---
title: We tested eight new decision models. None got more right than the one we use. | Plainlamp
description: Since our test of 27 decision models, OpenRouter has listed nine new ones. Eight can answer our question, among them the next version of the model our pipeline uses and one from Google. We read them on the same 339 unpublished descriptions. No new model got more right than the one we use; four were behind by a gap too big for chance, and the test cannot tell the other four apart from it.
url: https://plainlamp.com/blog/new-decision-models
---

13 November 2026, 21 min read

# We tested eight new decision models. None got more right than the one we use.

_Updated on 14 November 2026 with the eighth model, Decisions Fast, and with Decider Flash V1.3's reads of the first test's 38 descriptions._

Plainlamp turns a hard document, such as a paper, a policy or a standard, into a short narrated video lesson and a guide. A small AI model helps at the start. When a person describes their document in their own words, it sorts the words into a kind: paper, policy, standard, rules, or none. Such a model is called a decision model: it answers a multiple-choice question about a piece of text and says how sure it is.

On 2 November we wrote about [testing 27 decision models](https://plainlamp.com/blog/decision-models) on that question. On 4 November we added a second test, on 339 new descriptions.

Since then, OpenRouter, the service we call these models through, has listed nine decision models we had not tested. Eight of them can answer our question. One is the next version of the model our pipeline uses. So we tested again: the same question, the same 0.6 line, and the same 339 descriptions, which we have never published.

## The new models

| Model (maker) | Listed on OpenRouter | What it is | Listed price, US$ per million input tokens | An endpoint with zero data retention |
| --- | --- | --- | --- | --- |
| Decider V1.3 (Perplexity) | 12 November | the next version of the model our pipeline uses | free endpoint only | yes |
| Decider Flash V1.3 (Perplexity) | 12 November | its smaller sibling | free endpoint only | yes |
| Gemini Decision 1 (Google) | 11 November | a decision model from Google | 0.10 | none listed |
| HeyGen Decision 1 (HeyGen) | 13 November | a decision model from HeyGen | 0.14 | none listed |
| HeyGen Decision 1 Turbo (HeyGen) | 13 November | its lower-cost variant | 0.09 | none listed |
| Drex Max Finance (Nace.AI) | 10 November | by its listing, a version of Drex Max tuned for finance | 0.06 | yes |
| Reflex (Morph) | 3 November | a decision model from Morph | 0.042 | yes |
| Decisions Fast (Parallel) | 6 November | by its listing, a decision model that researches the live web | free endpoint only (from 14 November: US$0.001 a call) | none listed |

- **How we found them:** OpenRouter's list of decision models on 13 November 2026 had 41 entries. We left out the five aliases, one free copy of a model that is on the list twice, and every model we had tested before. Nine new models were left. Eight of them answer on the route our test uses (a route is the web address a kind of call goes to).
- **Free endpoints:** an endpoint is one provider's offer of a model, with its own price and data rules. On 13 November Decider V1.3, Decider Flash V1.3 and Decisions Fast were listed only at a price of zero (a "free endpoint"), with no paid one. Free endpoints have a limit per minute and a limit per day. So calls to a free endpoint started at least 3.1 seconds apart. Decider V1.3 and Decider Flash V1.3 read the 339 descriptions on 13 November; Decider Flash V1.3's reads of the first test's 38 waited for the next day's allowance, and so did Decisions Fast.
- **A change on 14 November:** OpenRouter no longer listed Decisions Fast as a free endpoint. It listed it at US$0.001 a call, under the same dated version. We read it that day, at that price.
- **A finance model:** OpenRouter's listing describes Drex Max Finance as a version of Drex Max tuned for finance, for documents such as invoices and statements. Our question is not about finance. We tested it because it is a new decision model on the list, and because Drex Max was one of the four that our second test could not tell apart from the model we use.
- **One we could not test:** Clairic, from Cloudflare, is on the list, but OpenRouter offers it only on its Extractions route. Our test calls the Decisions route, which answered "No endpoints found" for it.
- **One no longer on OpenRouter:** Mercury Decide, from Inception, was in both tests of our first post. On 13 November it was no longer on OpenRouter's list, and a call to it got "No endpoints found".

