Decision models: TypeSafe AI's Jev (decisions instead of text)

status: watch, opened 2026-09-21, last moved 2026-10-22. AI models

The thread: the "System-One" model category — AI returning calibrated decisions (app state + typed query in, no text generation) instead of prose, pitched at automation economics (hallucination and cost cuts).
2026-09-21: opened. Origin identified: founder Diogo Almeida (ex-OpenAI researcher); a demand surge last week briefly knocked out TypeSafe's API; Theo Browne's framing — "roughly as intelligent as a switch statement" — and that's the point. Prior color: X launch Sep 18-19 (OpenRouter beta, ~40× GPT speed claim, unverified; Jev talk at AIE).
Watch for: independent evals of decision-mode vs text-gen on classification/routing workloads, pricing, whether agent plumbing (routing, tool calls) is the real wedge, fast-follower entrants.
2026-09-22: two follow-on writeups: Ecosistema Startup (Spanish) — Jev wired as an agent under prompt injection; Business Standard explainer of the framing. No independent evals yet; the injection surface is the new color — decision models deployed as agents inherit the prompt-injection surface the category was pitched as avoiding.
2026-09-22 (X layer, ai-social dig): OpenClaw shipped decision- model support in core + plugins last week (@jlehman_; blog: openclaw.ai/blog/decision-models-in-openclaw) — "Everything's coming up Jev!" The category's first harness-native adoption: the anti-plumbing pitch now distributes through agent plumbing.
2026-09-22 (evening): four channels live within ~4 days of launch (independent guide jevaiguide.com, checked Sep 19): (1) TypeSafe API (first-party: jev-1.13.0/jev-latest, api.typesafe.ai/v1/systemone; 64k context per request, 32k cap on state + longest question; $0.042/1M input, output free; gated by waitlist-approved account); (2) Vercel AI Gateway (typesafe-ai/jev; AI SDK 7+ experimental_evaluate; free until Sep 25, 2026 promo, card on file required); (3) OpenRouter (typesafe/jev-1.13 + jev-latest redirect; 32k context; $0.042/1M input / $0 output; OpenRouter carries a docs guide, cookbooks, and a Sep 19 "Jev vs LLM" tutorial — aggregator-grade distribution); (4) Cloudflare Workers AI (typesafe/jev, Workers/REST, 32k). All four serve the same Jev 1.13; request shapes differ (Vercel calls the yes/no type "boolean", TypeSafe calls it "noul" — not swappable by URL). Harness layer: OpenClaw core+plugins. Guide also documents integrations for browser/computer use and Claude Code & Codex. Economics note: $0.042/M input undercuts the cheap-frontier tier (GLM Flash $0.15, GPT-6 Luna $0.10) — decision models price like a feature, not a model. Still no independent evals vs text-gen on routing/classification workloads; ~40× GPT speed claim still unverified. Openness: closed source, no weights released (guide FAQ, Sep 19) — Ollama search returns zero Jev models; the product IS the typed-decisions API (Noul/ Choice/Score + state), so even open weights wouldn't be usable through a chat-completions surface. Integration path is API-only: OpenRouter ($0.042/M) or TypeSafe first-party (64k context, SDKs, waitlist); natural wedge = relevance gate (relevant/filtered/stale-update as Score/Choice) with the text model kept for the reason and summary prose; TVA false-positive classification is a Choice/Noul fit.
2026-09-23: the category's openness axis inverts — Nokia open-sourced AnyJev (marktechpost, single outlet, flag): a training-free layer that "turns any open LLM into a calibrated decision model," claiming 6.8× automatable-traffic lift. Same window: GLiClass (Knowledgator's open-source classifier, RT'd: "open-source Jev, before Jev existed") landed in OpenClaw via ONNX — local lightweight decisions instead of full LLM calls. Reading: Jev's closed-API scarcity is being attacked from both ends (a telecom giant's training-free layer + an OSS predecessor getting harness-native distribution). The moat to watch: if open calibration layers replicate the API's function at zero cost, the decision-model category's pricing power evaporates. Independent comparisons (Jev vs AnyJev vs GLiClass on routing) now exist as a testable question — still nobody has run them.
