Big ideas
Enabling new use cases for decision models beyond high-volume cheap decisions
With Eevo, Upright extends use cases of decision models to consequence-based decision-making, beyond high-volume cheap decisions. This post explains how.
Juho Ojala
CTO & Co-Founder, Upright
Published Oct 27, 2026
Decision models are a class of large language models built for picking from a set of pre-specified options in a single pass, rather than generating a sequence of tokens (text, or other).
They natively produce probability distributions across those options, meaning each option is associated with a probability (number from 0-100%) that reflects how likely each option is the "right one".
Workload-wise, the needed computation is comparable to a similar-sized large language model generating a single token, which makes decision models fast and cheap.
The combination of being fast while natively outputting probabilities over options makes them a great fit for high-volume decisions that require natural language processing or otherwise fuzzy logic.
The initial use cases for decision models have been two-fold:
(1) "smart if statements" in software, replacing complex and hard-to-maintain hard-coded decision logic with simple LLM tasks defined in natural language
(2) replacing generative LLMs in classification tasks with a cheaper and faster alternative
Importantly, the main selling point for decision models against traditional generative LLMs has so far been cost and speed, not quality of outputs.
Indeed, compared to top-tier generative LLMs, output quality of decision models is in most cases inferior to top-tier generative LLMs like Claude Opus 5.5 or latest OpenAI models. While our decision model Eevo beats Jev and other decision models at poker, you should absolutely not use it for playing poker against Claude, at least not directly.
Given that the selling point of decision models is cost and speed as opposed to output quality, it might seem counterintuitive that the primary use case for Eevo, Upright's consequential decision model, is helping AI agents (running top-tier frontier models) make better decisions.
The cost and speed of decision models, however, is exactly what makes this practical. Eevo is built to work together with your agent. The agent, with access to all your context, and its trillions of parameters, is best equipped to explain the decision context, outline the options, and describe the consequences of the options. With that information, the agent calls Eevo's native consequential decision model API. Eevo's harness then runs hundreds or thousands of individual calls to the underlying "raw" Eevo decision model to size the consequences using pairwise comparisons, model their overlap, and compute expected value for all stakeholders, taking around 10 seconds.
Without the remarkable efficiency of decision models, such an approach would be impractical. In other words, efficiency enables spending lots of effort, which enables rigour, which enables quality. In AI parlance, this is called test-time compute scaling, which refers to spending more compute when doing work to get better quality, as opposed to spending more compute during training.
To work well for the above, our recipe for training Eevo is to start with a strong base model, and then post-train it with tasks around comparative sizing of consequences, expected value tasks, systems thinking, fermi estimation, game theory, modelling overlap of consequences, and adjacent tasks.
We originally intended Eevo to be used only for consequence-based decision-making in agents, and not like a general-purpose decision model. When we shared Eevo to our customers, many shared that Eevo seems to perform well also in more general-purpose decision model -like tasks. To our surprise, we learned that the training generalizes well also to seemingly distant tasks, such as playing Pacman.