## The question and the line

The model gets the person's words and five options, each with a short definition: paper, policy, standard, rules, or none. It gives one answer and says how sure it is, from 0 to 1. We call that number its confidence. A read is one call to the model with one description.

Our pipeline takes an answer at 0.6 or more. On "none", or under 0.6, it does not take the model's answer, and it records no kind from the person's words. In this post, as in the first, we call that a stop. For each read, we ask what our pipeline would do with the answer:

- **right:** it takes the right kind, or it stops when no kind fits;
- **wrong kind taken:** it takes a wrong kind. This is the harm: a wrong kind is recorded for the document;
- **stopped, though a kind fit:** the right kind exists, but the model said "none" or was under 0.6. The answer is not taken when it could have been, but no wrong kind is recorded.

A stop is right when no kind fits, and a miss when one does. When we count a model's stops below, we mean the misses: stops where a kind fit.

The full question and its five definitions are in [our first post](https://plainlamp.com/blog/decision-models).

## The results on unpublished words

![Table of Decider V1.2, the model we use, and eight new decision models on 339 unpublished descriptions. Right of 339, wrong kind taken, stopped though a kind fit, failed reads, and the verdict against Decider V1.2: Gemini Decision 1 332, 0, 7, 0, could be chance; Decider V1.2 336, 1, 2, 0; Decider V1.3 336, 1, 2, 0, could be chance; Reflex 330, 1, 8, 0, could be chance; Drex Max Finance 334, 1, 3, 1, could be chance; Decider Flash V1.3 323, 6, 10, 0, behind; HeyGen Decision 1 313, 13, 5, 8, behind; HeyGen Decision 1 Turbo 308, 13, 5, 13, behind; Decisions Fast 311, 15, 13, 0, behind.](https://plainlamp.com/blog/new-decision-models/decision-update-1-scores-1080.webp)

We read the new models on the 339 descriptions of our second test. We wrote them for that test and have not published the set. Our first post quotes three of them with their answers; the rest are in no public place. The rule for each read is the one above.

- **Fixed first:** before any new model read the 339, we wrote down in our own code (our repository is private) the eight models, the rule and the test. Only the verdicts of that test were fixed first; the other counts describe the result.
- **The test:** it compares each new model with Decider V1.2. It counts only the descriptions that one of the two got right and the other did not (an exact sign test). We call a gap too big for chance when chance alone gives a split at least as uneven less than 1 time in 20. Then we allow for making eight comparisons at once (Holm's method), which makes a gap harder to call too big.
- **The other side of each comparison** is Decider V1.2's reads of 4 November, from the second test.
- **The reads:** one read of each description on 13 November 2026 (Decisions Fast on 14 November), and a second read of 50 of them.

The order of the table is our first post's: fewest wrong kinds taken first, then fewest failed reads, then the most right. So a model with fewer wrong kinds stands above a model with more right answers.

| Model (maker) | Right (of 339) | Wrong kind taken | Stopped, though a kind fit | Failed reads | Only Decider V1.2 right | Only this model right | Against Decider V1.2 |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Gemini Decision 1 (Google) | 332 | 0 | 7 | 0 | 7 | 3 | could be chance |
| Decider V1.2 (Perplexity), the model we use | 336 | 1 | 2 | 0 | - | - | - |
| Decider V1.3 (Perplexity) | 336 | 1 | 2 | 0 | 1 | 1 | could be chance |
| Reflex (Morph) | 330 | 1 | 8 | 0 | 8 | 2 | could be chance |
| Drex Max Finance (Nace.AI) | 334 | 1 | 3 | 1 | 2 | 0 | could be chance |
| Decider Flash V1.3 (Perplexity) | 323 | 6 | 10 | 0 | 14 | 1 | behind, too big for chance |
| HeyGen Decision 1 (HeyGen) | 313 | 13 | 5 | 8 | 24 | 1 | behind, too big for chance |
| HeyGen Decision 1 Turbo (HeyGen) | 308 | 13 | 5 | 13 | 29 | 1 | behind, too big for chance |
| Decisions Fast (Parallel) | 311 | 15 | 13 | 0 | 27 | 2 | behind, too big for chance |