2026-09-26: the first third-party number lands on Jev itself — Theo - t3.gg (via steipete RT, ~$1,000 of tokens): benchmarking the Jev Router: "Performance on DeepSWE was roughly the same as GPT-6 Astra on low. It costs slightly more, and it took almost 5x longer to run." First independent numbers are a wash: same performance, more cost, more latency. Category pricing-power question now has a practitioner datapoint; single- benchmark caveat (DeepSWE only) applies.
2026-09-29: mainstream color arrives — Business Insider hackathon piece ("the new AI model taking San Francisco by storm"): the category's scene layer a day after the independent numbers were a wash (Theo, Sep 26). The hype-to-benchmark gap is now the story inside the category; the testable question (Jev vs AnyJev vs GLiClass) still has no runner.
2026-09-29 (evening): a fast-follower answers the watch — Liquid AI shipped d1 (marktechpost, single outlet): "a decision model that returns calibrated probabilities with zero output tokens" — one call for classification/routing/scoring. The category's second named vendor, and the first from an established lab (Liquid's edge-architecture line, not a TypeSafe bootstrap). Same caveats: no independent evals, no runner for Jev vs AnyJev vs d1 — the testable set just grew by one.
2026-09-30: the fast-follower list reaches the biggest name — OpenAI shipped a Decisions API at DevDay (Sep 29; syndicated coverage hit the feed Sep 30, MSN framing: "a Jev clone that confirms the importance of fast, cheap intelligence"): typed decisions in ~150ms on GPT-6 Luna from a fixed answer list. Origin press so far is small-outlet (byteiota, pasqualepillitteri.it) — verify against OpenAI's product page before grading specs. If real, the category now has OpenAI in it, the pricing-power question (open layers replicating the API's function) sharpens further, and the runner question becomes Jev vs AnyJev/d1 vs Decisions API.
2026-10-01: the category's biggest open-weights entrant — Perplexity open-sourced its decision model (Arav, ~2:30 PM ET
developers account): pplx-decider-v1-27b, multimodal, served through a Decisions API at $0.04/M input (vs Jev $0.042 — "fraction of the price" is rounding; the real pitch is open weights), self-graded 85.71% across benchmarks, claimed "higher than Jev." The open-weights counter to Jev's closed API moves from a telecom's training-free layer (AnyJev) to a scaled vendor's native model. Same day: inline TradingView charts in Computer; Amex-curated skill library for business-card members; Seedance 2.5 as a Computer skill (perplexity-index adjacency — the agentic stack widening). Runner question is now five-way — Jev vs AnyJev vs d1 vs OpenAI Decisions API vs pplx-decider — still no independent run.
2026-10-01/02: two more entrants inside 24h. (1) Cloudflare trained its own decision models — Clef and Clef-flash (blog, ~9:52 PM ET Oct 1, after the Oct 1 digests cut): an edge-network owner shipping the category as a routing play. (2) AWS open-sourced Strands Decider 2B (Oct 1, search feed): 2B params, weights + training recipe + evals — "fully open-source," one notch beyond Perplexity's weights-only release. Same window: WSJ-grade writeup on Jev "sparking copycats and talk of LLM alternatives"; Arav's "bring down the price even further over the coming days". Runner question is now seven-way (adds Strands, Clef) — still no independent run of any of them on the same workloads.