Beyond anecdotes, Eevo is also a strong general-purpose decision model in public benchmarks, such as the Jev Decision Index (0.4.1)
| Benchmark |  Upright Eevo‑1 |  Perplexity Decider‑1 |  TypeSafe Jev 1.14 |  Cloudflare Clef 27B |  OpenAI Decisions |  Microsoft Decision‑1 |
|---|---|---|---|---|---|---|
| Decision IndexJev Decision Index 0.4.1 (public part) | Upright Eevo-1: index 74.8 | Perplexity Decider-1: index 67.7 | TypeSafe Jev 1.14: index 60.7 | Cloudflare Clef 27B: index 61.7 | OpenAI Decisions: index 47.4 | Microsoft Decision-1: index 56.9 |
| Upright Eevo-1: average rank 1.0 | Perplexity Decider-1: average rank 2.4 | TypeSafe Jev 1.14: average rank 3.1 | Cloudflare Clef 27B: average rank 4.7 | OpenAI Decisions: average rank 4.8 | Microsoft Decision-1: average rank 5.0 | |
| Economic rationalitySTEER Public eval | 81.3 | 76.6 | 71.7 | 68.8 | 64.8 | 66.4 |
| Bayesian reasoningCogBench (urn) Public eval | 98.1 | 97.7 | 97.6 | 90.3 | 83.9 | 82.7 |
| Strategic reasoning with incomplete informationPokerBench (pre-flop) Public eval | 87.7 | 77.3 | 74.3 | 70.2 | 81.0 | 76.3 |
| Game-theoretic reasoningPayoff games Upright eval | 93.3 | 81.2 | 81.2 | 75.8 | 71.5 | 72.1 |
| Risk and sequential decisionsBackgammon doubling cube Upright eval | 82.1 | 69.3 | 74.7 | 62.1 | 68.1 | 55.3 |
| Expected-value reasoningEV decisions Upright eval | 84.1 | 81.6 | 80.0 | 66.6 | 65.9 | 67.2 |
| Upright Eevo-1: average rank 1.2 | Perplexity Decider-1: average rank 2.6 | TypeSafe Jev 1.14: average rank 4.0 | Cloudflare Clef 27B: average rank 3.3 | OpenAI Decisions: average rank 5.7 | Microsoft Decision-1: average rank 4.2 | |
| Hard multi-step reasoningBBH Public eval | 94.7 | 78.6 | 90.1 | 61.8 | 51.3 | 64.6 |
| Reasoning over long narrativesMuSR Public eval | 84.6 | 53.5 | 56.5 | 74.2 | 30.7 | 34.7 |
| Causal reasoningCLadder Public eval | 90.6 | 81.6 | 77.2 | 88.2 | 24.0 | 50.6 |
| Arithmetic word problemsGSM8K Public eval | 85.4 | 80.5 | 67.8 | 52.1 | 23.2 | 57.7 |
| Reasoning about codeCRUXEval Public eval | 91.9 | 77.7 | 59.9 | 78.6 | 46.0 | 46.6 |
| Commonsense reasoningWinoGrande Public eval | 91.3 | 88.0 | 82.2 | 87.1 | 47.9 | 79.2 |
| Chess move choiceChessBench Public eval | 25.7 | 17.9 | 9.5 | 17.8 | 10.0 | 11.3 |
| Select-all-that-apply questionsSATA-Bench Public eval | 60.9 | 22.1 | 26.9 | 32.6 | 30.2 | 35.0 |
| Contract inferenceContractNLI Public eval | 80.1 | 76.9 | 57.7 | 73.0 | 66.9 | 66.4 |
| Adversarial inferenceANLI Public eval | 73.6 | 63.3 | 61.5 | 55.0 | 51.3 | 64.5 |
| Commonsense completionHellaSwag Public eval | 98.0 | 95.5 | 92.5 | 97.6 | 87.9 | 92.8 |
| Aspect-based sentimentACOS Public eval | 66.4 | 39.1 | 31.0 | 31.0 | 11.7 | 19.8 |
| Financial sentimentFinEntity Public eval | 92.7 | 91.7 | 81.9 | 94.3 | 73.8 | 87.8 |
| Sarcasm detectioniSarcasmEval Public eval | 56.4 | 67.4 | 39.9 | 31.0 | 19.1 | 49.9 |
| Stance detectionVAST Public eval | 65.1 | 70.3 | 60.5 | 39.0 | 36.0 | 53.7 |
| Clinical-trial inferenceNLI4CT Public eval | 75.0 | 72.3 | 71.1 | 66.9 | 56.6 | 67.6 |
| Hallucination detectionRAGTruth Public eval | 70.2 | 67.7 | 48.4 | 57.6 | 41.0 | 44.1 |
| Intent classificationBANKING77 Public eval | 92.5 | 90.9 | 79.7 | 94.0 | 76.3 | 86.7 |
| Intent and out-of-scope detectionCLINC150 Public eval | 97.6 | 94.4 | 90.2 | 97.4 | 88.8 | 86.3 |
| Reasoning-intensive retrievalBRIGHT Public eval | 41.6 | 43.6 | 40.1 | 40.6 | 32.7 | 40.1 |
