17 min leer

Open Source vs Frontier Models for Enterprise AI Agents

By submitting, you consent to our use of your data. Privacy Policy.

Categoría

Agentes de IA

Compartir artículo

Where a frontier model earns its price, what depending on one risks, how to train an open model on its logs to run the processes your business depends on, and how Beam splits the work between them.

📌 Key takeaways

  • A frontier model earns its price on work people send back. On a design deck, a 10-point accuracy gain saves about $25 of review time per deck for $1.24 of extra tokens in our model, a 20-to-1 return.

  • On extraction with a properly written prompt, a smaller model reaches the same accuracy and the gain disappears. The prompt is the real work of discovery.

  • Price is the small risk and access is the big one: a year of doubled prices on a 20,000-invoice step costs $9,000; a month without the model costs about $140,500.

  • Dedicated GPUs for an open model beat the frontier API above about 135,000 tasks a month on one step. Below that, pay-per-use hosting of the same open weights costs about $1.30 per 1,000 tasks.

  • In Beam the model is a setting on each step, so the extraction step and the CFO deck step do not have to share a provider.

Most companies I talk to treat the model decision as a brand choice: a frontier lab or open source. The pilot runs on the best frontier model, it works, and the next question is how fast to roll it out.

The trouble starts in production. Once a payroll check or an invoice run depends on one frontier model, the business depends on someone else's API. A price hike is annoying. A blocked account, a trade restriction or a retired model stops the process. Yet frontier models are still better, and on some work a 10 percent gain in accuracy is worth a lot. So instead of picking a side, we worked out which work belongs on the frontier and which belongs on a model you own. My answer: use frontier models for very complex work, keep their output logs, and use them to train open models that run the processes your business depends on. The model on your production line is now a means of production, and you should own it.

This piece covers the split between frontier APIs and open models you run yourself, for the steps of an enterprise agent: where the frontier earns its price, what depending on it risks, how to own the model, and whether open source is the security risk it is sold as. It does not cover choosing among frontier models, which is in How We Select the Right Frontier Model for Each Problem, or the market map of open-weight releases, which is in Open-Weight vs Closed Models. Every number is either published, with a source, or modelled from assumptions stated in the text. Figures as of 15 November 2026.

The key idea: rent the frontier for hard work and own the model that runs the line. Frontier models handle design, decks and hard reasoning under zero retention, with logs saved. Those logs train an open model that runs extraction, classification and checks in your own environment.

Frontier for the hard work, logs saved. An open model trained on those logs for the processes that must run on Monday.

Where does a frontier model earn its price?

A frontier model earns its price on work people send back, because there a 10-point accuracy gain removes rounds of review that cost far more than the tokens. On routine work with a well-written prompt, a smaller model reaches the same accuracy and the gain disappears.

The mechanism is back and forth. On design work, Claude Fable, the frontier model we use most, needs far fewer rounds of human feedback than Claude Haiku, a smaller model. It gets closer on early passes, so a person stops correcting, regenerating and sending the work back. The rate card is easy to see. The review hours cost more and are easier to leave out.

To size it, I modelled a 10 percent gain as 10 more outputs in every 100 accepted on each pass, on four kinds of work (modelled; a smaller model against a frontier model):

Work

Accepted per pass

Sent back per 100 tasks

Review time saved per 1,000 tasks

Extra token cost per 1,000 tasks

Design deck

30% to 40%

233 to 150

$25,000

$1,237

Hard reasoning memo

60% to 70%

67 to 43

$5,357

$192

Extraction, vague prompt

85% to 95%

18 to 5

$929

$31

Extraction, written prompt

96% on both

4 on both

$0

$31

On a deck, the frontier cuts send-backs by more than a third: about $25 of review time saved per deck for $1.24 of extra tokens. That 20-to-1 return is why design, decks and hard reasoning stay on the frontier.

Send-backs per 100 tasks for a smaller model versus a frontier model on four kinds of work. Design deck: 233 versus 150. Hard reasoning memo: 67 versus 43. Extraction with a vague prompt: 18 versus 5. Extraction with a written prompt: 4 on both.

The frontier's gain is largest where work gets sent back most, and disappears once the extraction prompt is written properly (modelled).

The last two rows carry the non-obvious result. With a vague prompt, the frontier still wins on extraction, because its broad knowledge fills the gaps in the instruction. Write the prompt properly, with the rules, the definition of a correct field, the exceptions and a few examples, and the gap closes. In our deployments that writing is the real work of discovery: agreeing with the business what "done well" means.

The prompt is not free. At 40 hours of a specialist's time, about $3,600, it pays back on tokens alone after roughly 100,000 tasks, five months at 20,000 invoices a month. The bigger payoff is that the step no longer depends on one model's knowledge. Any model that follows the prompt can run it, and that portability is what the next question is about.

🧭 How we see it at Beam

We publish 98% accuracy across more than 10 million agent tasks, most of them this routine kind of work: extraction, classification and checks against written rules. The discovery work that produces the prompt is what makes those steps portable.