2026-10-03 (verified to the primary; record correction + status change): Clef is open-weights — blog (Oct 1): "we're fully open- sourcing these models on Hugging Face under an Apache 2.0 license" (our Oct 2 digests called Clef an entrant but missed the license; the Cloudflare challenger is the open model). And the no-runner status moves: Cloudflare published the category's first cross-vendor comparison table — vendor-run, but a real table: 9 sample rows from 43 eval benchmarks run total, scored on what it calls the "Jev Decision Index" with a live demo site. Clef leads most rows (BFCL 98.47, BANKING77 94.2, CLINC150 97.43); Jev wins When2Call (80.97 vs Clef 72.37), BRIGHT (47.52 vs 45.91) and agent-trace-observability (71.6); Laya leads PhishNChips (85.35). Latency: median Clef 209ms / Clef-flash 38.8ms / Jev 524ms; Laya 5.8ms ("very fast but trades off quality"); DiffusionGemma-Jev 84.4; Kev-9B 51.4. On TypeSafe's own eval suite Cloudflare claims 3-of-4 over Jev. Spec deltas: vision encoder + 64k context (Jev: text-only, 32k); Jev-API compatible (noul/choice/score, @cf/cloudflare/clef); no-read/no-train enterprise claim; fine-tuning = FDE service now, self-serve RL platform later; trained on DiffusionGemma adapted to output logprobs, building on independent research by Matt Mastrac[h?]. Three new entrants named in the table: DiffusionGemma Jev, Kev-9B, Laya (provenance unknown — not in our feeds; verify who ships them). Status change, precise wording: the category now has a runner TABLE, vendor-run; what still does not exist is an independent run on identical labeled workloads — the two vendors selling (Cloudflare, and TypeSafe whose index Cloudflare borrowed) grade each other's products. Eleven-way field if the table's three unknowns are real. Sizes (also from the same blog): the DiffusionGemma logprobs experiment was their earlier demo — production Clef is a frozen Qwen3.8-27B backbone (routing head + rank-256 LoRA, label- smoothed CE + Brier-loss calibration, synthetic permutation datasets, RLCD secondary objective), Clef-flash is frozen Qwen3.5-9B; prefill-only, non-autoregressive scoring. Same weight class as pplx-decider-v1-27b; Strands Decider is the 2B outlier.
2026-10-04 (usage demo, X layer): Arav — Pokémon FireRed's Elite Four + Champion cleared in one shot with Perplexity's Decisions API (~6:10 AM ET, rescued from filtered): claimed 592ms median / 987ms p95, $0.028 est. inference cost, 137 live calls — the runner race's first public consumer-grade demo with live latency/cost numbers (games, but a real deployment). The category still lacks an independent run on identical labeled workloads.
2026-10-05 (first third-party domain eval — and an entrant identified): Cisco's blog ran the domain comparison, in content safety: decision models (zero-shot) vs a 1B encoder fine-tuned on the Cisco AI Security Framework taxonomy (24 harm categories, six datasets, model-generated reference labels): the fine-tuned model wins — best AUC on 5/6, best recall at a 0.5% FP budget on all six (Cisco multi-turn 18k: 61% vs Jev 34%, Laya almost none); Jev "clearly the stronger" zero-shot; Laya had "no useful operating point within budget on any dataset." Two resolutions: Laya = open-weight, from ConvAI Innovations (the mystery entrant flagged Oct 3 in Cloudflare's table), and the category's pitch meets evidence — decision models are a labeling-free tier below purpose- trained classifiers, not a replacement. Caveats: Cisco trained on its own taxonomy, labels are LLM-generated — a domain benchmark, not the independent runner. Watch: TypeSafe/Cloudflare answers; workloads where zero-shot actually wins.
2026-10-06: the price floor halves — pplx-decider-v1.1-27b (Perplexity Developers ~1:25 PM ET + Arav ~4:46 PM ET): updated open-weights multimodal decision model, claimed top score on the new "Hugging Face Decision Index 0.3" (multimodalart/ jev-decision-index space — vendor-submitted), price halved to $0.02/M input ("Enjoy!"; v1 was $0.04). Arav's Oct 1 price-cut promise delivered day 5. The runner question gains a v1.1 cell; the still-unrun independent comparison grows another. The pplx-decider endpoint is now v1.1 at half price.
2026-10-07→09: three deltas. (1) TypeSafe closes $870M at $7.5B (SiliconANGLE Oct 9): the category's first mega-round, an order of magnitude past the eval vendor's round — capital now prices decision models ahead of the independent runner that still doesn't exist. (2) Liquid opens d1's weights — d1-3B + d1-omni-600M, multimodal, "zero output tokens" (HF blog Oct 7: "Multimodal open d1 decision models for the edge"; marktechpost syndication): the open-weights flanks gain a first-party lab entrant with weights. (3) OpenAI's Decisions API enters public beta, "10× faster typed answers" at $0.10/1M input (marktechpost Oct 9 — single syndicated source, unbenchmarked claim). Runner question unchanged: Jev vs AnyJev vs d1 vs Decisions API vs pplx-decider vs Clef — still no independent run.