- **Too big for chance, even allowing for eight comparisons (Holm's method):** Decider Flash V1.3, HeyGen Decision 1, HeyGen Decision 1 Turbo and Decisions Fast are behind Decider V1.2: 14 descriptions to 1, 24 descriptions to 1, 29 descriptions to 1 and 27 descriptions to 2.
- **HeyGen's two models** each took a wrong kind 13 times, and both read all seven insurance contracts as "policy", like the trap of our first test. One of the 13 is IFRS 9 read as "rules", which you could argue for.
- **Decider Flash V1.3,** the smaller sibling of Decider V1.3, took a wrong kind 6 times; three of them were insurance contracts read as "policy". Of the 14 descriptions only Decider V1.2 got right, nine were stops and five wrong kinds.
- **Decisions Fast,** from Parallel, took a wrong kind 15 times, the most of the eight, and stopped 13 times though a kind fit. It took a wrong kind on five of the seven insurance contracts. Of the 27 descriptions only Decider V1.2 got right, 13 were stops and 14 wrong kinds. It was right on two that Decider V1.2 missed. By its listing it researches the live web, and every one of its reads came back with web citations. For this model we give counts only, not its answers to single descriptions.
- **HeyGen's failed reads:** 8 reads of HeyGen Decision 1 and 13 of HeyGen Decision 1 Turbo failed. HeyGen refused each for rate (its own "rate limit exceeded"), on the second try too. Our code called HeyGen's two models at the same time and did not pace paid endpoints: 95 of their calls needed a second try, and 21 failed on that too. So these failures are about how fast we called, not answers of the models. Our rule also gives a free endpoint more time after such a refusal (30 seconds, up to six times) than a paid one (2 seconds, once). By the rule we fixed before the runs, a failed read counts as not right. Outside that plan, when we saw the failures, we read those 21 descriptions again, one model at a time: all answered (20 right, 1 stopped, 0 wrong). With those reads the gaps would be 16 to 1 and 17 to 1, still too big for chance.
- **Could be chance:** Gemini Decision 1 (7 to 3), Decider V1.3 (1 to 1), Reflex (8 to 2) and Drex Max Finance (2 to 0). The test cannot tell these four apart from Decider V1.2.
- **No model got more right than Decider V1.2,** at our 0.6 line. Decider V1.3 got the same 336. At other lines it would differ (see the limits).
- **Decider V1.3 against V1.2:** the same counts: 336 right, one wrong kind, two stops. Their results differ on two descriptions. Both times the two versions chose the same, right kind, one above our 0.6 line and one below. On "la méthode EBIOS Risk Manager pour l'analyse des risques cyber", V1.3 was 0.81 sure and V1.2 only 0.30. On a description of an internal document that an IT department calls a "standard" (the right kind is "policy"), V1.2 was 0.69 sure and V1.3 0.57. Both took "policy" for the privacy notice on a shopping website.
- **At a higher line, V1.3 would be ahead.** Where both versions chose the right kind, V1.3 gave the higher confidence on 171 descriptions and the lower on 102. At a line of 0.8, V1.3 would be right on 10 descriptions that V1.2 misses, and V1.2 on none that V1.3 misses. These are checks we made after the result, and our pipeline uses 0.6.
- **Gemini Decision 1 never took a wrong kind.** It was right on all three descriptions Decider V1.2 missed. Its seven misses were all stops: four descriptions of IFRS standards, three of which say a law requires the standard, each chosen as "standard" at 0.55 to 0.59, just under our line, and three studies, such as a PhD thesis and a master's dissertation.
