# Enso — the intelligence layer of Hanzo OS

> Never compromise on performance. Or price. Enso reads every request and puts the right models on it — frontier accuracy when the work is hard, a fraction of the cost when it isn’t. One OpenAI- and Anthropic-compatible API, in Flash, Pro and Ultra. Proprietary, on Hanzo Cloud; the open-weights Zen family stays free to self-host.

Enso · the intelligence layer of Hanzo OS

# Never compromise on performance. Or price.

Enso is a multi-agent intelligence system, delivered as one model. It reads every request and puts the right models on it — the frontier when the work is hard, the fast ones when it isn’t.

[Start using Enso](https://console.hanzo.ai)[See the technology](https://hanzo.ai/models/enso)

Enso is proprietary and available only via Hanzo Cloud. The open-weights Zen family stays free to self-host.

Available throughOpenAI + Anthropic APIClaude Code / Codex via Hanzo CLIHanzo SDKsMCP

What is Hanzo Enso

## A multi-agent system, delivered as one model

Instead of hand-designing team roles and workflows, Enso learns to assemble agents from a pool and coordinate them through efficient, non-obvious collaboration patterns — automatically, per task.

01

### One API — every model, every modality

Text, code, vision, documents, images, audio, and video through one OpenAI- and Anthropic-compatible endpoint. Enso unifies the frontier and open models across every modality — the first hyper-modal interface. You write one integration, not ten.

02

### Superior on complex, multi-step work

Built for coding, reasoning, research, and other quality-critical workflows. Enso delivers stronger, more reliable results on hard, multi-step tasks than any single model.

03

### You control the agent pool

Opt specific providers or models out of Enso’s pool to meet data, privacy, compliance, or org requirements — with a full audit trail of which models ran, on your organization’s cloud.

Built into Hanzo OS

## Intelligence, action, outcome. One system.

Enso does not stop at an answer. It works through Hanzo agents, calls governed tools, executes in isolated sandboxes, reads your company context, deploys through Hanzo Cloud, and the result comes back measured by Observability and Insights — so the loop from intent to outcome closes inside one system rather than across five vendors.

[Agents](https://hanzo.ai/agents)[Sandboxes](https://hanzo.ai/tabs)[Company context](https://hanzo.ai/team)[Hanzo Cloud](https://hanzo.ai/cloud)[Observability](https://hanzo.ai/o11y)[Insights](https://hanzo.ai/insights)

Enso vs Zen

## Proprietary orchestration, on open foundations

### Enso

Proprietary · Hanzo Cloud only

- Learned orchestration over the best models
- Flash · Pro · Ultra presets
- One OpenAI + Anthropic endpoint
- Managed, metered, audited on Hanzo Cloud

[Explore Enso benchmarks](https://hanzo.ai/models/enso)

### Zen

Open weights · run anywhere

- Open-weight frontier models
- Chat, code, and agents
- Self-host on your own hardware
- Free — or managed on Hanzo Cloud

[Explore Zen models](https://hanzo.ai/models/zen)

The technology

## Research-driven coordination for multi-agent intelligence

Enso is grounded in Hanzo’s research on learned model orchestration (HIP-0510): how a system can learn to assemble, route, and coordinate expert agents for each task instead of relying on hand-designed workflows.

Learned router

### Microsecond routing

A lightweight coordinator scores every request and dispatches it to the right model in microseconds — routing overhead you can ignore, applied to every call.

Learned coordinator

### Roles, turns, and verification

Enso assigns Thinker / Worker / Verifier roles and adaptively delegates across coding, math, reasoning, and knowledge tasks — coordinating diverse model pools to outperform any single worker.

[Learn more about Enso](https://docs.hanzo.ai/docs/services/models)

How to use

## Three presets — price × performance

Ultra, Pro, and Flash trade cost against accuracy in that order. Pick the one that fits the work, and switch between them without changing your integration — it is one endpoint either way.

### Enso Ultra

Flagship

enso-ultra

$5 → $25/MTok in → out

Maximum quality

Our most accurate model, for work where being wrong is expensive — research reproduction, security analysis, and long autonomous runs. It leads several public benchmarks, and costs less per output token than the frontier models it beats.

