Blog

Decision Models: Who Launched What, and What to Ask

Jev, Clef, Strands Decider, OpenAI’s Decisions API, Microsoft-Decision-1: decision models went from one product to dozens in less than a month. Here is who launched what, the three approaches behind them, and the questions worth asking before you pick one.

Updated decision-models ai 7 min read

By David Gorges Co-authored with AI AI helped with research, structure and first drafts. David Gorges checked every fact, reworked the text and stands behind every sentence.

Managing Director

In Decision Models in Symfony AI, we wired Jev into a Symfony app and put the threshold in our own code. This post is its companion: who launched what in the first weeks, the three approaches that came out of the rush, and the questions worth asking before you choose a provider.

The field moves weekly, so we keep this post up to date. The header tells you when we last touched it, and the sources at the end show what we checked.

Twenty-five days, from Jev to Microsoft

Here is what happened between Jev’s launch on September 15 and October 9, 2026:

  1. September 15

    TypeSafe AI launches Jev

    TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, opens Jev in early access, alongside a $40 million seed round led by DCVC at a $200 million valuation. heise reports on the launch the same day.

  2. September 19–25

    Symfony AI gets a TypeSafe bridge

    Yoan Bernabeu submits the TypeSafe bridge for Symfony AI. Christopher Hertel merges it on September 21, and version 0.14.0 ships it on September 25. Pydantic AI, LangChain, Windmill, and ToolJet add their own Jev integrations around the same time.

  3. September 20–22

    JevBench and the first open models

    The community launches JevBench, which had reached 1,624 decisions and 106 systems in the snapshot we used. Laya, from Convai Innovations under Apache 2.0, and Kev 4B, Jared Palmer’s Qwen3.5-4B-based model, arrive with Jev-compatible APIs.

  4. September 25

    Kev against Jev, and PostHog’s Jeeves

    Opper tests Kev against Jev on 362 fresh tasks. Accuracy is within two percentage points, but calibration is considerably weaker. PostHog releases Jeeves, a 9B model that can optionally reason before deciding. It outperforms Jev on JevBench, 0.935 versus 0.866, with a reported median latency of 3.3 seconds on a 325-question development set.

  5. September 26

    An unchanged LLM as a decision model

    Privatemode uses prefilling and log probabilities to turn an unchanged GLM-5.3-Flash into a decision model. It reports results on par with Jev across 28 datasets, with image understanding as an extra capability.

  6. September 28

    DiffusionGemma on vLLM

    Red Hat demonstrates Google’s DiffusionGemma 26B-A4B as a decision model on vLLM. A single denoising step fills the answer fields, with reported throughput above 300 decisions per second.

  7. September 29

    OpenAI’s Decisions API and Liquid’s d1

    OpenAI announces its Decisions API at DevDay: a constrained GPT-6 Luna variant with text and image input. By October 7, the documentation lists it in public beta: POST /v1/decisions with predicate, choice, and score questions, $0.10 per million input tokens. Its score is a probability-weighted position on a scale, not a confidence value. Liquid AI launches d1 the same day: Jev-style primitives and TypeSafe-compatible SDKs as a hosted service, free during preview.

  8. October 1

    Cloudflare releases Clef

    Cloudflare publishes Clef and Clef-flash as open weights under Apache 2.0 and hosts both on Workers AI. Clef sits on a frozen Qwen3.8-27B backbone, Clef-flash on Qwen3.5-9B. Cloudflare calls them fully Jev-API compatible and reports a median latency of 38.8 milliseconds for Clef-flash. In its own benchmarks, Clef beats Jev on 8 of 10 tasks.

  9. October 1

    AWS open-sources Strands Decider

    AWS releases Strands Decider 2B as open source, including training data and scripts. It replaces the language head of a Qwen3.5-2B backbone with a pointer head that scores the supplied options. Reported median latency is about 115 milliseconds on an RTX 3090 and about 153 milliseconds for small tasks on an M3 MacBook. TechCrunch headlines: “Amazon releases its own Jev clone as decision models flood the web.”

  10. October 9

    TypeSafe raises $870 million, Microsoft enters

    TypeSafe announces an $870 million Series A led by Andreessen Horowitz at a $7.5 billion valuation, with Sequoia and DCVC, 24 days after the launch; Jev is still in early access. The same day, Microsoft launches Microsoft-Decision-1 in Foundry, built on Qwen3.5-9B and priced like Jev at $0.042 per million input tokens, output free. Microsoft reports 83.5 percent mean accuracy across 36 benchmarks it selected itself.

Three approaches

Three approaches come out of that rush:

  • Purpose-built hosted models: Jev, d1, OpenAI’s Decisions API, and Microsoft-Decision-1. Bounded decisions with no setup, but your data goes to a provider, and billing, availability, and calibration differ from one to the next.
  • Open-weight models with Jev-compatible APIs: Strands Decider, Clef, Kev, Laya, Jeeves, H2O-Lightning-4B, and dozens more, mostly a Qwen or ModernBERT backbone with a pointer head. You run them yourself; Ollaya packages them in an Ollama-like experience.
  • An LLM you already have: prefill the answer and read the log probabilities. Slower and pricier per call, but multimodal, and there is no new model to operate.