| Product search relevanceAmazon ESCI Public eval | 53.9 | 52.9 | 46.3 | 46.5 | 43.2 | 49.4 |
| Phishing detectionPhishNChips Public eval | 94.4 | 49.3 | 67.5 | 58.7 | 3.8 | 31.8 |
| Multi-hop claim verificationHoVer Public eval | 81.2 | 72.5 | 53.5 | 31.2 | 41.9 | 50.4 |
| Upright Eevo-1: average rank 1.6 | Perplexity Decider-1: average rank 2.6 | TypeSafe Jev 1.14: average rank 4.7 | Cloudflare Clef 27B: average rank 2.2 | OpenAI Decisions: average rank 6.0 | Microsoft Decision-1: average rank 3.9 | |
| Function callingBFCL Public eval | 97.8 | 97.0 | 95.1 | 97.9 | 92.2 | 95.5 |
| Tool retrievalToolRet Public eval | 64.1 | 62.4 | 59.6 | 64.3 | 53.0 | 60.5 |
| Next API callAPI-Bank Public eval | 88.8 | 87.2 | 86.4 | 91.8 | 85.0 | 86.4 |
| Device controlHome appliance simulator Public eval | 100.0 | 90.9 | 45.5 | 80.7 | 42.1 | 50.0 |
| When to call a toolWhen2Call Public eval | 88.6 | 87.1 | 75.5 | 63.3 | 60.7 | 82.3 |
| Upright Eevo-1: average rank 1.3 | Perplexity Decider-1: average rank 2.3 | TypeSafe Jev 1.14: average rank 4.5 | Cloudflare Clef 27B: average rank 3.1 | OpenAI Decisions: average rank 5.3 | Microsoft Decision-1: average rank 4.4 | |
| Poetic formBPoMP Public eval | 98.0 | 90.5 | 80.8 | 94.1 | 81.3 | 86.9 |
| HumourHumicroedit Public eval | 37.8 | 31.6 | 23.5 | 33.1 | 22.8 | 24.1 |
| Chord recognitionPOP909 Public eval | 76.9 | 60.2 | 15.3 | 14.8 | 17.1 | 41.9 |
| Colour preferencecfcolor Public eval | 36.0 | 41.1 | 30.1 | 34.0 | 21.8 | 24.7 |
| Group preference predictionHabermas Machine Public eval | 46.8 | 40.3 | 19.6 | 54.5 | 17.9 | 14.2 |
| Caption matchingNew Yorker Public eval | 83.7 | 73.7 | 65.7 | 62.8 | 60.7 | 62.8 |
| Upright Eevo-1: average rank 2.0 | Perplexity Decider-1: average rank 3.5 | TypeSafe Jev 1.14: average rank 1.0 | Cloudflare Clef 27B: average rank 4.5 | OpenAI Decisions: average rank 4.8 | Microsoft Decision-1: average rank 5.2 | |
| Graduate-level scienceGPQA Diamond Public eval | 46.3 | 40.1 | 67.4 | 31.3 | 34.0 | 28.6 |
| Academic knowledgeMMLU-Pro Public eval | 71.5 | 68.4 | 79.9 | 61.7 | 52.4 | 60.4 |
| Expert-level questionsHLE Public eval | 2.1 | 0.0 | 3.5 | 0.0 | 0.0 | 0.0 |
Sources: Jev Decision Index (board, 22 Oct 2026) · our evaluation runs (24–26 Oct 2026) · OpenAI Decisions and Microsoft Decision-1 on the Decision Index: our runs of each vendor’s decision API (24 Oct 2026) · STEER (Raman et al. 2024, CC BY 4.0) · CogBench (Coda-Forno et al., ICML 2024) · Habermas Machine (Google DeepMind, CC BY 4.0).
Average rank = mean rank per category, among models with a result; ties averaged.
Compared to trillion-parameter generative frontier models, decision models are relatively easy and cheap to train and operate. This makes it possible for organisations to have control and sovereignty over their critical AI infrastructure, running models like Eevo on-premise, or even creating their own decision models optimized for their use, having full control over the whole stack.
At Upright, we are flexible on how organisations use Eevo. Organisations can choose whether to use Eevo's public cloud version (eevo.io) or deploy it on-premise (managed by us or the organization). For advanced use cases, we also create custom versions of Eevo optimized for your use case, with the same flexible deployment options.
For more information on Eevo, read the announcement, or try it yourself.
For more information, contact juho.ojala@uprightproject.com.
October 27th, 2026
Juho Ojala
CTO & Co-Founder, Upright
Share:
