AI Agents WATCH CLOSELY OPPORTUNITY

Holo4: an AI model that operates business software, for operations teams

ShiniFrontiers · Source: H Company

Section 1 of 6 FREE PREVIEW

What problem does it solve?

What it is: Holo4 is an AI model from H Company that carries out tasks on a computer, much as an employee would. It can read what is on the screen and click and type, and it can also connect directly to other software through an API, a built-in link between systems.

Who it's for: businesses whose staff spend time moving work between several software systems and who want an AI to take those tasks over.

The catch: it fully completes only about four in ten long tasks, so its work still needs checking, and an engineer has to set it up before anyone can use it.

A. Work that crosses several systems — for operations teams

Problem: Most business tasks touch more than one system. A single task might start in an email, pass through a supplier's website and end in the accounting system. Most AI tools can work with only one kind of system, so people still do the steps in between.

What it does: Holo4 can handle every step of such a task. Where a system has an API, Holo4 uses it; where it does not, Holo4 works through the screen as a person would.

B. Lower running costs, and data that can stay in-house — for finance and IT heads

Problem: The most capable AI models charge per task, and long tasks are expensive. Every screenshot they work on is also sent to the AI provider.

What it does: In H Company's test of long computer tasks, the stronger version of Holo4 completed 41.5% of tasks at $1.22 each, while the best-performing model completed 48.7% at $8.48. Holo4's cheaper version can also be installed on the company's own servers, so screenshots never leave the business.

Note: For commercial use, the stronger version is available only through H Company's paid online service.

See the difference

The situation	Today	With Holo4
Work that crosses several systemsMonth-end: bringing statements from several bank websites into the accounting system	A finance executive logs in to each bank's website and re-keys the figures	Holo4 signs in to each website, reads the statement and enters the figures in the accounting system
Lower running costs, and data that can stay in-housePaying for long computer tasks	$8.48 per task with the best-performing model, 48.7% completed	$1.22 per task with Holo4's stronger version, 41.5% completed
Lower running costs, and data that can stay in-houseAutomating screens that show customers' financial data	Every screenshot is sent to the AI provider	Holo4's cheaper version runs on the company's own servers
Work that crosses several systemsDeciding whether it is good enough	Vendor demos and headline test scores	Watch a step-by-step recording of every test H Company ran, before committing

The first row is our illustration of the kind of work H Company describes. The other rows use H Company's published figures, licences and test recordings.

What it saves: Staff time spent moving work between systems, and about 85% of the AI cost per long task compared with the best-performing model in H Company's test.
What it costs: Free to download. On H Company's online service, 2 cents to $1.22 per task in its tests, plus an engineer's time to set it up.
How ready: Working in the real world

Released 28 September 2026, online and as a download. All results come from H Company's own tests; nobody else has checked them yet.

Join free to unlock: how it works, a step-by-step setup with its costs, what you gain, and where it falls short, for every Friday idea.

Join free to read it all →

Section 2 of 6 MEMBERS

How does it work?

Join free to read this section

Holo4 — how it fits together.

Join free to unlock: how it works, a step-by-step setup with its costs, what you gain, and where it falls short, for every Friday idea.

The big picture

Most AI tools are like an employee who can only use the phone: if a department has no phone line, the task stops and comes back to someone's desk. Holo4 uses the phone line when there is one, and when there is not, it walks over and works on that department's computer itself. Technically it is a computer-use agent model: it looks at the screen, chooses one action, and repeats until the task is done.

Holo4 makes the decisions but cannot press any keys by itself. A small helper program, which engineers call the harness, takes a screenshot, for example of a bank's statement page showing this month's closing balance, sends it to Holo4, and then carries out the action Holo4 asks for, such as typing that balance into the accounting system. H Company provides such a program, called hai-agents, and designs Holo4 to work with it.

What it takes to put Holo4 to work

The four pieces a business needs, from the person who asks for a task to the software where the work happens. Green is new; orange is what the business builds.

The request. A member of staff or a workflow tool: describes the task in plain language, for example: enter this month's closing balance from the bank's website into the accounting system. Tasks are described, not programmed step by step.

The helper program. Helper program (harness): takes a screenshot, for example of the bank's statement page, sends it to Holo4, carries out the action Holo4 asks for, and repeats. Holo4 cannot act without it, so this is where the engineering work goes.

