What Decider 1 is
Decider 1 is meraGPT’s System One model, announced on 22 September 2026. You send a state, which can be any JSON, and a set of typed questions. It answers every question in one call and generates no text. meraGPT, run by Singapore-registered Okyasoft Pte Ltd, serves it only through its own API, and the weights are proprietary. On 9 November 2026 meraGPT released Decider 1.1, which now answers every Decider request. Decider 1 is not related to Mapika’s decider, an open model with a similar name.
What it returns
The request schema is the same as Jev’s. A Noul returns the probability of yes. A Choice takes 2 to 255 labels and returns a probability per label plus the most likely one. A Score takes 2 to 32 levels and returns a probability per level and the expected level. One request holds up to 64 questions, and the docs say the state plus any one question must fit in 32,768 tokens. Since version 1.1 the state can include up to eight images, and each question reads the state on its own. Decider 1 took 10 labels, 10 levels and 4,096 tokens for the whole request, with text only.
What it is good at
meraGPT names three jobs: routing and triage, guardrails and checks inside agents, and scoring against a rubric such as severity or urgency. The first two match support inbox triage, intent and model routing and LLM guardrails.
Every quality figure comes from meraGPT’s own runs on typed-decisions, a synthetic benchmark published under the Hugging Face organisation LocalLLaMA. meraGPT’s submission to systemonemodels.org on 29 September 2026 says its team built the benchmark. The dataset card calls the benchmark independent and does not say who built it.
On the card’s 400 test cases and 2,000 decisions, from a single run, Decider 1 scores 0.768 accuracy, 0.096 KL, 0.052 Brier and 0.180 ECE, with a p50 of 526 ms end to end. Jev 1.13.0 scores 0.727, 1.442, 0.148 and 0.144 on the same cases. The card also lists Liquid AI’s d1 at 0.742 accuracy, measured on 30 September.
The card builds its gold answers from three samples of one teacher model, whose self-agreement is 0.735, and says “scores well above 0.735 mean a model is learning the teacher’s quirks.” Decider 1’s 0.768 is above that line. meraGPT’s own pages call the reference an ensemble of teacher models. The card puts Jev’s KL gap down to Jev placing nearly all its probability on one answer, against a soft three-sample gold. No third-party measurement had been published by 30 September 2026, and Decider 1 does not appear on Benchmark Heaven’s JevBench v1.5.4.
For Decider 1.1, meraGPT’s model page gives 0.746 accuracy and a KL of 0.116 on typed-decisions, against 0.726 and 1.471 for Jev 1.14 in the same run. Its 9 November post puts When2Call, a test of whether to call a tool, ask or refuse, at 77.9, up from 42.0 for Decider 1. These are meraGPT’s own runs too, and the dataset card has no Decider 1.1 row.
What it is not for
meraGPT rules out open-ended answers, facts the model would need to look up, locating things in an image or copying out its text, and general knowledge and multi-step reasoning, where it says larger models are well ahead. The docs warn that unrelated items batched into one state sway each other’s answers.
Access
Decider 1.1 is available now at POST https://meragpt.com/v1/systemone, as decider-1, which always serves the latest 1.x, or decider-1.1. It costs $0.04 per million input tokens in meraGPT’s docs, where the launch post gave $0.03, and output is free. Signing in adds $1 of free credit, then you buy prepaid credit from $10. The playground needs no account. TypeSafe’s Python and TypeScript SDKs work once TYPESAFE_BASE_URL is set to https://meragpt.com with a meraGPT key. It was not on OpenRouter on 12 November 2026.