- Research & paper reproduction
- Security assessment
- Deep, long-running tasks

### Enso Pro

Default

enso · the default

$4 → $20/MTok in → out

Balanced — the everyday default

The default for real work — coding, code review, and agents that have to answer while someone waits. Priced for everyday scale. Opt out of specific providers to meet data and compliance constraints.

- Coding & code review
- Responsive agents
- Provider opt-out controls

### Enso Flash

enso-flash

$2 → $4/MTok in → out

Fastest, most economical

The high-volume default. Lowest latency and lowest cost, for everyday chat, classification, extraction, and simple agent steps.

- High-volume, low latency
- Cheapest per request
- Great default for chat & tools

[Compare every measured benchmark](https://hanzo.ai/models/enso)

Quantitative results

## Frontier capability, measured — without single-vendor risk

Enso reaches frontier-level results by routing each request to the right model in microseconds. Real, measured numbers — not a fabricated benchmark table.

<15µs

Routing overhead

per request

500+

Callable models

frontier + open Zen, one API

1 API

OpenAI + Anthropic

drop-in, or via Hanzo CLI

Enso delivers frontier capability without the risk of single-vendor export controls or lock-in — the router always dispatches to a currently-available model in its pool.

Efficiency & savings

## The savings are the product

Enso delivers frontier accuracy and bills a fraction of what always calling a top model costs. enso-ultra reaches the field’s best 98.0% GPQA-Diamond for $25/MTok — 5.5× under the priciest frontier model, which scores 93.2%.

98.0%

GPQA-Diamond · enso-ultra · 194/198 correct · Hanzo-measured

5.5× cheaper

than gpt-5.2-pro, which scores 93.2%

69%

saved vs always calling a top model, at 20% hard

3 tiers

One GPQA set of 198, three routing pools: 184 · 190 · 194 correct

Measured, stated plainly: enso-ultra 98.0%, enso-pro 96.0%, enso-flash 92.9% on GPQA-Diamond — each a distinct price/quality tier, all through one API.

01

### Accuracy at cost

The goal is the top-left: high accuracy, low cost. enso-ultra sits there — 98.0% GPQA-Diamond (Hanzo-measured), at $25/MTok against fable-5 at $42 and gpt-5.2-pro at $139, both of which score lower. Pro and Flash trade accuracy for even lower cost.

enso-ultra 98.0%enso — 96.0% GPQA-Diamond · $20/MTokenso 96.0%enso-flash — 92.9% GPQA-Diamond · $4/MTokenso-flash 92.9%enso-ultra 98.0%enso — 96.0% GPQA-Diamond · $20/MTokenso 96.0%enso-flash — 92.9% GPQA-Diamond · $4/MTokenso-flash 92.9%

A solid ring is Hanzo-measured; a dashed ring is vendor-reported. enso-ultra (98.0%) leads the field on accuracy at $25/MTok — the win is accuracy-per-dollar across the three tiers.

02

### Frontier accuracy without frontier prices

Top-tier results without paying a top-tier rate on every request. Output price per million tokens across models in the field:

fable-5not used to coordinate

$42

glm-5.2

$3.73

kimi-k2.6

$2.71

deepseek-v4-pro

$2.5

gemma-4-31b

$0.44

~95×cheaper coordination — for competitive quality

fable-5 costs $42/MTok and scores 81.3% solo on our harness — a premium coordinator that is expensive and worse. The cheap models Enso coordinates run $0.44–$3.73/MTok: up to ~95× cheaper for the same coordination job.

03

### Pay for what each request needs

Simple requests cost little; only the hardest work costs more. You pick the tier, and Enso keeps every request inside that price/quality contract.

simple request

$0.006

the common case, on Flash

harder request

$0.030

escalates to Ultra, only when needed

always-max baseline

$0.035

paying top rate every time

Everyday requests cost a fraction of always paying the top rate — extra spend goes only to the hard fraction that needs it. Quality and cost are a property of the tier you choose, not a caller parameter. Per-request costs at billed token rates, 1K in / 1K out.

04

### What it saves you

Estimate the monthly bill for your volume: always calling a top model, versus Enso routing most requests to a cheap one and escalating only the hard fraction.