- **Gemini's confidence is on another scale.** Each model gives a probability for the answer it chose (its top probability) and a confidence. Every other model here works its confidence out from the top probability: (top probability − 0.2) / 0.8. So for them, 0.6 means a probability of 0.68. Gemini gives the top probability itself as its confidence, so for Gemini our 0.6 line is a lower bar. We used the confidence each model gives, as we fixed before the runs.
- **Gemini with the others' formula:** it would be right on 327 (12 stops, still no wrong kind). It would be behind Decider V1.2 on 11 descriptions and ahead on 2. On its own, that gap would be too big for chance (about 1 time in 45), but not once we allow for the eight comparisons. We planned this count before the runs as a description; it is not one of the fixed verdicts.
- **Drex Max Finance got as many right as Drex Max** did on 4 November, and as many wrong kinds: 334 right and one wrong kind. Their results differ on four descriptions, two each way (a comparison we made after the result). One read of Drex Max Finance failed with a server error, which our code does not try again; read again later, outside the plan, it was right. Its host, DeepInfra, answered slowly that day: a median of 6.5 seconds a read on the 339 descriptions.
- **Reflex** was behind on 8 descriptions and ahead on 2. Seven of the eight were stops, which record no wrong kind. Its one wrong kind: it took "rules" for a job advert for a compliance analyst.
- **Wrong kinds, the harm:** Gemini Decision 1 took none; Decider V1.3, Reflex and Drex Max Finance took one each; Decider Flash V1.3 took 6; HeyGen Decision 1 and HeyGen Decision 1 Turbo took 13 each; Decisions Fast took 15. Decider V1.2 took one.
- **The privacy notice and the insurance contracts:** Decider V1.2's one wrong kind was "policy" for the privacy notice on a shopping website. Decider V1.3, Decider Flash V1.3, HeyGen Decision 1, HeyGen Decision 1 Turbo and Drex Max Finance took "policy" for it too; Gemini Decision 1 and Reflex said "none". Decider V1.3, Gemini Decision 1, Drex Max Finance and Reflex were right on all seven insurance contracts, as Decider V1.2 was.
- **The hard descriptions decide, again:** on the 192 descriptions not marked hard, allowing for the eight comparisons, the test cannot tell any new model apart from Decider V1.2 (a check we made after the result).
- **Steady:** on the 50 descriptions read twice, each model's result was the same both times, except Decider Flash V1.3 (1 changed) and Decisions Fast (1 changed), leaving aside the five whose first read had failed (two for HeyGen Decision 1 and three for HeyGen Decision 1 Turbo).
- **Decider V1.2, read again:** the comparison uses Decider V1.2's reads of 4 November. On 13 November we read it once more, the same dated version: the same result on all 339 descriptions (on one, its answer changed, both times under our line, so our pipeline stops either way). Its confidence was not the same number on 146 of them; the largest move was 0.07.
- **Our own test:** AI writers (Claude, from Anthropic) wrote these descriptions to our brief, and AI readers, also Claude, set the answers. Nobody drew them from what real people type. We wrote the set after we had chosen Decider V1.2, and our brief asked for hard cases like the trap of the first test, which Decider V1.2 had passed. On the hardest ones, the right answer follows our own rules, which the models never saw.

