Eevo-1
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity5.7/10
Self-serve hosted API with prepaid credit and free starter decisions, full docs with an OpenAPI file, an MCP server and a console. Has a SystemOne-compatible /api/v1/systemone. No weights, no gateway listings, no public SLA.
- Capability6/10
Choice, score and noul through the SystemOne-compatible API, plus a per-stakeholder expected-value mode. Not on JevBench, the Decision Index or Artificial Analysis. Temperature-fitted, no ECE. Direct mode is 'sub-second' with no p50 published.
- Adoption2/100
Launch posts on 14 and 19 Oct with low engagement (single to low double digits of likes), one Hacker News thread. MCP clients are the only integration. No named production users.
Vendor claims
- Direct mode, held-out items in both option orders: STEER 78.8 vs pplx-decider-v1.2 76.6 and Jev 1.14 71.7; PokerBench pre-flop 88.7 vs 77.3 and 74.3; behind Clef 27B on Habermas (56.9 vs 67.9)[Vendor claim — not independently verified]
- Default mode takes about 5 seconds and costs about $0.20 per decision; Direct mode answers in under a second[Vendor claim — not independently verified]
- Outperforms current frontier models on coherent consequence sizing while being orders of magnitude faster and cheaper[Vendor claim — not independently verified]
publicado (OK do Fabricio 19/10 BRT). Scores are calculated from scratch.
Read this first
- Closed API, no weights. Upright does not name the base model or the parameter count. Everything about how Eevo works comes from Upright’s own announcement and docs.
- Every benchmark number here is Upright’s. The comparison table was run by Upright, partly on public sets (STEER, CogBench, PokerBench, Habermas) and partly on its own evals. Eevo is not on JevBench, the Decision Index or the Artificial Analysis index (checked 19 Oct).
- Two products in one endpoint family. The headline feature, expected value per stakeholder, is not a System One question type. The part that competes with the rest of this catalog is the SystemOne-compatible API.
What it is
Eevo is a hosted decision model from Upright, a company of about 30 people in Helsinki that has spent nine years building models that size the positive and negative impacts of companies. Upright calls it a consequential decision model: you give it a decision, the options and whose interests count, and it returns, for each stakeholder, the option with the highest expected value and a probability for every option.
Upright announced it on 14 Oct 2026, with posts from its team on X. We put it on watch that day because the API docs were not readable yet. The docs are now public, and they include a TypeSafe-style endpoint, so it meets the inclusion criteria.
There are three ways to call it:
| Mode | Endpoint | What you get | Speed and cost (Upright’s figures) |
|---|---|---|---|
| Default | POST /api/v1/decisions |
You list the consequences under each option; Eevo compares them in pairs per stakeholder, removes double counting and returns each option’s expected value | Typically about 5 s and about $0.20 (29 to 450 model calls) |
| Direct | POST /api/v1/decisions with mode |
Eevo works out the consequences itself; one call per stakeholder, best option plus probabilities | “Sub-second” |
| SystemOne-compatible | POST /api/v1/systemone |
choice, score and noul questions about any state, up to 32 per request |
Billed on input tokens |
A client written for another SystemOne-compatible model should work by changing the base URL and key, according to the tutorial. By default each choice is asked in up to three option orders and averaged, billed once.
Specs
| Attribute | Value |
|---|---|
| Vendor | Upright (Upright Project), Helsinki |
| Model id | eevo-1 |
| Access | Hosted API with API key, MCP server, web console |
| Price | $2.50 per 1M input tokens; output tokens free. First account gets 5 Default and 50 Direct decisions free; then prepaid credit |
| Limits | SystemOne API: 1–32 questions, state up to 50,000 characters. Decisions: 2–8 options, 1–8 stakeholders, context up to 20,000 characters |
| Training (Upright’s description) | Soft targets on expected-value tasks and game-theoretic problems; a single temperature fitted afterwards on questions with known answers |
| Docs | llms.txt index · SystemOne tutorial · pricing · OpenAPI |
Evidence (UnverifiedClaim)
Upright’s announcement reports Direct-mode accuracy on held-out items, scored with the options in both orders and averaged:
| Benchmark | Eevo-1 | pplx-decider-v1.2 | Jev 1.14 | Clef 27B | OpenAI Decisions | Microsoft-Decision-1 |
|---|---|---|---|---|---|---|
| STEER (economic rationality, public) | 78.8 | 76.6 | 71.7 | 68.8 | 64.8 | 66.4 |
| CogBench urn (Bayesian, public) | 97.5 | 97.7 | 97.6 | 90.3 | 83.9 | 82.7 |
| PokerBench pre-flop (public) | 88.7 | 77.3 | 74.3 | 70.2 | 81.0 | 76.3 |
| Payoff games (Upright eval) | 95.8 | 81.2 | 81.2 | 75.8 | 71.5 | 72.1 |
| Backgammon doubling cube (Upright eval) | 83.0 | 69.3 | 74.7 | 62.1 | 68.1 | 55.3 |
| EV decisions (Upright eval) | 82.8 | 81.6 | 80.0 | 66.6 | 65.9 | 67.2 |
| Habermas group preference (public, outside intended use) | 56.9 | 59.3 | 44.3 | 67.9 | 42.6 | 40.1 |
Upright says Jev 1.14 was run through TypeSafe’s API on 18 Oct and that OpenAI Decisions refused 304 STEER requests, which count as wrong. The table is narrow by design: it tests the kind of reasoning Eevo was trained for. It says nothing about routing, moderation or the general classification work that JevBench and the Decision Index measure.
Score working (19 Oct 2026, from scratch, docs/06 formulas)
- Maturity = availability 7 × 0.30 + docs 8 × 0.25 + integrations 4 × 0.25 + support 3 × 0.20 = 2.10 + 2.00 + 1.00 + 0.60 = 5.70 → 5.7.
- Capability = types 10 × 0.30 + calibration 4 × 0.25 + latency 4 × 0.25 + accuracy 5 × 0.20 = 3.00 + 1.00 + 1.00 + 1.00 = 6.00 → 6.0.
- Adoption = engagement 5 × 0.35 + ecosystem 2 × 0.35 + production 0 × 0.30 = 1.75 + 0.70 + 0 = 2.45 → 2.
Availability 7: self-serve sign-up, free starter decisions and prepaid billing, but a brand-new product with no weights. Docs 8: tutorials, how-to guides, a full reference with an OpenAPI file and an llms.txt index. Integrations 4: the SystemOne-compatible API and an MCP server; no OpenRouter, Vercel or framework listings found. Support 3: a company behind it, documented errors and rate limits, no public SLA. Types 10: choice, score and noul. Calibration 4: temperature fitting is described, but no ECE is published (the rule for a claim without ECE). Latency 4: no p50 or p99 anywhere, only “sub-second” for Direct mode. Accuracy 5: public benchmarks run by the vendor, no public board, so the proportional rule does not apply. Engagement 5, ecosystem 2, production 0.
Fit / anti-fit
Fit: agent steps where the right answer depends on who gains and who loses, and teams that want a logged record of each trade-off; quick typed questions through a SystemOne-compatible client.
Anti-fit: high-volume classification on price (at $2.50/M input it costs far more per token than most hosted decision models), anything that needs open weights or self-hosting, and any decision that rests on Upright’s table alone. Upright’s own docs say Eevo knows nothing beyond the request and cannot tell when a deciding fact is missing.