,"context":"64k tokens per request (state + all questions); 32k for state plus longest question; text only","latency":"Vendor-reported (via OpenRouter Ori Eval, as cited on madewithjev.com): median 154 ms vs 860 ms for GPT-5.6 Luna. Cloudflare measured Jev median 524.1 ms (p95 536.0 ms) across 43 benchmarks; Red Hat measured median 348.1 ms (prompt injection) and 360.4 ms (content safety) over a remote API from a UK client.","price":"$0.042 per 1M input tokens; output tokens free (direct API). Same $0.042/1M listed on Eden AI's page for typesafe/jev-latest.","access":["Hosted API (api.typesafe.ai/v1/systemone)","OpenRouter (typesafe/jev-1.13)"],"open_weights":false,"license":null,"params":null,"status":"Direct API: status not stated in docs. Eden AI page describes its Jev decisions endpoint as alpha (shapes, IDs, pricing may change without deprecation).","benchmarks":[{"name":"Red Hat prompt injection (EvalHub NeMo Guardrails), accuracy, Jev-1.13.0 via API","score":"86.35% (rank 4 of 9)","source":"https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails"},{"name":"Red Hat content safety (toxicity-profanity-safety), accuracy","score":"86.20% (rank 1 of 9)","source":"https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails"},{"name":"Nimble authors' held-out set agreement (Bespoke Labs figure)","score":"93.21% for Jev 1.13.0","source":"https://ollama.com/library/nimble"},{"name":"Nimble authors' mean accuracy over 13 public datasets (Bespoke Labs figure)","score":"76.0%","source":"https://ollama.com/library/nimble"},{"name":"Cloudflare Jev Decision Index, e.g. BANKING77 macro-F1 (Cloudflare figure)","score":"79.74 (Clef 94.20)","source":"https://blog.cloudflare.com/clef-decision-models/"},{"name":"MediaPipe Decision Maker test (author-reported on madewithjev.com)","score":"92% for Jev vs 75% for MediaPipe alone","source":"https://madewithjev.com"}],"summary":"Jev is TypeSafe's closed, hosted decision model that returns probabilities for yes/no, pick-one and score questions at $0.042 per million input tokens.","url":"https://docs.typesafe.ai/api","sources":[{"label":"Jev API docs (endpoint, question types, 255 choice options, score 2-10 levels)","url":"https://docs.typesafe.ai/api"},{"label":"Jev models page (64k/32k context, pricing, rate limits, text only)","url":"https://docs.typesafe.ai/models"},{"label":"Made with Jev launch page (15 Sep 2026, latency/other claims, weights unpublished)","url":"https://madewithjev.com"},{"label":"Eden AI decisions docs (typesafe/jev-latest, alpha warning, 255 options, price)","url":"https://www.edenai.co/docs/v3/llms/decisions"},{"label":"OpenRouter Jev recipe (confirms typesafe/jev-1.13 id)","url":"https://openrouter.ai/labs/jev/compile"},{"label":"Red Hat benchmark","url":"https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails"},{"label":"Cloudflare Clef blog","url":"https://blog.cloudflare.com/clef-decision-models/"},{"label":"Ollama Nimble page","url":"https://ollama.com/library/nimble"}],"short":{"price":"$0.042/1M input tokens (output free)","latency":"154 ms median (vendor)","access":["API","OpenRouter"],"status":"Not stated (Eden AI: alpha)","context":"64k tokens per request; text only"}},{"id":"openai-decisions","name":"OpenAI Decisions API (GPT-6 Luna)","vendor":"OpenAI","kind":"decision","released":null,"question_types":["yes_no","choice","score"],"max_options":null,"context":"Not stated numerically; accepts text and inline base64 images (no hosted image URLs or file_id)","latency":"Vendor claim: about 10x faster than the Responses API; no absolute figures. Note: OpenRouter's Ori Eval, as cited on madewithjev.com, reports median 860 ms for GPT-5.6 Luna (third-party-reported, different setup).","price":"$0.10 per 1M input tokens; no output, cache-read or cache-write charges; regional processing premiums and long-context multipliers apply","access":["Hosted API (POST /v1/decisions)"],"open_weights":false,"license":null,"params":null,"status":"Public beta; OpenAI expects GA 'in the coming weeks'","benchmarks":[],"summary":"OpenAI's Decisions API answers predicate, choice and score questions with probabilities using the gpt-6-luna model, currently in public beta.","url":"https://developers.openai.com/api/docs/guides/decisions","sources":[{"label":"OpenAI Decisions guide","url":"https://developers.openai.com/api/docs/guides/decisions"},{"label":"Made with Jev (Luna latency from OpenRouter eval)","url":"https://madewithjev.com"}],"short":{"price":"$0.10/1M input tokens","latency":"~10x faster than Responses (vendor)","access":["API"],"status":"Public beta","context":"Text and inline base64 images"}},{"id":"clef","name":"Clef","vendor":"Cloudflare","kind":"decision","released":"2026-10-01","question_types":["yes_no","choice","score"],"max_options":null,"context":"65,536 tokens on Workers AI (blog says 64k); multimodal, up to 4 images per request; 1-64 questions per request","latency":"Vendor-reported by Cloudflare across 43 eval benchmarks: median 209.3 ms, p95 238.6 ms","price":"$0.24 per 1M input tokens; no charge for output tokens (Workers AI model page)","access":["Hosted API (Workers AI @cf/cloudflare/clef)","Open weights (Hugging Face)","Local via SGLang or Ollama"],"open_weights":true,"license":"Apache-2.0","params":"27B (post-trained from Qwen3.8-27B)","status":"Released as hosted Workers AI model