Requests / month1,000,000

Hard fraction (needs escalation)20%

Always top model

$35,000/mo

Enso (routed + adaptive)

$10,800/mo

You save

69%

$24,200/mo

Model: a top model bills the premium rate on every request ($0.035/req); Enso serves the easy majority on Flash ($0.006/req) and only the hard fraction escalates to Ultra ($0.030/req). Billed token rates at a 1K-in/1K-out request — your mix sets the exact number.

Built for

## What teams build with Enso

### Coding & code review

Enso finds the bugs a single model misses — comprehensive reviews that surface twenty issues where others flag three. Drop it into your existing coding tools unchanged.

### Research & autonomy

Point Enso at a paper or a patent landscape and it works autonomously — reading, implementing, training, evaluating, and connecting sources across dozens of documents in hours, not days.

### Security assessment

From one scoped instruction, Enso drives an end-to-end assessment — recon, injection and auth checks, and a clean report with evidence and retest steps — staying strictly inside scope.

### Orchestration at scale

Frontier-level output with unusually strong persona and identity stability across long sessions — the property that matters most for production agent products.

Pricing

## Pay for intelligence, not integrations

Usage-based, per-organization billing on Hanzo Cloud. When one agent handles a task you pay the standard rate for that model; when Enso coordinates several, you’re charged a single rate based on the top-tier model involved — never stacked fees.

Pay-as-you-go

For production workloads that need maximum reliability. Consumption-based tokens, served at higher priority, with transparent per-request cost you can predict and export.

- Single rate — no stacked model fees
- Per-request orchestration trace
- Per-org usage & cost export

Subscription

For casual, everyday hands-on use. Every tier includes Flash, Pro, and Ultra — upgrade when you need longer, heavier, or more frequent sessions.

- All three presets on every tier
- Standard · Pro · Max usage tiers
- Upgrade or downgrade anytime

[See full pricing](https://cloud.hanzo.ai/pricing)

FAQ

## Questions, answered

Where can I use Hanzo Enso?

Enso is proprietary and available ONLY through Hanzo Cloud — a single endpoint that speaks both the OpenAI and Anthropic API styles natively. Point your existing OpenAI or Anthropic client at the Hanzo base URL and call an `enso-*` model id — or use it in Claude Code or Codex via the Hanzo CLI. (The open-weights Zen family, by contrast, is free to run on Hanzo Cloud or self-host anywhere.)

What are Flash, Pro, and Ultra?

The three default Enso presets: Flash for fast, high-volume work; Pro as the balanced everyday default for coding and agents; Ultra for maximum quality on hard, high-stakes problems. All behind one API — switch by changing the model id. Zen and other models remain available too.

How is Enso different from the Zen models?

Zen is the family of OPEN-WEIGHT models built by Zoo Labs Foundation that you can self-host. Enso is Hanzo’s PROPRIETARY orchestration layer on top — a learned router that assembles and coordinates the best available models (Zen and frontier) per task. Enso runs only on Hanzo Cloud; Zen runs anywhere.

Can I control which models or providers Enso uses?

Yes. Opt specific providers or models out of Enso’s pool to satisfy data-residency, privacy, or compliance requirements. Every request records which models actually ran.

Will my data be used to train models? Can I opt out?

Off unless you turn it on. We do not use your inputs or outputs to train our models unless you explicitly opt in, and where an organization administers accounts the administrator controls that for everyone in it. Enso runs inside your Hanzo Cloud organization with a full audit trail. The terms are at /legal/research-contribution.

Can I see which underlying models Enso used for each query?

Yes. Each response carries the orchestration trace — the models selected, the roles they played, and the routing decisions — visible in the console and via the API.

Is Enso generally available?

Yes. Enso is available now on Hanzo Cloud and is the default for new chats and API requests — every default request routes through the Enso router, which selects the right tier (Flash, Pro, or Ultra) per task. Zen and other models stay available for explicit selection. Enterprise and dedicated deployment are available on request.

## Ready to build with Hanzo Enso?

Enable Enso for your Hanzo Cloud organization, or talk to us about enterprise and dedicated deployment.

[Start using Enso](https://console.hanzo.ai)[Contact sales](https://cloud.hanzo.ai/contact/sales)[Read the research](https://docs.hanzo.ai/docs/services/models)