Evidence Checked October 12, 2026
Landscape, October 2026
Decision models went from one vendor to a crowded field in about a month. Here's what changed since our benchmark runs in early October: the Claude lineup, OpenAI's Decisions API, and the open-weight models that, by their own published numbers, may no longer be mid.
Provenance. Our own measurements are on Four ways compared. Everything on this page comes from vendor documentation, release notes, public leaderboards and press as of October 12, 2026, and we haven't reproduced it. Vendor benchmark claims are marked as claims. The field changes daily.
The Claude lineup
Three new Claude models shipped in the last four weeks. All have a 1M-token context window and up to 128K output tokens.
| Model | ID | Released | Input / output per M tokens | Role in a cascade |
|---|---|---|---|---|
| Claude Fable 5.1 | claude-fable-5-1 | Sept 1, 2026 | $10 / $50 | The hardest escalations: long-horizon audits, novel failures |
| Claude Opus 5.5 | claude-opus-5-5 | Sept 22, 2026 | $4 / $20 | The default LLM tier: inspect, decide, fix the cause below. Cheaper than Opus 5 ($5 / $25) |
| Claude Sonnet 5.5 | claude-sonnet-5-5 | Sept 28, 2026 | $2 / $10 | A cheaper LLM tier for routine escalations and plan audits |
| Claude Haiku 5.5 | claude-haiku-5-5 | Oct 7, 2026 | $0.10 / $0.50 (≤100K-token prompts) | A judgment tier, or the watch-band reviewer. Anthropic positions it for classification, extraction and routing |
- Haiku 5.5 costs $0.50 / $2.50 for prompts over 100K tokens. The Batch API is 50% off (Haiku 5.5: $0.05 / $0.25).
- Anthropic has no decision or classification endpoint and no log-probabilities. The closest mechanism is structured outputs plus strict tool use, which is what our Haiku adapter used.
- Haiku 5.5 is a remarkable price drop: the previous Haiku 4.5 was $1 / $5, ten times more.
- Sources: Claude pricing, models overview, release notes.
Jev: unchanged
jev-1.13.0is still the only model;jev-latestandjev-previewboth point to it. No 1.14 yet.- Still $0.042 per million input tokens, output free; still text-only; 64k context per request.
- Rate limits are "adjusting dynamically" under demand, per TypeSafe.
- The Python SDK reached v0.7.3 (October 9) with concurrency and TLS fixes. Jev is also listed on OpenRouter as
typesafe/jev-1.13. - Source: docs.typesafe.ai/models.
OpenAI: Decisions API and Luna
- Decisions API has been in public beta since October 6, with GA promised "in the coming weeks." Still only
gpt-6-lunaand no pinned snapshot. Still $0.10 per million input tokens, with no charge for output or caching. It accepts images as inline base64. Rate limits aren't documented. - Luna is also a general model.
gpt-6-luna(Sept 22) is OpenAI's cheapest general text model: $0.10 in, $0.50 out per million, 1.05M context, on Chat Completions, Responses and Batch. Its predecessorgpt-5.6-lunais still listed at $0.20 / $1.20. - OpenAI has not adopted the
/v1/systemonerequest shape that others are converging on; you still need an adapter. - Sources: Decisions guide, gpt-6-luna, changelog.
Open weights: have they gotten better?
By the published numbers, yes, noticeably. In our early-October runs, the open options (APUS-OpenJev-4B, Strands Decider 2B) sat clearly below Jev. Since then:
| Model | Released | What's new | Reported result |
|---|---|---|---|
| Strands Decider, Gemma 4 family (E2B, E4B, 12B, 26B-A4B) | Oct 9 | Apache-2.0; also a new 2B reference model. Gemma 4 support is on GitHub main, not yet on PyPI | JevBench public: 180, 187, 200, 207 / 231 (Jev: 200). Median latency 102–245 ms on an H100 |
| APUS-OpenJev-v1 9B and 35B-A3B | ~Sept 23 | New sizes beside the 4B we tested; GGUF and MLX 4-/8-bit | Not yet on our harnesses |
| Cloudflare Clef (27B) and Clef-flash (9B) | Oct 1 | Apache-2.0; text, JSON, images and video in; also hosted on Workers AI | Vendor claims wins over Jev on several tasks (not independently replicated) |
| H2O-Lightning-4B | v1.2.3, Oct 10 | Apache-2.0, Qwen3.5-4B base, takes images | Vendor claims JevBench v1.6.1 composite 72.5 against Jev's 71.5 |
| Liquid AI d1-3B, d1-omni-600M | Oct 7 | Text and images (omni: audio, experimental) | Vendor claims 8 ms on an RTX 4090; made for edge devices |
What it takes to run them
The models that beat Jev on JevBench are big. Before you plan on "free," check what the hardware costs. A model's weights take roughly parameters × bytes per parameter: 2 bytes at BF16, 1 at 8-bit, 0.5 at 4-bit. Then add a few GB for the KV cache and the runtime. These are rule-of-thumb estimates, not measurements:
| Model class | 4-bit | 8-bit | BF16 | Realistic hardware |
|---|---|---|---|---|
| 2B–4B (Decider 2B, APUS 4B, H2O-Lightning-4B, Gemma 4 E2B/E4B) | 1–2 GB | 2–4 GB | 4–9 GB | Any recent laptop |
| 9B–12B (APUS 9B, Decider 12B) | 5–7 GB | 9–13 GB | 18–24 GB | A 16–24 GB laptop if quantized; BF16 wants a 24–32 GB GPU |
| 26B-A4B mixture-of-experts (Decider 26B) | ~14 GB | ~27 GB | ~52 GB | All 26B must sit in memory even though only ~4B work per token, so it's fast once it fits. A 24 GB GPU at 4-bit; a 48 GB+ machine at 8-bit; an 80 GB GPU at BF16 (Strands' latency figures are from an H100) |
| 31B dense (the top of the Capability board) | ~17 GB | ~33 GB | ~62 GB | Every parameter works on every token, so it's several times slower per request than a 4B. A 24 GB GPU or a 32 GB+ Mac squeezes 4-bit; a 48 GB card (L40S, A6000) or a 64 GB Mac for 8-bit; an 80 GB H100 for BF16 |
| 35B-A3B mixture-of-experts (APUS 35B) | ~19 GB | ~37 GB | ~70 GB | Like the 26B, but larger: a 32 GB+ machine at 4-bit |
- On a 24 GB Apple-silicon laptop (our test machine): the 2B–4B models run comfortably, and 9B–12B fit if a quantized build exists. The 26B-A4B at 4-bit is borderline, because macOS lets the GPU use only part of unified memory by default (roughly two-thirds to three-quarters). 31B dense and 35B-A3B don't realistically fit. Downloads are large too: budget the model's size in free disk, twice over while converting.
- Check the runtime, not just the size. Some decider servers run on PyTorch, which on a Mac means BF16 on the GPU unless someone publishes a quantized MLX or GGUF build. A 12B in BF16 (~24 GB) won't fit a 24 GB Mac.
- Quantizing changes the answers. 4-bit weights shift probabilities. Treat a quantized model as a new model: re-grade your questions and re-fit thresholds (they belong to a model, runtime and rubric triple).
- Rented GPUs have a price too. An 80 GB H100 rents for several dollars an hour. At our measured hosted price, Jev answers tens of thousands of typical requests for that. Open weights win on privacy and at high, steady volume, not automatically on cost.
The JevBench Capability board
JevBench v1.6.1's live board (October 12) lists Jev 1.13.0 as the unranked reference at Capability 77.1 (Calibration 90.6). The top open models:
| Rank | Model | Size | Capability |
|---|---|---|---|
| 1 | Quyet-1.0-Large | 31B | 81.7 |
| 2 | decisio v0.8.0 (Gemma 4 31B) | 31B | 79.6 |
| 3 | deck-31B | 31B | 77.6 |
| 4 | René-1 31B FP8 | 31B | 76.0 |
| 5 | H2O-Lightning-4B v1.1 | 4B | 75.0 |
| — | Jev 1.13.0 (reference) | — | 77.1 |
Source: benchmarkheaven.com/jev-models; harness at github.com/fstandhartinger/jevbench.
How to read this
JevBench is a public benchmark, and public sets leak into training. Our own lesson was that benchmark wins didn't predict mail wins: the 2B decider tied Jev on several JevBench families and still trailed it badly as a sorter feature source. A 31B model is also not a laptop model: see what it takes to run them. The honest next step is to run the Gemma 4 Decider 12B/26B and APUS 9B through the same wine, mail and page harnesses. The Decision Skeleton makes that a one-module job: swap the backend, re-run the eval.
Other entrants and the emerging standard
- Microsoft Decision-1 (Foundry, Oct 9): hosted only, text only, 32K context, priced like Jev at $0.042 per million input tokens. Microsoft claims best accuracy on 36 benchmarks.
/v1/systemoneis becoming a de facto API. Reports say AWS, Upstage and Ollama have aligned on TypeSafe's request shape. Ollama v0.35+ and llama.cpp's server can serve it locally, and Pydantic AI has aSystemOneModel. A decision model you can swap without an adapter is now realistic for everyone except OpenAI.- Other hosted decision APIs are appearing (Upstage Solar Decide, Perplexity, and others) from single-source announcements; treat them as unverified.
- Google, Meta, Mistral and Nvidia have no first-party decision model yet; Gemma 4 and Qwen 3.5/3.8 are the common bases for others' fine-tunes.
What this changes in our advice
- Write against the
/v1/systemoneshape and keep backends swappable. The cost of switching is now an eval run, not a rewrite. - Re-test open weights before calling them mid. If a 12B–26B open model matches Jev on your data, privacy and zero marginal cost become free wins.
- Haiku 5.5 changed the "small LLM" math. At $0.10 / $0.50, the price gap to decision models is now 5–7×, not 50×. Repeatability, latency and calibration, not price, are the reasons to prefer a decision model.
- Pin what you can. More vendors means more silent model updates. Thresholds belong to a (model, runtime, rubric) triple.