and open weights; beta status not stated on model page","benchmarks":[{"name":"Cloudflare-run Jev Decision Index, BANKING77 macro-F1 (vendor figure)","score":"94.20 (Jev 79.74)","source":"https://blog.cloudflare.com/clef-decision-models/"},{"name":"Cloudflare-run Jev Decision Index, BFCL case exact (vendor figure)","score":"98.47 (Jev 95.75)","source":"https://blog.cloudflare.com/clef-decision-models/"},{"name":"Cloudflare-run Jev Decision Index, When2Call accuracy (vendor figure; Jev ahead)","score":"72.37 (Jev 80.97)","source":"https://blog.cloudflare.com/clef-decision-models/"},{"name":"Decision Index 0.2.1 (chance-corrected, community page)","score":"not listed for Clef in fetched table","source":"https://llmconfigurator.com/en/guides/decision-models"}],"summary":"Cloudflare's open-weight 27B multimodal decision model, hosted on Workers AI at $0.24 per million input tokens.","url":"https://developers.cloudflare.com/workers-ai/models/clef","sources":[{"label":"Cloudflare blog announcement (date, latency, benchmarks, Apache 2.0)","url":"https://blog.cloudflare.com/clef-decision-models/"},{"label":"Workers AI Clef model page (price, 65,536 context, limits)","url":"https://developers.cloudflare.com/workers-ai/models/clef"},{"label":"Hugging Face Clef (27B, Apache-2.0, /v1/systemone via SGLang)","url":"https://huggingface.co/Cloudflare/clef"},{"label":"llmconfigurator guide (Ollama 0.35 added clef, clef-flash)","url":"https://llmconfigurator.com/en/guides/decision-models"}],"short":{"price":"$0.24/1M input tokens","latency":"209.3 ms median (vendor)","access":["API","Open weights","Local"],"status":"Released; beta not stated","params":"27B (from Qwen3.8-27B)","license":"Apache-2.0","context":"65,536 tokens (Workers AI)"}},{"id":"clef-flash","name":"Clef Flash","vendor":"Cloudflare","kind":"decision","released":"2026-10-01","question_types":["yes_no","choice","score"],"max_options":null,"context":null,"latency":"Vendor-reported by Cloudflare across 43 eval benchmarks: median 38.8 ms, p95 122.4 ms","price":null,"access":["Hosted API (Workers AI @cf/cloudflare/clef-flash)","Open weights (Hugging Face)","Local via Ollama"],"open_weights":true,"license":"Apache-2.0","params":"Backbone Qwen3.5-9B (total parameter count not stated)","status":"Released as hosted Workers AI model and open weights","benchmarks":[{"name":"Cloudflare-run Jev Decision Index, BANKING77 macro-F1 (vendor figure)","score":"90.93 (Jev 79.74)","source":"https://blog.cloudflare.com/clef-decision-models/"},{"name":"Cloudflare-run Jev Decision Index, CLINC150+OOS macro-F1 (vendor figure)","score":"66.77 (Jev 89.27)","source":"https://blog.cloudflare.com/clef-decision-models/"}],"summary":"The smaller, faster Cloudflare decision variant, with a vendor-reported median latency under 40 ms.","url":"https://blog.cloudflare.com/clef-decision-models/","sources":[{"label":"Cloudflare blog announcement","url":"https://blog.cloudflare.com/clef-decision-models/"},{"label":"Workers AI Clef page (clef-flash accepted as model param; no price/limits given)","url":"https://developers.cloudflare.com/workers-ai/models/clef"},{"label":"Hugging Face Clef (Clef-Flash is smaller variant)","url":"https://huggingface.co/Cloudflare/clef"},{"label":"llmconfigurator guide (Ollama 0.35)","url":"https://llmconfigurator.com/en/guides/decision-models"}],"short":{"latency":"38.8 ms median (vendor)","access":["API","Open weights","Local (Ollama)"],"status":"Released (hosted + open weights)","params":"Qwen3.5-9B backbone","license":"Apache-2.0"}},{"id":"nimble","name":"Nimble 9B","vendor":"Bespoke Labs","kind":"decision","released":null,"question_types":["yes_no","choice","score"],"max_options":26,"context":"256K window listed by Ollama, but each question's prompt must fit 8,192 tokens; up to 64 questions and 64 KiB request body","latency":"Vendor claim (Ollama page): answers in under 100 ms on a MacBook Pro with M5 Max","price":"Free to run locally (open weights)","access":["Open weights / local (ollama pull nimble)","Ollama /v1/systemone endpoint"],"open_weights":true,"license":"Apache-2.0","params":"9B (fine-tuned from Qwen3.5-9B)","status":"Available in Ollama library (Ollama 0.35 added it, 29 Sep 2026)","benchmarks":[{"name":"Mean accuracy across 13 public datasets (Bespoke Labs figure)","score":"75.7% (Jev 1.13: 76.0%)","source":"https://ollama.com/library/nimble"},{"name":"Held-out set agreement (Bespoke Labs figure)","score":"90.12% (292/324) vs Jev 93.21%","source":"https://ollama.com/library/nimble"},{"name":"Public benchmark averages by type (Bespoke Labs figure)","score":"choice 81.6%, boolean 80.2%, score 54.6%","source":"https://ollama.com/library/nimble"},{"name":"Decision Index 0.2.1 (Nimble 9B v2 repo; chance-corrected)","score":"39.57 (ECE 0.024)","source":"https://llmconfigurator.com/en/guides/decision-models"}],"summary":"Nimble is an Apache-2.0 9B open decision model from Bespoke Labs that runs locally through Ollama's Jev-compatible System One endpoint.","url":"https://ollama.com/library/nimble","sources":[{"label":"Ollama Nimble library page","url":"https://ollama.com/library/nimble"},{"label":"llmconfigurator guide (Ollama 0.35 date, Index score)","url":"https://llmconfigurator.com/en/guides/decision-models"}],"short":{"price":"Free (open weights)","latency":"Under 100 ms on M5 Max (vendor)","access":["Open weights","Local (Ollama)"],"status":"Available in Ollama library","params":"9B (from Qwen3.5-9B)","license":"Apache-2.0","context":"256K listed; 8,192 tokens/question"}},{"id":"llama-guard-4","name":"Llama Guard 4","vendor":"Meta","kind":"guardrail","released":null,"classifies":"Prompt and response safety against 14 MLCommons-based hazard categories (S1 Violent Crimes ... S14 Code Interpreter Abuse); text and images","output":"Text: 'safe' or 'unsafe' followed by violated category codes (e.g. 