What does depending on a frontier model actually risk?

Two risks, and only one of them is big. Price is the small one: in our model a full year of doubled prices on an invoice step costs $9,000. Losing access for a month costs about $140,500.

Take one extraction step at 20,000 invoices a month (modelled):

  • Model cost. At $0.0375 an invoice on a frontier model, the step costs $750 a month. If the price doubles for a year, that is $9,000 extra.

  • Lost access. If the model is blocked, the team processes invoices by hand at 8 minutes each and $50 an hour: $133,300 a month. Rebuilding the step on a replacement model and checking it takes about 80 hours, another $7,200.

One month without the model costs as much as 15 years of a doubled price. A one-week outage costs four years of it.

Price versus access on one extraction step at 20,000 invoices a month. A year of doubled prices: $9,000. One month without the model: about $140,500, made of $133,300 of manual processing and $7,200 to rebuild the step.

A month without the model costs as much as 15 years of a doubled price (modelled).

I call price the small risk because it is short-term noise around a falling line. A launch discount ends and the quarter's bill jumps, but over the long run the cost of intelligence keeps falling as chips get more efficient and model architectures do more with less. Andreessen Horowitz estimated that for an LLM of equivalent performance, the cost is decreasing by about 10x every year (a16z, LLMflation, Nov 2024). At that rate, the capability you pay $750 a month for today costs about $75 a year from now, usually on a newer model.

Access is the big risk because it can end without warning: a security review, a trade war, or nationalism, where a country reserves its best models for its own people and companies. We have had a provider cut off our access. A retirement is the quieter version. Anthropic, for example, commits to at least 60 days' notice before retiring a publicly released model (Anthropic, Model deprecations). Sixty days is enough to migrate a step that has a written prompt and an evaluation set. It is tight for a step whose quality depends on one model's knowledge.

A second frontier provider helps, and a written prompt makes that switch faster. It does not help when the cause is a trade restriction that hits every lab in one country, and the backup is still someone else's API. Put a 10% yearly chance on losing access for a month with no tested replacement, and the expected loss on this one step is about $14,000 a year. That is the number to compare against the cost of owning. The full pricing of retirement, repricing, cutoff and out-of-scope actions is in The Hidden Risks of Building AI Agents on One Frontier Model.

How do you own the model that runs your processes?

You own it by training an open model on the frontier's output logs and running it in your own environment. Keep the frontier for the hard work, and let the open model run the processes your business depends on.

In the old economy, the means of production were raw materials, labour and factories. Lose the ore, the crew or the mill, and output stopped, whatever the order book said. Owning them is what made a company independent.

An AI-enabled business still has people and systems of record. But a growing share of the work those people did by hand now runs through a model. The model that extracts the invoice, classifies the case or drafts the reply is on the production line. When it lives only behind a frontier API, the company is renting its factory. A price hike changes the rent. A blocked key shuts the factory.

The means of production, then and now. Then: raw materials, labour and factories. Now: people, systems of record, and the model that runs the step. A model behind a frontier API is a rented factory; a model in your own environment is one you own.

The model on the production line is a means of production. Behind someone else's API, you are renting it.

The path is concrete. Run the frontier on the hard work and on a sample of the routine work. Save the output logs and have a person check them. Fine-tune an open model on the checked logs, test it against the frontier on the same tasks, and move the step over when its accuracy is within a point of the frontier's. Refresh it when the frontier improves and the new logs are worth learning from. If the frontier model goes away, the open model keeps running.

Why train at all, if a written prompt closes the gap? The prompt gets a small model close. Training on checked logs is what lets a small open model you host hold the frontier's accuracy, and the same logs become your test set. It is cheaper than it sounds. In 2023 Stanford researchers trained Alpaca, a 7B model, on 52,000 examples generated by an OpenAI model for under $600 (Stanford CRFM, Mar 2023). It was a research model, not licensed for commercial use, but the cost point stands. In our worked example below, the logs, the review and the fine-tune runs cost about $2,800 of a $10,000 setup.

What owning costs to run depends on volume. On one step with 5,000 tokens in and 500 out (modelled):

Where the model runs

Cost per 1,000 tasks

At 20,000 tasks a month

At 500,000 tasks a month

Frontier API

$37.50

$750

$18,750

Smaller closed API

$7.50

$150

$3,750

Open model, pay per use in your cloud

$1.30

$26

$650

Open model, dedicated GPUs, fine-tunes and upkeep

flat $5,050 a month

$5,050

$5,050

Dedicated hardware beats the frontier above about 135,000 tasks a month on a step. Below that, running the same open weights on a pay-per-use host keeps them yours for about a dollar per thousand tasks, and you can move them to your own hardware when volume grows.