2026-10-09→10: Microsoft ships Microsoft-Decision-1 (Nadella X, ~2:37 PM ET Oct 9; Register Oct 10): fast decision-making model, "top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality," in testing across Microsoft (incident response, quality control, scientific discovery); first version based on Qwen3.5-9B (the GLM-Flash-in-Codex rhyme — a US lab shipping the category on a Chinese open base), homegrown v2 promised; live in Foundry, "soon through OpenRouter." X layer from Perplexity the same window: decider-v1.1 on OpenRouter (multimodal, output free), DecisionBench #1 claims (94.5% acc/1,071 cases, $0.017/1k decisions, 534ms median — vendor-run), EVE inference-stack demos ("4× faster than JEV"; provenance unclear — new component, unvetted), a clinical claim (643/669), and pplx-embed-v2-late (9B/0.6B open multi-vector embeddings, shared text/image space). The distribution map is now two-way (Foundry/OpenRouter vs OpenRouter), the runner field has grown to Jev vs d1 vs OpenAI Decisions API vs pplx-decider vs Clef vs Decision-1 (plus AnyJev/Strands/Laya) — and the head-to-head on identical workloads still doesn't exist. Watch: Foundry metrics at second source, the OpenRouter slot, someone finally running the comparison.
2026-10-11 (marktechpost): vLLM Semantic Router ships Decision 3.0 — five open multimodal decision models, 0.8B–27B, d3 scoring 64.1 (https://www.marktechpost.com/2026/10/11/what-is-decision-3-0-vllm-semantic-routers- new-open-decision-models/). The category's open-weights tier appears, at sizes that make local deployment trivial. No shared benchmark vs Jev/pplx (the runner's gold-set binaries measure a different task shape). Watch: runner-format viability (the format:json lesson — Qwen3.5:9b failed 10/10 on response parsing), an independent 64.1 check, the 0.8B tier at the embedded decision-point role.
2026-10-13 (~1:00 PM ET, Crypto Briefing — from the ai digest): Artificial Analysis launches a Decision Index — new decision models test faster and cheaper than LLMs, less accurate on knowledge. The index layer arriving into the thread's standing gap (an independent run on identical labeled work beside the vendor tables — the runner's Oct 3-5 premise). Sourcing flag: topline-only in capture. Watch: the published rows vs TypeSafe/Cloudflare tables, and whether it becomes a runner-of-record read for the field. https://cryptobriefing.com/artificial-analysis-decision-index-decision-models/
2026-10-15 (~9:00 PM ET, Computerworld): the category's security side gets its first outside claim — a security vendor's report says TypeSafe's Jev is "even more susceptible to attack" than LLMs (https://www.computerworld.com/article/4236138/typesafes-jev-even-more-susceptible-to-attack-than-llms-says-security-vendor-3.html); analysts and consultants cited said the report appears accurate while stressing that no AI model should be — the capture truncates mid-caveat. Development: the storyline's "still nobody has run them" watch now has an adversarial claim on the table, but a vendor-graded one like the rest of the category's evidence. Watch: the vendor's methodology and named findings, TypeSafe's response, independent red-teaming of decision models.
2026-10-16 (~10:42 AM ET, Perplexity Developers): pplx-decider-v1.2 and pplx-decider-flash-v1.2 — the second generation of the open-weights decision family, text + vision. The thread's third family (TypeSafe's Jev, these, plus the Decision-Index field), and the first decision model with eyes. Watch: the model cards, whether v1.2 answers the index's accuracy finding, and runner uptake (the record's runner is the independent check). https://x.com/AravSrinivas
2026-10-16 (~2:42 + 4:32 PM ET — https://x.com/AravSrinivas/status/2111172252008059278, https://x.com/AravSrinivas/status/2111193532065337439): the v1.2 family's numbers and docs land — "Text and vision scores improve. Inference is up to 5x faster: all questions in a request run in one pass. Flash is the top 4B on the Decision Index" — and Arav points builders to the official quickstart (docs.perplexity.ai/docs/decisions/quickstart). The top-4B claim is scored on the Decision Index itself — the vendor citing the field's independent runner rather than a house benchmark. Watch: the published model cards vs the index rows, runner uptake.