'unsafe S9')","params":"12B (dense, pruned from Llama 4 Scout)","open_weights":true,"license":"Llama 4 Community License","access":["Hugging Face (gated)"],"price":null,"context":null,"languages":"English plus multilingual text; evaluated on French, German, Hindi, Italian, Portuguese, Spanish, Thai","latency":null,"benchmarks":[{"name":"Meta in-house test set, English, output filtering (vendor claim): recall 69%, FPR 11%, F1 61%","score":"F1 61%","source":"https://huggingface.co/meta-llama/Llama-Guard-4-12B"},{"name":"Meta in-house test set, multilingual (vendor claim)","score":"recall 43%, FPR 3%, F1 51%","source":"https://huggingface.co/meta-llama/Llama-Guard-4-12B"}],"summary":"Meta's 12B multimodal safe/unsafe classifier that names which of 14 hazard categories were violated.","url":"https://huggingface.co/meta-llama/Llama-Guard-4-12B","sources":[{"label":"Hugging Face model card","url":"https://huggingface.co/meta-llama/Llama-Guard-4-12B"}],"short":{"access":["Hugging Face (gated)"],"classifies":"Prompt/response safety, 14 categories","output":"'safe'/'unsafe' + category codes","params":"12B (dense)","license":"Llama 4 Community License"}},{"id":"shieldgemma","name":"ShieldGemma (2B / 9B / 27B)","vendor":"Google","kind":"guardrail","released":"2024-07-31","classifies":"Whether a prompt or response violates a supplied policy; 4 harm types: sexually explicit, dangerous content, hate speech, harassment","output":"Text starting 'Yes' or 'No'; optional score = probability from Yes/No token logits","params":"2B (listed as 3B in safetensors metadata); 9B and 27B siblings","open_weights":true,"license":"Gemma license (gated, accept Google terms)","access":["Hugging Face (gated)"],"price":null,"context":null,"languages":"English only","latency":null,"benchmarks":[{"name":"ShieldGemma-2B optimal F1 / AU-PRC, SG Prompt (vendor claim)","score":"0.825 / 0.887","source":"https://huggingface.co/google/shieldgemma-2b"},{"name":"ShieldGemma-2B, OpenAI Mod","score":"0.812 / 0.887","source":"https://huggingface.co/google/shieldgemma-2b"},{"name":"ShieldGemma-2B, ToxicChat","score":"0.704 / 0.778","source":"https://huggingface.co/google/shieldgemma-2b"},{"name":"ShieldGemma-2B, SG Response","score":"0.743 / 0.802","source":"https://huggingface.co/google/shieldgemma-2b"}],"summary":"Google's Gemma-2-based text policy classifier that answers Yes/No on four harm types.","url":"https://huggingface.co/google/shieldgemma-2b","sources":[{"label":"Hugging Face model card (release date from Gemma release log; paper 2024-07-31)","url":"https://huggingface.co/google/shieldgemma-2b"},{"label":"Gemma release log","url":"https://ai.google.dev/gemma/docs/releases"}],"short":{"access":["Hugging Face (gated)"],"classifies":"Policy violation; 4 harm types","output":"'Yes'/'No' + optional score","params":"2B (9B, 27B siblings)","license":"Gemma license (gated)"}},{"id":"shieldgemma-2","name":"ShieldGemma 2","vendor":"Google","kind":"guardrail","released":"2025-03-10","classifies":"Image safety against 3 policies: sexually explicit, dangerous content, violence/gore","output":"Probability from Yes/No next-token likelihood per policy","params":"4B","open_weights":true,"license":"Gemma license (gated)","access":["Hugging Face (gated)"],"price":null,"context":null,"languages":null,"latency":null,"benchmarks":[{"name":"Google internal synthetic benchmark (~500 examples/policy), optimal F1 (vendor claim): sexually explicit 88.6, dangerous 93.7, violence/gore 85.0","score":"85.0-93.7 F1","source":"https://huggingface.co/google/shieldgemma-2-4b-it"}],"summary":"Google's image-only safety classifier built on Gemma 3 4B.","url":"https://huggingface.co/google/shieldgemma-2-4b-it","sources":[{"label":"Hugging Face model card","url":"https://huggingface.co/google/shieldgemma-2-4b-it"},{"label":"Gemma release log (ShieldGemma 2 listed 2025-03-10 with Gemma 3)","url":"https://ai.google.dev/gemma/docs/releases"}],"short":{"access":["Hugging Face (gated)"],"classifies":"Image safety, 3 policies","output":"Yes/No probability per policy","params":"4B","license":"Gemma license (gated)"}},{"id":"granite-guardian-4-1","name":"Granite Guardian 4.1 8B","vendor":"IBM","kind":"guardrail","released":"2026-04-29","classifies":"Harm (social bias, jailbreaking, violence, profanity, unethical behavior), RAG hallucination (context relevance, groundedness, answer relevance), function-calling hallucination, bring-your-own criteria","output":"Yes/No inside <score></score>; optional <think> reasoning","params":"8B","open_weights":true,"license":"Apache 2.0","access":["Hugging Face","Ollama (per third-party guide)"],"price":null,"context":null,"languages":"English only","latency":null,"benchmarks":[{"name":"OOD safety aggregate F1 (model card)","score":"0.79 no-think / 0.78 think","source":"https://huggingface.co/ibm-granite/granite-guardian-4.1-8b"},{"name":"LM-AggreFact