The AI model. Holo4: reads the screenshot and decides the next step, for example: the closing balance is here; type it into the ledger. The cheaper version can run on the business's own server.. $1.22 per long task with the stronger version, against $8.48 with the best-performing model.

The business's software. Websites, desktop programs and Android apps: used through the screen, including systems with no API. Systems with an API: connected to directly, for example the accounting system. A safe workspace for code: where Holo4 can write and run small programs, for example to rearrange data. One model works across all of them, so a task does not stop where a system has no API.

Holo4: technical view

What the test results mean for a buyer

Holo4 does not beat the best-performing model, but it comes close for a fraction of the price. In OSWorld 2.0, a public benchmark of long computer tasks, the stronger version completed 41.5% of tasks at $1.22 each, while the best-performing model completed 48.7% at $8.48. The cheaper version costs less again but completed only about one long task in eight. H Company trained Holo4 on about 10,000 practice tasks whose answers could be checked automatically, then with reinforcement learning, rewarding it for each task it completed.

Model	Long tasks completed	Cost per task
Claude Opus 5.5, the best-performing model in the test	48.7%	$8.48
Holo4, stronger version (27B)	41.5%	$1.22
Qwen3.8 27B, the model the stronger version was built on	19.4%	$3.49
Holo4, cheaper version (35B-A3B)	12.3%	$0.61

Test	Stronger version	Cheaper version	Best result in H Company's table
Business tasks done through APIs (AutomationBench)	45.4% · $0.05	34.5% · $0.02	50.3% · $3.05 (Opus 5)
Phone apps (AndroidWorld)	85.1% · $0.08	77.6% · $0.07	88.8% (Fable 5)
Shorter computer tasks (OSWorld)	85.2% · $0.08	80.8% · $0.05	86.1% (Qwen3.8 Max)
New tasks kept out of training: 47 websites	80.2%	70.8%	Not tested
New tasks kept out of training: 14 connected business tools	89.4%	85.4%	Not tested
New tasks kept out of training: 17 desktop programs	72.0%	64.8%	Not tested

Why not use an AI that only connects through APIs?

An AI that only connects through APIs works well wherever such a link exists, but it stops at any system without one. An AI that only works through the screen has the opposite problem, because some systems have no screen to read. Holo4 does both, and can write small programs to join the two within one task.

	API-only AI	Screen-only AI	Holo4
A system with no API	❌ No way in	✅ Works through the screen	✅ Works through the screen
A system with no screen	✅ Connects directly	❌ Nothing to read	✅ Connects, or writes a program
A task that needs both	⚠️ Needs a second tool	⚠️ Needs a second tool	✅ One model, one task

Choosing between the two versions

H Company released two Holo4 versions. It also released Holotron4 Nano, a much weaker model on long tasks, and still offers its earlier Holo3 models; neither is covered here.

	Stronger version (Holo4 27B)	Cheaper version (Holo4 35B-A3B)
Price on H Company's service, per million tokens in / out	$0.40 / $3.00	$0.30 / $2.00
Business use on your own servers	❌ Not allowed (CC BY-NC 4.0 licence); business use only through H Company's service	✅ Allowed (Apache 2.0 licence)
Running it yourself	Not for business use	An engineer downloads the open-weight files from Hugging Face and runs them with free tools (llama.cpp or vLLM), following H Company's guide; screenshots then never leave the business
How it is built	Uses all of its parts for every word	A mixture of experts: uses only about a tenth of its parts for each word, which makes it faster and cheaper
H Company's advice	Complex tasks with many steps	Start here; it is fast enough for back-and-forth work

One task, step by step

Step	The helper program sends Holo4	Holo4 replies	What happens
1	The task: enter this month's closing balance from the bank's website into the accounting system, plus a screenshot of the bank's sign-in page	Click the user-ID box and type the test account's ID	The helper program types it and signs in
2	A screenshot of the account summary page	Click the link for this month's statement	The statement page opens
3	A screenshot of the statement page, which shows the closing balance	Record that closing balance in the accounting system, through its API	The accounting system confirms the entry
4	The confirmation from the accounting system	The task is complete	The operations manager checks the entry

Join free → I have an account

Section 3 of 6 MEMBERS

How to set it up, and what it costs

Join free to read this section

Holo4 — steps and cost.

Join free to unlock: how it works, a step-by-step setup with its costs, what you gain, and where it falls short, for every Friday idea.

Before you start: is it a fit?

	Holo4
Good fit if	Staff spend significant time moving work between several software systems
	At least one of those systems has no API
	The business already pays an AI provider for this kind of work and wants to cut the cost
	The screens involved show data that should stay inside the business
	A person can check the results before they take effect, for example payments approved by a manager
Who acts on it	The head of operations chooses the tasks, the head of technology decides where Holo4 runs, and an engineer sets it up
Not yet if	Mistakes would reach customers or the accounts unchecked: even the stronger version completes about four long tasks in ten
	The business wants the stronger version on its own servers for commercial use, which its licence does not allow

See it work before you commit

H Company has published a step-by-step recording of every test behind its scores, and anyone can replay them in a web browser at `trajectories.hcompany.ai`. Watching where Holo4 succeeds and where it gets stuck is the quickest way to judge whether it suits your work.

Part A: Decide the scope

Choose two or three tasks for a trial (the head of operations) Pick tasks that cross more than one system and whose results a person can check.
Decide where Holo4 will run (the head of technology) Both versions are available on H Company's paid online service. On your own servers, only the cheaper version may be used for business.
Start with the cheaper version, as H Company advises (the engineer) Switch to the stronger version for long tasks or unfamiliar software.

Part B: Make a first request

You need	Why
An engineer	To connect Holo4 to the business's software
An account on H Company's platform, with credits	Holo4 is on the paid plan; the free plan offers only the older Holo3.1
A test copy of each system, with test data	Each step sends a screenshot to Holo4

Get an access key (the engineer) Create a key on H Company's platform (portal.hcompany.ai) and add credits.
Install OpenAI's standard software (the engineer) H Company's service accepts the same requests as OpenAI's, so OpenAI's standard software works with one web address changed.
Send a screenshot (the engineer) Use a screenshot of one of the business's own systems. To try the cheaper version, change the model name to holo4-35b-a3b. Expect: a reply describing the screen and what Holo4 would click to sign in.

Part C: Run the trial

Follow H Company's "Build an agent" guide (the engineer) It explains the helper program: the task and the screen go in, and one action comes out, again and again until the task is done.
Start from one of H Company's ready-made examples (the engineer) H Company provides working examples for a web browser, for running code, for connected business tools, and for a mix of all three.
Measure the trial on your own tasks (the head of operations) Run the chosen tasks on test data and record how many finish correctly and what each one cost. Expect: your own completion rate and cost per task, to compare with H Company's figures.

What it costs

Item	Cost
Downloading Holo4	Free. The cheaper version may be used for business on your own servers; the stronger version may not.
Using H Company's service	Per million tokens; see the Holo model family table in section 2
One task, in H Company's tests	2 to 5 cents for business tasks done through APIs; 61 cents to $1.22 for long computer tasks
People	An engineer's time to set it up and maintain it, and staff time to check results

Join free → I have an account

	What changes	What it can mean for the business
Reach	Systems with no API can be automated through their screens	Fewer steps left for staff to re-key by hand
Cost	$1.22 per long task, against $8.48 with the best-performing model in H Company's test	About seven times as many tasks for the same AI budget
Control	The cheaper version can run on the company's own servers	Screens with sensitive data stay inside the business
Evidence	Every published test can be replayed step by step	A decision based on watching it work, not on headline scores

Limit	Why it matters	What to do
Completes a minority of long tasks	The stronger version completed 41.5% and the cheaper version 12.3% in H Company's long-task test.	Start with short tasks that a person checks.
Results are H Company's own	Every Holo4 score comes from H Company's own tests, and nobody else has checked them.	Treat your own trial as the test that counts.
Licence of the stronger version	Its download may not be used commercially (CC BY-NC 4.0).	Use it through H Company's paid service, or run the cheaper version in-house.
Needs an engineer	Holo4 chooses actions; a helper program has to carry them out.	Plan engineering time and start from H Company's examples.
Hardware needs not stated	The sources read for this paper do not say what server Holo4 needs.	Read H Company's guide before buying hardware.

Holo4: an AI model that operates business software, for operations teams

What problem does it solve?

How does it work?

The big picture

What it takes to put Holo4 to work

What the test results mean for a buyer

Why not use an AI that only connects through APIs?

Choosing between the two versions

One task, step by step

How to set it up, and what it costs

Before you start: is it a fit?

See it work before you commit

Part A: Decide the scope

Part B: Make a first request

Part C: Run the trial

What it costs

What do you gain?

What are the limits?

What this project can't do yet

What it does not do

Sources