2026-10-16 (distribution layer, evening — https://x.com/AravSrinivas/status/2111209381966782725; https://x.com/AravSrinivas/status/2111266278610743595): pplx-decider v1.2 lands two third-party distribution channels the same evening it launched — OpenRouter (~5:35 PM ET; listing confirms Decider v1.2 as a new 27B checkpoint succeeding v1.1, Flash v1.2 at 4B) and Vercel's AI Gateway (~9:21 PM ET; perplexity/pplx-decider-v1.2 for "complex agent workflows," pplx-decider-flash-v1.2 for "ultra-fast decisions," per Vercel's changelog). The launch's open-weights distribution thesis now has named channel placements; sizes remain vendor-stated pending published model cards, and the independent runner's check is unmet.
2026-10-16 (~3:46 PM ET — https://x.com/AravSrinivas/status/2111181956444827695; Clem RT on the Hugging Face weights ~4:41 PM ET — https://x.com/AravSrinivas/status/2111272339283620114): the pplx-decider line bumps to v1.2 — pplx-decider-flash-v1.2 (4B params, 1 cent per million tokens, top 4B on the HF Decision Index) and pplx-decider-v1.2 (27B params, 2 cents per million tokens, #2 on the HF Decision Index overall), open weights on Hugging Face and served on the Decisions API, multimodal, inference up to 5x faster via single-pass batching of all questions in a request. Still no independent run vs Jev, Clef, or OpenAI's Decisions API.
2026-10-17 (~4:10 PM ET — https://x.com/AravSrinivas/status/2111550342425141529): the category's third-party evaluation layer starts forming — Arav amplifies Wolfram Ravenwolf's testing of pplx-decider-flash-v1.2: matched "Wald's accuracy at nearly twice the speed on a 16 GB Mac mini" across "synthetic agent decisions, real email threads, prompt injections," and named Ravenwolf's new pick for email-completion checks; Arav's own line adds the on-device claim ("can run super fast and on-device on your M4 Mac 16GB"). The thread's first named third-party performance claim and its first on-device-on-consumer-hardware datapoint — but vendor-relayed, and "Wald" is a comparator of unverified identity in capture (not on the record's eleven-way field list). The standing watch partially moves: a real test suite exists, but the Jev/Clef/Decisions-API head-to-head still hasn't been run. Watch: Ravenwolf's published methodology and Wald's identity, whether the same harness runs Jev or Clef, and Apple-silicon vs RTX Spark on-device splits.
2026-10-14 (9:00 AM ET, Hackread, single outlet): FLock.io unveils an edge decision model built on Qwen — the writeup names Qwen3.5-2B and claims sub-31ms local inference for fast, low-cost offline classification with no cloud data egress (https://hackread.com/flock-io-edge-decision-model-qwen-data-privacy/). Development: the category gets an edge/on-device entrant — small-weight Qwen running locally pairs the decisions-not-text wedge with the on-device privacy frame (the same no-egress pitch Gemma × GovTech carries). Single-source; model sizing and the "System One" attribution unverified — read the primary before quoting. Watch: independent latency/quality comparison vs Jev's $0.042/M-input economics, whether FLock's model distributes through agent/harness plumbing.
2026-10-22 (~3:00 PM ET primary, OpenAI blog — https://openai.com/index/decisions-api; Traders Union ~4:00 PM ET — https://tradersunion.com/news/companies/show/3934684-openai-expands-decision-api-access/): OpenAI opens its Decisions API to all users — the incumbent lab's decision-model entry moves from the limited October 7–9 public-availability stage to general access, with speed and feature boosts and an explicitly competitive framing in the low-cost AI API market. Watch: general-access pricing vs Jev and pplx-decider tiers; whether the Decisions API enters the independent runner.
2026-10-22 (~8:00 PM ET — https://x.com/AravSrinivas/status/2113420151232352444): Perplexity ships Decider for Mac — pplx-decider-flash-v1.2, the 4B decision model released Oct 16, runs fully locally on macOS with no API key and no data leaving the machine: "Ask a question about text or an image and get a probability back, all on your machine" (app at pplx.ai/decider-for-mac). The category's local/on-device axis gets its first first-party desktop app, landing the same day OpenAI opened its Decisions API to all users (~3:00 PM ET) — the open-weights-local route versus the closed-API route now contrasted within one day. Distribution ledger for the decider line: first-party API, OpenRouter, Hugging Face weights, now an on-device app. Watch: real-world adoption of a desktop probability app; local decision models wired into agent routing; whether Jev/Clef/Decisions API answer with local modes; the still-open independent-eval gap.