RAG avg balanced accuracy","score":"0.760 no-think / 0.764 think","source":"https://huggingface.co/ibm-granite/granite-guardian-4.1-8b"},{"name":"FC Reward Bench balanced accuracy","score":"0.79 no-think / 0.78 think","source":"https://huggingface.co/ibm-granite/granite-guardian-4.1-8b"}],"summary":"IBM's Apache-licensed 8B judge for harm, RAG groundedness and agent tool-call risks, English only.","url":"https://huggingface.co/ibm-granite/granite-guardian-4.1-8b","sources":[{"label":"Hugging Face model card (release 'April 2026')","url":"https://huggingface.co/ibm-granite/granite-guardian-4.1-8b"},{"label":"Granite 4.1 family date 2026-04-29 (third-party coverage)","url":"https://chatforest.com/reviews/ibm-granite-4-1-dense-enterprise-llm-family-review/"}],"short":{"access":["Hugging Face","Ollama"],"classifies":"Harm, hallucination, custom criteria","output":"Yes/No in <score> tags","params":"8B","license":"Apache 2.0"}},{"id":"nemotron-3-5-content-safety","name":"Nemotron 3.5 Content Safety","vendor":"NVIDIA","kind":"guardrail","released":"2026-06-02","classifies":"Prompt and response safety (text and optional image) using Aegis 2.0 taxonomy, plus custom user-defined policies with optional reasoning","output":"Text lines: User Safety (safe/unsafe), Response Safety, Safety Categories (comma list); optional <think> trace in custom-policy mode","params":"4B (Gemma-3-4B-it base)","open_weights":true,"license":"OpenMDW License v1.1 plus Gemma Terms of Use (NGC page lists NVIDIA Nemotron Open Model License instead)","access":["Hugging Face","NVIDIA NIM container (NGC)","build.nvidia.com"],"price":null,"context":"128K tokens","languages":"12: English, Arabic, German, Spanish, French, Hindi, Japanese, Thai, Dutch, Italian, Korean, Chinese","latency":"No numeric latency on model card","benchmarks":[{"name":"Model card (vendor): Aegis 2.0 / Wildguard prompt F1 ~0.85-0.86","score":"see card","source":"https://huggingface.co/nvidia/nemotron-3.5-content-safety"},{"name":"Red Hat benchmark, prompt injection, default vs custom policies (third party)","score":"69.37% vs 84.84% accuracy; median 239.6/240.4 ms","source":"https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails"},{"name":"Red Hat benchmark, content safety","score":"84.67% default / 85.07% custom accuracy; ~230-242 ms","source":"https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails"}],"summary":"NVIDIA's 4B multilingual, multimodal guard that also accepts custom policies.","url":"https://huggingface.co/nvidia/nemotron-3.5-content-safety","sources":[{"label":"Hugging Face model card","url":"https://huggingface.co/nvidia/nemotron-3.5-content-safety"},{"label":"Search listing for NGC/NIM and license discrepancy","url":"https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/models/nemotron-3.5-content-safety"},{"label":"Red Hat article","url":"https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails"}],"short":{"access":["Hugging Face","NVIDIA NIM","build.nvidia.com"],"classifies":"Prompt/response safety (Aegis 2.0)","output":"User/Response Safety + categories","params":"4B (Gemma-3-4B-it base)","license":"OpenMDW v1.1 + Gemma Terms","context":"128K tokens"}},{"id":"qwen3guard-gen","name":"Qwen3Guard-Gen (0.6B / 4B / 8B)","vendor":"Alibaba (Qwen)","kind":"guardrail","released":"2025-10-16","classifies":"Prompt and response safety with 3 severity levels (Safe/Controversial/Unsafe); 8 categories (Violent, Non-violent Illegal Acts, Sexual Content, PII, Suicide & Self-Harm, Unethical Acts, Politically Sensitive Topics, Copyright Violation), Jailbreak (input only), refusal detection for responses","output":"Text lines: 'Safety:', 'Categories:', and 'Refusal:' for responses. A Stream variant adds token-level moderation during generation","params":"0.6B, 4B, 8B","open_weights":true,"license":"Apache 2.0","access":["Hugging Face","vLLM / SGLang"],"price":null,"context":"Serving examples use 32768 tokens (not a stated spec)","languages":"119 languages and dialects","latency":null,"benchmarks":[{"name":"Paper claims state-of-the-art prompt/response classification (vendor claim); no numbers on the model card or arXiv abstract. Shieldstral's card lists Qwen3Guard-8B best on WildGuardTest prompt F1 88.2 and BeaverTails response 85.9 (third-party table)","score":"88.2 / 85.9 F1","source":"https://huggingface.co/mistralai/Shieldstral-1.0"}],"summary":"Alibaba's multilingual guard family that returns a three-level verdict plus categories, with a streaming variant.","url":"https://huggingface.co/Qwen/Qwen3Guard-Gen-8B","sources":[{"label":"Hugging Face model card","url":"https://huggingface.co/Qwen/Qwen3Guard-Gen-8B"},{"label":"arXiv technical report (date 2025-10-16)","url":"https://arxiv.org/abs/2510.14276"},{"label":"Shieldstral card comparison table","url":"https://huggingface.co/mistralai/Shieldstral-1.0"}],"short":{"access":["Hugging Face","vLLM / SGLang"],"classifies":"3 severity levels, 8 categories","output":"Safety/Categories/Refusal lines","params":"0.6B, 4B, 8B","license":"Apache 2.0","context":"32768 tokens (serving examples)"}},{"id":"gpt-oss-safeguard","name":"gpt-oss-safeguard (20b / 120b)","vendor":"OpenAI","kind":"guardrail","released":"2025-10-29","classifies":"Any developer-supplied written policy applied at inference time to messages, completions or full chats","output":"Classification