The interface converged faster than the models did. Jev compatibility is everyone’s selling point, and that’s good news for you: your choice of provider stays reversible.

Questions worth asking

Do you need a separate model? Your language model is already a decision model gets competitive decisions out of an unchanged Qwen3.5-9B by rotating the option letters and summing their log probabilities. It reports better calibration than Jev, with a mean ECE of 0.055 versus 0.099. Opper’s Kev test found the opposite pattern: same accuracy, weaker calibration. So measure calibration on its own; an accuracy leaderboard won’t tell you.

What does a decision cost? Jev and Microsoft-Decision-1 charge $0.042 per million input tokens, OpenAI $0.10, all with free output. But Opper counted about 257 fixed input tokens on every Jev request, which made Kev up to twelve times cheaper for short states. Compare the cost per useful decision, including overhead, retries, and the human review behind it.

Do the benchmarks settle it? Not yet. Most numbers come from vendors: Cloudflare has Clef ahead of Jev on 8 of 10 tasks, Microsoft reports 83.5 percent on benchmarks it picked, and on JevBench the open H2O-Lightning-4B scores 72.5 against hosted Jev’s 71.5. User reports contradict each other: in the Decisions API thread, one commenter’s results favored Jev and an OpenAI engineer reported different numbers; in the Series A thread, one commenter had Clef beating Jev while another found OpenAI “not close.” Keep the inputs, question wording, model version, and provider route with your results, and compare on the same labeled cases.

What about German? Laya lists seven languages with Latin scripts. We found no comparable coverage details, and no German benchmark results, for Jev or d1. If your tickets are in German, so is your evaluation set.

Who will maintain it? Diogo Almeida told TechCrunch the wave looked more like people implementing a cool architecture than building lasting products, and some of the 106 JevBench entries are weekend projects. For customer systems, ask who will support the model a year from now, or pick one you can maintain yourself.

What happens in a real loop? SREGym gave Codex a Jev second opinion on ten SRE problems, five runs each: passes rose from 20 to 24 out of 50, with four problems better, two worse, and four unchanged. That’s a small sample, and a reminder that ranking can’t find a cause nobody proposed. Test postconditions and false approvals separately.

Ready to try one? Decision Models in Symfony AI walks through the code: the TypeSafe bridge, a threshold in a Messenger handler, and a Workflow guard before a tool call. For new posts, follow the blog’s RSS feed.

Sources

28 sources
  1. Series A

    TypeSafe · October 9, 2026

  2. Microsoft leans on open weight model from Chinese AI lab to challenge Jev

    The Register · October 10, 2026

  3. Introducing Clef: our open-source decision models

    Cloudflare · October 1, 2026 · vendor benchmarks

  4. Introducing Strands Decider

    Strands Labs (AWS) · October 1, 2026

  5. Amazon releases its own Jev clone as decision models flood the web

    TechCrunch · October 1, 2026

  6. Jev (AI model)

    Wikipedia

  7. Jev aims to help machines make faster decisions

    heise · German

  8. [AI Bundle][Platform] Add TypeSafe bridge

    symfony/ai PR #2549

  9. OpenAI DevDay 2026 recap

    OpenAI

  10. OpenAI Decisions API

    OpenAI · official guide, checked October 7, 2026

  11. OpenAI DevDay 2026: every announcement

    dev.to

  12. AlphaSignal on the Decisions API

    AlphaSignal

  13. Decision models

    Liquid AI

  14. MarkTechPost on d1

    MarkTechPost · September 29, 2026

  15. PostHog Jeeves

    GitHub

  16. Jev vs. Kev

    Opper · September 25, 2026

  17. Laya

    brainfunctioncollapse.com

  18. Jev, Laya, and the emerging role of decision models

    Runware

  19. System One from GLM-5.3-Flash

    Privatemode · September 26, 2026

  20. Your language model is already a decision model

    GitHub

  21. Run decision models on vLLM with DiffusionGemma

    Red Hat · September 28, 2026

  22. JevBench v1.5.4

    Benchmark Heaven

  23. H2O-Lightning-4B on JevBench

    Benchmark Heaven · checked October 10, 2026

  24. Decisions API discussion

    Hacker News · conflicting early user reports

  25. Typesafe AI raises $870M at $7.5B

    Hacker News · anecdotal user reports

  26. “decision model,” past month

    Hacker News

  27. Jev-Driven SRE Diagnosis: What Worked and What Failed

    SREGym · September 17, 2026

  28. Optional Jev decision support

    SREGym on GitHub · integration documentation

Updated 10 October 2026.

Contents

Read next

ai decision-models symfony

Decision Models in Symfony AI: Put Confidence to Work

Which team should handle this ticket? Does this tool call need a human review? Decision models turn questions like these into structured answers with probabilities. Let’s connect Jev to Symfony AI and put the threshold where we can test it: in our application code.

php security symfony

Envelope Encryption in Symfony 8.2: Lock the Key, Not Just the Data

Symfony 8.2 encrypts a Doctrine column with one attribute. The trick behind it is called envelope encryption: lock the data with one key, and lock that key with a second one that never leaves your key service. Here's how it works, with a playground to try it.