## The first test's 38 descriptions, again

![The words “My car insurance policy document.” and how seven of the new decision models read them. 3 said “none” in every read: Decider V1.3, Reflex and Gemini Decision 1. 2 said “policy”, never sure enough to take: Decider Flash V1.3 and Drex Max Finance. 2 said “policy” sure enough to take, in every read: HeyGen Decision 1 Turbo and HeyGen Decision 1.](https://plainlamp.com/blog/new-decision-models/decision-update-2-trap-1080.webp)

We also read the new models on the 38 descriptions of our first test, as on 1 November: once, three more times, then once with zero data retention required. For Gemini Decision 1, HeyGen Decision 1 Turbo, HeyGen Decision 1 and Decisions Fast, OpenRouter had no endpoint with zero data retention and refused that last read, so they have four reads of each description: 152 in all. These 38 have been on our site, with their answers, since 2 November. So for models made later this is not a blind test. We show these reads for the trap and for steadiness; no verdict rests on them.

| Model (maker) | Right | Wrong kind taken | Stopped, though a kind fit | Failed reads | With zero data retention required | Median time | US$ per million calls |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Decider Flash V1.3 (Perplexity) | 190 | 0 | 0 | 0 | answered | 205 ms | free |
| Decider V1.3 (Perplexity) | 190 | 0 | 0 | 0 | answered | 300 ms | free |
| Drex Max Finance (Nace.AI) | 190 | 0 | 0 | 0 | answered | 6794 ms | 12.9 |
| Reflex (Morph) | 185 | 0 | 5 | 0 | answered | 323 ms | 9.4 |
| Gemini Decision 1 (Google) | 148 of 152 | 0 | 4 | 0 | no endpoint | 256 ms | 22.3 |
| HeyGen Decision 1 Turbo (HeyGen) | 148 of 152 | 4 | 0 | 0 | no endpoint | 271 ms | 25.1 |
| HeyGen Decision 1 (HeyGen) | 148 of 152 | 4 | 0 | 0 | no endpoint | 294 ms | 34.6 |
| Decisions Fast (Parallel) | 144 of 152 | 4 | 4 | 0 | no endpoint | 646 ms | 1000.0 |

- **The trap,** "My car insurance policy document.": Decider V1.3, Reflex and Gemini Decision 1 said "none" in every read; Decider Flash V1.3 and Drex Max Finance said "policy", never sure enough to take it; HeyGen Decision 1 Turbo and HeyGen Decision 1 said "policy" sure enough for our pipeline to take it in every read.
- **The thesis,** "a PhD thesis on supply chain resilience": Decider Flash V1.3, Decider V1.3, Drex Max Finance, HeyGen Decision 1 Turbo and HeyGen Decision 1 took it as a paper in every read; the other two said "none" in every read.
- **Right in every read:** Decider Flash V1.3, Decider V1.3 and Drex Max Finance, as Decider V1.2 was on 1 November.
- **Steady:** every model chose the same answer for a description in all its reads.
- **Decisions Fast** is in the table only: for this model we give counts, not its answers to single descriptions.

## What we chose

**We will move to Decider V1.3,** the next version of the model we use. The test gives no reason against it: it got the same counts as Decider V1.2, and at a higher line it would be ahead. The test gives no proof that it is better either: at our 0.6 line the two differ by one description each way. We move because it is the newer version.

Until that change is made, our pipeline uses Decider V1.2, in the dated version we chose, with zero data retention required on every call. We will keep both rules for V1.3: a dated version, and zero data retention. On 13 November OpenRouter listed V1.3 only as a free endpoint, which has a daily limit on calls.

Gemini Decision 1 is the one new model that never took a wrong kind, which is the harm we care most about. On that day OpenRouter listed no endpoint for it with zero data retention, and our pipeline requires one.

## What this test does not tell you

- The new models read each of the 339 descriptions once, and 50 of them twice. A read near our line can go either way.
- The 0.6 line decides this test too. At 0.7, Decider V1.3 and Drex Max Finance would get more right than Decider V1.2; at 0.8, Decider V1.3 and Reflex would (counts we made after the result).
- The new models were read on 13 November (Decisions Fast on 14 November), and Decider V1.2's side of each comparison is nine or ten days older. Read again on 13 November, it gave the same result on every description.
- A model's confidence is its own. Gemini's is on another scale than the others' (above), and a line that fits one model may not fit the next.
- A failed read counts as not right. HeyGen's failed reads came from how fast we called, not from the models (above).
- The first test's 38 descriptions are public, so the new models' reads of them are not a blind test.
- It is one question. A model that is best here may not be best on yours.
- OpenRouter's Decisions route is marked alpha: an early version that may still change. Models and prices change too. These are the ones of 13 November 2026 (Decisions Fast: 14 November).
- Times are from one computer in the Netherlands.