with an inspectable reasoning trace (harmony format)","params":"20b (21B total, 3.6B active); 120b (117B total, 5.1B active)","open_weights":true,"license":"Apache 2.0 plus gpt-oss usage policy","access":["Hugging Face","Ollama","LM Studio","vLLM","Groq via HF Inference Providers","Azure AI catalog"],"price":null,"context":null,"languages":null,"latency":"Reasoning model, slower and heavier than small classifiers (vendor says classifiers trained on tens of thousands of labels can beat it on complex tasks)","benchmarks":[{"name":"OpenAI internal multi-policy benchmark, accuracy (vendor claim): gpt-oss-safeguard-120b vs GPT-5","score":"46.3% vs 43.2%","source":"https://www.helpnetsecurity.com/2025/10/29/openai-gpt-oss-safeguard-safety-models/"},{"name":"Shieldstral card table (third party): OpenAI Moderation prompt F1 84.0; WildGuardTest response 80.7; HarmBench response 88.2 for GPT-OSS-Safeguard-20B","score":"80.7-88.2 F1","source":"https://huggingface.co/mistralai/Shieldstral-1.0"}],"summary":"OpenAI's open-weight reasoning models that judge content against whatever policy text you give them.","url":"https://huggingface.co/openai/gpt-oss-safeguard-20b","sources":[{"label":"Hugging Face model card","url":"https://huggingface.co/openai/gpt-oss-safeguard-20b"},{"label":"Launch coverage (date 2025-10-29, benchmark)","url":"https://www.helpnetsecurity.com/2025/10/29/openai-gpt-oss-safeguard-safety-models/"},{"label":"OpenAI technical report page (not fetched; referenced via search)","url":"https://openai.com/index/gpt-oss-safeguard-technical-report"},{"label":"OpenAI help center: not served via API","url":"https://help.openai.com/en/articles/11870455"}],"short":{"latency":"Reasoning model; slower (vendor)","access":["Hugging Face","Ollama","vLLM"],"classifies":"Developer-supplied written policy","output":"Classification + reasoning trace","params":"20b (21B total); 120b (117B total)","license":"Apache 2.0 + gpt-oss usage policy"}},{"id":"openai-omni-moderation","name":"OpenAI Moderation API (omni-moderation-latest)","vendor":"OpenAI","kind":"guardrail","released":null,"classifies":"13 categories: harassment, harassment/threatening, hate, hate/threatening, illicit, illicit/violent, sexual/minors, self-harm (+intent, +instructions), sexual, violence, violence/graphic; text and image","output":"JSON: flagged boolean, per-category booleans, per-category scores 0-1, applied input types","params":null,"open_weights":false,"license":"Proprietary API","access":["OpenAI API /moderations endpoint"],"price":"Free to use (per OpenAI docs)","context":"Image files up to 20 MB; text limit not stated","languages":null,"latency":null,"benchmarks":[],"summary":"OpenAI's free hosted moderation endpoint returning per-category flags and scores for text and images.","url":"https://developers.openai.com/api/docs/guides/moderation","sources":[{"label":"OpenAI moderation guide","url":"https://developers.openai.com/api/docs/guides/moderation"}],"short":{"price":"Free (per OpenAI docs)","access":["API"],"classifies":"13 categories; text and image","output":"JSON flags + scores 0-1","license":"Proprietary API","context":"Images up to 20 MB"}},{"id":"llama-prompt-guard-2","name":"Llama Prompt Guard 2 (86M / 22M)","vendor":"Meta","kind":"guardrail","released":null,"classifies":"Direct jailbreak / prompt-injection attempts in prompts","output":"Binary label benign vs malicious","params":"86M (mDeBERTa-based, listed 0.3B on Hub); 22M English-focused variant","open_weights":true,"license":"Llama 4 Community License (gated)","access":["Hugging Face (gated)"],"price":null,"context":"512 tokens (split longer inputs)","languages":"86M: English, French, German, Hindi, Italian, Portuguese, Spanish, Thai; 22M for English-only","latency":"92.4 ms per classification on A100, 512 tokens (86M, model card)","benchmarks":[{"name":"86M direct jailbreak detection, English AUC (vendor claim)","score":"0.998","source":"https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M"},{"name":"86M recall at 1% FPR, English","score":"97.5%","source":"https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M"},{"name":"86M multilingual AUC","score":"0.995","source":"https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M"}],"summary":"Meta's tiny prompt-attack detector with a 512-token window.","url":"https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M","sources":[{"label":"Hugging Face model card","url":"https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M"},{"label":"Meta Prompt Guard docs (22M/86M sizes)","url":"https://www.llama.com/docs/model-cards-and-prompt-formats/prompt-guard/"}],"short":{"latency":"92.4 ms per classification (A100)","access":["Hugging Face (gated)"],"classifies":"Jailbreak/prompt-injection attempts","output":"Binary: benign vs malicious","params":"86M (22M English variant)","license":"Llama 4 Community License","context":"512 tokens"}},{"id":"protectai-deberta-pi-v2","name":"ProtectAI deberta-v3-base-prompt-injection-v2","vendor":"Protect AI","kind":"guardrail","released":null,"classifies":"Prompt injection in input text","output":"Binary label 0 benign / 1 injection-detected","params":"~200M (DeBERTa-v3-base)","open_weights":true,"license":"Apache 2.0","access":["Hugging Face"],"price":null,"context":"512 tokens used in examples (not stated as limit)","languages":"English