Specifications
| Inputs | Text, imagesUp to 8 PNG, JPEG, WebP or GIF images per request, 8 MB each. |
| Question types | ChoiceScoreNoul |
| Max Choice options | 255 |
| Score levels | Up to 32 |
| Questions per call | 64 |
| Total context | 32,768 tokens |
| State budget | Not documented |
| Rate limit | No numeric rate limit is published. The docs say per-key concurrency limits apply and ask high-volume users to email first. A full fleet returns 429 capacity_saturated with a Retry-After of 5 seconds, or 30 seconds while extra capacity starts. The signed-out playground allows 10 runs a day per IP. |
| Endpoint | POST https://meragpt.com/v1/systemone |
| Licence | Not published as of 14 Nov 2026 |
| SDKs | Python: typesafe-sdkTypeScript: @typesafe-ai/sdk |
Decider 1.1 reads each question on its own over the state. The docs and the errors page say the state plus any one question can hold 32,768 tokens, or 16,384 in the playground. The decider-1 model page instead calls 32,768 tokens the maximum request for state and questions, and the errors page still describes input_too_long as going over a 4,096-token context. A request over the limit returns 400 input_too_long, which is not retryable. A Choice takes 2 to 255 labels and a Score 2 to 32 levels, and one request holds up to 64 questions. The state can hold up to 8 images (PNG, JPEG, WebP or GIF, 8 MB each) as base64 data URLs, each counting as up to 1,600 tokens; image links are not fetched. The playground page says it reads text only and that images go through the API. Decider 1 took 4,096 tokens for the state and all questions together, 10 labels, 10 levels and text only. Checked on 2026-11-12; the playground note on 2026-11-14.
Versions
- decider-1.1, 9 Nov 2026, Decider 1.1, now the model behind every Decider request. Code that calls decider-1 gets 1.1 with no change, and decider-1.1 names the version; the docs say decider-1 always serves the latest 1.x. It reads images, takes 32,768 tokens for the state plus any one question, a Choice of up to 255 labels and a Score of up to 32 levels, and answers each question on its own, so adding a question does not move the other answers. It takes an "I don't know" option when one is offered, and meraGPT says it picks that option more often than Decider 1, sometimes on questions it could have answered. Not on OpenRouter on 12 November 2026. Release notes
- state-decider-1, 22 Sep 2026, The launch version, with the alias sd-1. At launch meraGPT said model ids are never renamed and a change in behaviour ships as a new id. Its docs and /v1/models catalogue now list the model as decider-1, which serves Decider 1.1, and this model page URL redirects to /models/decider-1. meraGPT reserved meragpt/state-decider-1 for a future OpenRouter listing, but OpenRouter did not list it on 30 September or 12 November 2026. Release notes
Use cases
What people use Decider 1 for, one page per pattern.
Support inbox triage with System One models
Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket.
Intent and model routing with System One models
One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue.
LLM guardrails with System One models
Put one Jev request in front of an LLM and one behind it. Yes/no questions return the probability that each hazard holds, a Score rates how much harm complying would do, and your thresholds turn those numbers into pass, review, block, or a crisis path.
Examples built with Decider 1
The most-starred and most-viewed entries in the directory. All Decider 1 examples.

Building a custom agent harness with Pi and Decider 1
meraGPT's own post, following Elvis Saravia's "Building a Custom Harness with Pi and Jev". A Pi agent harness of about 130 lines of JavaScript calls Decider 1 through TypeSafe's JS SDK to pick the model, block shell commands that delete data or use the network, and send unfinished answers back.
meraGPT Decider 1 cookbooks
meraGPT's official cookbook for its Decider 1 model: seven recipes, each with a runnable curl, Python and TypeScript request and the unedited output. They cover support triage, confidence gating, phishing and output guardrails, RAG passage filtering, picking an agent's first tool and asking many questions in one call.
Benchmarks
Published scores, each linked to its source. A row marked vendor-run was measured by the vendor itself, so read it as one side's result. The compare hub sets these beside the other models.
| Benchmark | Metric | Score | Run by | Date |
|---|---|---|---|---|
| typed-decisions | Accuracy | 0.746 | vendor-run | 9 Nov 2026 |
| typed-decisions | Accuracy | 0.768 | vendor-run | 22 Sep 2026 |
| typed-decisions | ECE | 0.18 ECE | vendor-run | 22 Sep 2026 |
- typed-decisions, Accuracy: Decider 1.1, from meraGPT's own run on its model page; Jev 1.14 scored 0.726 on the same cases. KL from the reference 0.116 against Jev 1.14's 1.471. meraGPT's team built the benchmark, and the dataset card had no Decider 1.1 row on 2026-11-12.
- typed-decisions, Accuracy: Decider 1 (state-decider-1). 400 test cases, 2,000 decisions, one run by meraGPT; Jev 1.13.0 scored 0.727. meraGPT says its team built the benchmark.
- typed-decisions, ECE: Decider 1 (state-decider-1), same run; Jev 1.13.0 scored 0.144. Lower is better. meraGPT publishes no ECE for Decider 1.1.