## How to repeat this test

- **The call, the question, the options and the line** are those of our first post: POST https://openrouter.ai/api/alpha/decisions, one question under the key kind, no fallback (allow_fallbacks: false) and data_collection: "deny". An answer is taken at confidence 0.6 or more, on the confidence the route gives.
- **On the 339 descriptions,** every call also carried ignore: ["primeintellect"], as in our second test. Zero data retention (zdr: true) was required of the models on OpenRouter's zero-retention list that day: Decider V1.3, Decider Flash V1.3, Drex Max Finance and Reflex, and Decider V1.2. It was not asked of Gemini Decision 1, HeyGen Decision 1, HeyGen Decision 1 Turbo and Decisions Fast: OpenRouter listed no such endpoint for them, and the words are made up.
- **On the first test's 38 descriptions,** the settings were the first post's: zero data retention asked of no model in the first four reads and required of every model in the fifth, and no ignore.
- **The reads:** one read of each of the 339 descriptions, and a second read of 50: the 50 whose id (such as "R001") has the lowest SHA-256. Up to six models were read at once in one run, and a free endpoint's run went on beside the paid models' run.
- **Free endpoints:** their calls started at least 3.1 seconds apart. An HTTP 429 from a free endpoint is a refusal for rate, not an answer: our code waits 30 seconds and reads the same words again. This happened twice for Decider Flash V1.3.
- **Paid endpoints:** a call that was busy (HTTP 429), timed out (HTTP 408, or no answer in 60 seconds), hit a server error (HTTP 500, 502, 503, 524 or 529) or lost its connection was tried once more, 2 seconds later, and then counted as failed.
- **The hosts that answered:** Decider V1.3 on Perplexity, Decider Flash V1.3 on Perplexity, Gemini Decision 1 on Google AI Studio, HeyGen Decision 1 on HeyGen, HeyGen Decision 1 Turbo on HeyGen, Drex Max Finance on DeepInfra, Reflex on Morph, Decisions Fast on Parallel. **The versions that answered:** perplexity/pplx-decider-v1.3-20261112, perplexity/pplx-decider-flash-v1.3-20261112, google/gemini-decision-1-20261111, heygen/heygen-decision-1-20261112, heygen/heygen-decision-1-turbo-20261112, nace-ai/drex-max-finance-20261110, morph/reflex-20261103, parallel/decisions-fast-20261106.
- **Web search:** no read of Gemini Decision 1 came back with web citations.
- **Cost:** our calls for the 339 descriptions and the second reads cost US$0.434, the failed reads read again US$0.001, and the reads of the first test's 38 US$0.169. The free endpoints cost nothing. Of these sums, Decisions Fast cost US$0.541: 541 calls at US$0.001.
- **The first test's 38 descriptions,** the question and its five options are listed in [our first post](https://plainlamp.com/blog/decision-models). We have not published the set of 339 descriptions, only the few our two posts quote. The SHA-256 of their file is fae55eb01521335b4a4f0fda10cc9150bbcb604233d44d1c58b07eda7b785d3b, so we can show later that the set did not change.

The first post, with both earlier tests: [We tested 27 decision models. One was right every time.](https://plainlamp.com/blog/decision-models)

**About Plainlamp.** Plainlamp turns a difficult PDF into a short narrated video and a plain-English guide. Every point shows the page it comes from.

[See what a lesson looks like](https://plainlamp.com/#lesson)

[All posts](https://plainlamp.com/blog)