only","latency":"Red Hat benchmark: median 54.1 ms (Table 1; appendix says 80.4 ms), on MacBook Pro M1 CPU","benchmarks":[{"name":"Model card (vendor): 20,000 held-out prompts, accuracy / F1","score":"95.25% / 95.49%","source":"https://huggingface.co/protectai/deberta-v3-base-prompt-injection-v2"},{"name":"Red Hat benchmark (third party), prompt-injection accuracy; Qwen3.6-35B scored 89.31%","score":"89.01%","source":"https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails"}],"summary":"A small English-only DeBERTa classifier for prompt injection; the card says the project is archived.","url":"https://huggingface.co/protectai/deberta-v3-base-prompt-injection-v2","sources":[{"label":"Hugging Face model card (archived notice)","url":"https://huggingface.co/protectai/deberta-v3-base-prompt-injection-v2"},{"label":"Red Hat benchmark article","url":"https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails"}],"short":{"latency":"54.1 ms median (Red Hat)","access":["Hugging Face"],"classifies":"Prompt injection in input text","output":"Binary: 0 benign / 1 injection","params":"~200M (DeBERTa-v3-base)","license":"Apache 2.0","context":"512 tokens (in examples)"}},{"id":"azure-ai-content-safety","name":"Azure AI Content Safety (incl. Prompt Shields)","vendor":"Microsoft","kind":"guardrail","released":null,"classifies":"Text and image harm detection (sexual, violence, hate, self-harm) with multi-severity levels; Prompt Shields for user-input attacks; protected material; groundedness (preview); task adherence; custom categories","output":"Per-category severity levels; Prompt Shields returns attack detection","params":null,"open_weights":false,"license":"Proprietary API","access":["Azure AI Content Safety API","Content Safety Studio"],"price":"Free F0 tier: 5,000 text records and 5,000 images per month (Prompt Shields included in text). Standard S0 per-1,000 prices not shown on the pricing page as fetched","context":null,"languages":"Trained/tested on Chinese, English, French, German, Spanish, Italian, Japanese, Portuguese; protected material, groundedness and custom categories (standard) are English only","latency":null,"benchmarks":[],"summary":"Microsoft's hosted moderation and prompt-attack API with severity scores.","url":"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview","sources":[{"label":"Azure docs overview","url":"https://learn.microsoft.com/en-us/azure/ai-services/content-safety/overview"},{"label":"Azure pricing page","url":"https://azure.microsoft.com/en-us/pricing/details/cognitive-services/content-safety/"}],"short":{"price":"Free F0: 5,000 text + 5,000 images/mo","access":["API","Content Safety Studio"],"classifies":"Text/image harm; Prompt Shields","output":"Per-category severity levels","license":"Proprietary API"}},{"id":"shieldstral-1-0","name":"Shieldstral 1.0","vendor":"Mistral AI","kind":"guardrail","released":"2026-08-04","classifies":"Text, image or text+image content against a natural-language policy","output":"Single yes/no token; softmax over logits gives a continuous score","params":"3B (4B in file metadata)","open_weights":true,"license":"Apache 2.0","access":["Hugging Face","Mistral API (per third-party coverage)"],"price":null,"context":"Trained to 32k tokens (recommended limit)","languages":"12: English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, Arabic, Russian","latency":"No latency figures; fits in 16GB VRAM","benchmarks":[{"name":"Model card (vendor): WildGuardTest prompt F1 88.1, ToxicChat 84.1, Aegis v2 86.2, HarmBench 99.4","score":"84.1-99.4 F1","source":"https://huggingface.co/mistralai/Shieldstral-1.0"},{"name":"Red Hat benchmark, prompt injection (Table 1; appendix conflicts)","score":"72.02%; median 187.9 ms","source":"https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails"},{"name":"Red Hat benchmark, content safety (Table 2; appendix conflicts)","score":"74.80%; median 191.9 ms","source":"https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails"}],"summary":"Mistral's 3B policy-adaptive multimodal safety scorer.","url":"https://huggingface.co/mistralai/Shieldstral-1.0","sources":[{"label":"Hugging Face model card","url":"https://huggingface.co/mistralai/Shieldstral-1.0"},{"label":"Release date 2026-08-04 (third-party coverage)","url":"https://rits.shanghai.nyu.edu/ai/mistral-releases-shieldstral-a-3b-policy-adaptive-safety-classifier"},{"label":"Red Hat article","url":"https://developers.redhat.com/articles/2026/10/02/benchmarking-ai-decision-models-against-traditional-guardrails"}],"short":{"access":["Hugging Face","Mistral API"],"classifies":"Content vs natural-language policy","output":"Yes/no token + softmax score","params":"3B (4B in file metadata)","license":"Apache 2.0","context":"32k tokens (recommended)"}},{"id":"jev-router","name":"TypeSafe Jev Router","kind":"router","vendor":"TypeSafe","released":"2026-09-25","routes_between":"Models across the OpenRouter ecosystem (specific models not named on the page)","decides":"Which model and which reasoning effort answers each request","how_decides":"Runs on Jev, TypeSafe's first System One model; adapts as the conversation evolves (vendor)","access":["OpenRouter API (typesafe/jev-router)","Chat Completions endpoint"],"price":"Not stated; OpenRouter lists price as -1 (variable)","open_weights":null,"license":null,"status":"Live on OpenRouter; 1M context listed","latency":null,"benchmarks":[],"summary":"Picks the best model and reasoning effort for each request, balancing quality, speed and cost (vendor).","url":"https://openrouter.ai/typesafe/jev-router","sources":[{"label":"OpenRouter model page","url":"https://openrouter.ai/typesafe/jev-router"},{"label":"OpenRouter models API (created 1790363560 = 2026-09-25; context 1000000; pricing -1)","url":"https://openrouter.ai/api/v1/models"},{"label":"OpenRouter API guide","url":"https://openrouter.ai/typesafe/jev-router/api"}],"short":{"routes_between":"OpenRouter model ecosystem","decides":"Model + reasoning effort","price":"Variable (OpenRouter lists -1)","access":["OpenRouter API"],"status":"Live; 1M context","license":"Unknown"}},{"id":"openrouter-auto","name":"OpenRouter Auto Router","kind":"router","vendor":"OpenRouter","released":"2023-11-08","routes_between":"Ranked candidates per task type from models on OpenRouter, filtered by your allow/exclude lists, cost tier and policies","decides":"Which model answers (cost tier low/medium/high/xhigh/max); beta track tests new behaviors","how_decides":"Classifies each prompt into about 30 task types, then ranks models by their share of OpenRouter spend for that task over a trailing 7 days; docs do not mention Not Diamond (vendor)","access":["OpenRouter API (openrouter/auto)","openrouter/auto-beta (early access)"],"price":"No additional fee; pay the selected model's standard rate","open_weights":false,"license":null,"status":"Live; beta track available","latency":null,"benchmarks":[],"summary":"Routes each prompt by task type using what the OpenRouter community collectively spends on each model.","url":"https://openrouter.ai/docs/guides/routing/routers/auto-router","sources":[{"label":"OpenRouter Auto Router docs","url":"https://openrouter.ai/docs/guides/routing/routers/auto-router"},{"label":"OpenRouter models API (created 1699401600 = 2023-11-08)","url":"https://openrouter.ai/api/v1/models"}],"short":{"routes_between":"OpenRouter models by task type","decides":"Model, by cost tier","price":"Selected model's rate; no fee","access":["OpenRouter API"],"status":"Live (beta track too)","license":"Proprietary service"}},{"id":"not-diamond","name":"Not Diamond","kind":"router","vendor":"Not Diamond","released":null,"routes_between":"Models you configure; pre-trained or custom router (full list not on docs page)","decides":"Which model to call; returns a recommendation, you make the call via your own gateway","how_decides":"Pre-trained router or custom routers trained on your data; mechanism not detailed","access":["REST API","Python SDK","TypeScript SDK"],"price":"$0.05 per million tokens routed (pay-as-you-go, vendor pricing page); enterprise custom","open_weights":false,"license":null,"status":"Live; pay-as-you-go access by application per pricing page","latency":null,"benchmarks":[],"summary":"Hosted router that recommends the best model per request, with optional custom routers and prompt optimization.","url":"https://www.notdiamond.ai","sources":[{"label":"Not Diamond pricing","url":"https://www.notdiamond.ai/pricing"},{"label":"Not Diamond docs","url":"https://docs.notdiamond.ai/docs/what-is-not-diamond"},{"label":"Third-party summary (no open-source or self-host found)","url":"https://tooldirectory.ai/tools/not-diamond"}],"short":{"routes_between":"Models you configure","decides":"Which model to call","price":"$0.05 per 1M tokens routed","access":["REST API","Python / TS SDK"],"status":"Live","license":"Proprietary service"}},{"id":"routellm","name":"RouteLLM","kind":"router","vendor":"LMSYS (lm-sys)","released":"2024-06-26","routes_between":"Two models: a strong and a weak one you choose","decides":"Whether to send each query to the strong or the weak model","how_decides":"Routers trained on preference data: mf (matrix factorization), sw_ranking (weighted Elo), bert classifier, causal_llm classifier","access":["Open-source Python library","OpenAI-compatible server"],"price":"Free software; you pay the models you call","open_weights":true,"license":"Apache-2.0","status":"Open source; repo active (5.6k stars)","latency":null,"benchmarks":[{"name":"Cost reduction at 95% of GPT-4 performance on MT Bench (vendor)","score":"up to 85%","source":"https://github.com/lm-sys/RouteLLM"}],"summary":"Open-source framework that routes each query between a strong and a weak model using routers trained on preference data.","url":"https://github.com/lm-sys/RouteLLM","sources":[{"label":"RouteLLM GitHub","url":"https://github.com/lm-sys/RouteLLM"},{"label":"Paper (v1 2024-06-26)","url":"https://arxiv.org/abs/2406.18665"}],"short":{"routes_between":"Strong vs weak model","decides":"Strong or weak model","price":"Free; pay the models","access":["Python library","OpenAI-compatible server"],"status":"Open source","license":"Apache-2.0"}},{"id":"azure-model-router","name":"Azure Foundry Model Router","kind":"router","vendor":"Microsoft","released":null,"routes_between":"About 38 models from OpenAI, Anthropic, xAI, DeepSeek, Meta and Fireworks (current version 2025-11-18), or a subset you pick","decides":"Which model answers; routing mode Quality, Balanced (default) or Cost; reasoning vs non-reasoning model","how_decides":"Trained language model that analyzes complexity, reasoning and task type in real time; routing decision uses text input only","access":["Foundry model deployment","Global Standard / Data Zone Standard"],"price":"Input prompts charged at the rate on the Azure pricing page; amount not retrieved","open_weights":false,"license":null,"status":"Live; version 2025-11-18 active, older versions frozen","latency":null,"benchmarks":[],"summary":"A trained model deployed like any Foundry model that picks the underlying LLM per prompt, with quality, balanced and cost modes.","url":"https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-router","sources":[{"label":"Microsoft Learn: model router","url":"https://learn.microsoft.com/en-us/azure/ai-foundry/openai/concepts/model-router"}],"short":{"routes_between":"OpenAI, Claude, xAI, DeepSeek +","decides":"Model + routing mode","price":"See Azure pricing page","access":["Foundry deployment"],"status":"Live; v2025-11-18","license":"Proprietary service"}},{"id":"bedrock-prompt-routing","name":"Amazon Bedrock Intelligent Prompt Routing","kind":"router","vendor":"Amazon Web Services","released":"2025-04-22","routes_between":"Two models within one family (default routers use Nova or Llama; configured routers add models such as Claude)","decides":"Which of two same-family models answers, using a response-quality-difference threshold and a fallback model","how_decides":"Predicts the response quality of each model for the prompt, then picks the best quality/cost combination; English-optimized","access":["Bedrock console","AWS CLI / SDK (CreatePromptRouter)"],"price":"Routing price not stated on page; see Bedrock pricing page","open_weights":false,"license":null,"status":"GA since 2025-04-22","latency":null,"benchmarks":[{"name":"Cost reduction without losing accuracy (vendor)","score":"up to 30%","source":"https://aws.amazon.com/bedrock/intelligent-prompt-routing"}],"summary":"Routes between two models in the same family by predicting which will give the better response for the cost.","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-routing.html","sources":[{"label":"AWS docs","url":"https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-routing.html"},{"label":"AWS product page","url":"https://aws.amazon.com/bedrock/intelligent-prompt-routing"},{"label":"AWS GA announcement","url":"https://aws.amazon.com/about-aws/whats-new/2025/04/amazon-bedrock-intelligent-prompt-routing-generally-available"}],"short":{"routes_between":"Two models, same family","decides":"Which of two models","price":"See Bedrock pricing","access":["Bedrock console","AWS CLI / SDK"],"status":"GA","license":"Proprietary service"}},{"id":"vllm-semantic-router","name":"vLLM Semantic Router","kind":"router","vendor":"vLLM project","released":null,"routes_between":"Configured backends across local, private and cloud models","decides":"Which model handles each call, per your quality, latency and cost policy","how_decides":"Policy-driven decision layer; earlier project posts describe a lightweight classifier (ModernBERT) judging intent and complexity","access":["Open-source (GitHub)","Envoy / Kubernetes integration"],"price":"Free software; you pay your backends","open_weights":true,"license":"Apache-2.0","status":"Open source; v0.4 reported 2026-09-24","latency":null,"benchmarks":[{"name":"Accuracy gain on MMLU-Pro vs direct vLLM inference (authors' paper)","score":"+10.2 points","source":"https://arxiv.org/abs/2510.08731v1"},{"name":"Latency cut vs direct vLLM inference (authors' paper)","score":"47.1%","source":"https://arxiv.org/abs/2510.08731v1"}],"summary":"Open-source decision layer that routes each request to a model based on policy and request complexity.","url":"https://github.com/vllm-project/semantic-router","sources":[{"label":"GitHub repo","url":"https://github.com/vllm-project/semantic-router"},{"label":"Paper","url":"https://arxiv.org/abs/2510.08731v1"},{"label":"vLLM blog","url":"https://vllm.ai/blog/2025-09-11-semantic-router"}],"short":{"routes_between":"Your configured backends","decides":"Model per call, by policy","price":"Free; pay your backends","access":["GitHub","Envoy / Kubernetes"],"status":"Open source; v0.4","license":"Apache-2.0"}},{"id":"arch-router","name":"Arch-Router","kind":"router","vendor":"Katanemo","released":null,"routes_between":"Models you map to your own domain/action routes","decides":"Which user-defined route (and so which model) fits the query","how_decides":"1.5B generative model fine-tuned from Qwen2.5-1.5B-Instruct that maps a query to your domain and action policy, returning a JSON route name","access":["Hugging Face weights","Arch gateway (open source)"],"price":"Free weights; you pay the models you route to","open_weights":true,"license":"Katanemo license (katanemo-research)","status":"Open weights; paper published 2025-06-19","latency":null,"benchmarks":[],"summary":"Small open-weights model that routes a query to the model matching your own routing preferences.","url":"https://huggingface.co/katanemo/Arch-Router-1.5B","sources":[{"label":"Hugging Face model card","url":"https://huggingface.co/katanemo/Arch-Router-1.5B"},{"label":"Paper","url":"https://arxiv.org/abs/2506.16655"}],"short":{"routes_between":"Models mapped to your routes","decides":"Route (domain/action) per query","price":"Free; pay your models","access":["Hugging Face","Arch gateway"],"status":"Open weights","license":"Katanemo license"}}]