Monthly cost of one step by volume and where the model runs. Frontier API: $750 at 20,000 tasks, $18,750 at 500,000. Smaller closed API: $150 and $3,750. Open model pay per use: $26 and $650. Open model on dedicated GPUs: flat $5,050. The dedicated line crosses the frontier line at about 135,000 tasks a month.

Dedicated GPUs beat the frontier API above about 135,000 tasks a month on one step. Below that, pay-per-use hosting of the same weights (modelled).

Two limits are worth stating. First, most "open source" models are open-weight: you can run them, but you usually cannot reproduce the training. For this decision that is enough. Ownership means the critical process still runs on Monday. Second, some providers' terms bar using their outputs to develop competing models. OpenAI's API terms list using Output "to develop models that compete with OpenAI" among the things you cannot do (OpenAI, API terms). Read the clause with your counsel before you train on the logs.

🧭 How we see it at Beam

In Beam, frontier outputs already serve as the examples a smaller model learns from, and the same checked outputs become the test set a step is scored against.

Is open source the security and IP risk it is sold as?

Partly. The loud version bundles a real mechanism with a broader claim, and the IP risk that is left sits in the questions you ask, not the documents you send.

The real mechanism is this. A closed provider sees every request and can refuse a malicious one centrally. Open weights in your environment have no central switch, so you run your own filters. That is ordinary security work, and it belongs on your side of the wall. It is also the answer to give a regulator: the filters and logs you run on your own model.

The broader claim is that open source as a category is unsafe and belongs off the approved list. My read is that this is partly a competitive play and partly a regulatory one, especially when it is aimed at open models from China whose weights run on the customer's own servers. Much of what people fear a model will say is already available through search. The defence worth having is the best defensive tooling, and frontier models will keep leading there. Use them for defence. An invoice agent can still run on weights you host.

There is also a cultural angle. Jensen Huang has argued that every country should build sovereign AI, codifying its own language, culture and data into a model of its own (World Government Summit, Feb 2024). A country that trains its own model can also reserve it for its own companies. That is the blocked API, seen from the other side of the border.

On IP, documents are the easy part. Use the zero-retention terms providers publish, and for most companies the productivity is worth more than the remaining risk. What zero retention does not cover is intent. Picture an engineer at a carmaker asking whether a specific battery chemistry suits a new gearbox. The design file never left, but the pattern of questions reveals the bet. Assume the provider can learn from the pattern of your questions, and keep the questions that reveal strategy on a model you run.

What should you do first?

Take one workflow you already run at volume and sort its steps in two: work people send back stays on the frontier, and work with a rule you can write moves to a model you own.

Which steps go where?

If the step...

Run it on

Why

Designs, writes a deck or reasons through a hard case

Frontier, under zero retention, logs saved

A 10-point gain removes rounds of review

Extracts, classifies or checks against written rules

Open model with a written prompt

Same accuracy once the prompt is written

Runs every day and stops the business if it stops

Open model trained on frontier logs, in your environment

Losing access costs far more than any price change

Asks questions that reveal strategy

A model you run

The query pattern is the leak

How do you start small?

  1. Pick one workflow and list its steps with monthly volumes.

  2. Write the prompt for the highest-volume routine step: rules, a definition of a correct field, exceptions, a few examples.

  3. Run the frontier on 5,000 sample tasks, have a person check the outputs, and keep them as logs and as a test set.

  4. Run an open model on the same test set. Move the step when its accuracy is within a point of the frontier's, and keep the frontier as the fallback.

  5. Check monthly: accuracy per step, cost per step, and how many days it would take to move each step to another model.

What should you ask your team or vendor?

  • Which steps depend on one specific model, and what happens the day it is gone?

  • How long would it take to move each step to another model?

  • Do we keep the output logs, and who owns them?

  • Is our provider's zero-retention term in our contract, and what do its terms say about training on outputs?

  • Which steps run in our own environment today?

  • What does each step cost per 1,000 tasks, and how many times is work sent back?

What does this look like on an invoice agent?

Take an illustrative accounts payable agent at 20,000 invoices a month, with six steps. Classification, extraction, matching to the purchase order and drafting the vendor email move to an open model with written prompts. Two steps stay on the frontier: reasoning through the 10% of invoices that are exceptions, and the monthly spend deck for the CFO. Both are work people send back, and their logs train the next version of the open model.

An illustrative six-step invoice agent split between models. Classification, extraction, purchase-order matching and the vendor email run on an open model with written prompts. Exception reasoning and the monthly CFO deck stay on the frontier, and their logs train the next version of the open model. Monthly model cost falls from $1,757 to $304.

Four of six steps keep running if the frontier model disappears tomorrow. Monthly model cost falls from $1,757 to $304 (modelled).

In our model, the monthly model cost falls from $1,757 to $304. The split costs about $10,000 once: $7,200 to write four prompts (the steps share rules, so 20 hours each), $500 of frontier runs for the logs, $1,800 for a person to check a sample of them and $500 for fine-tune runs. Tokens alone pay that back in about seven months, or about four counting only the extraction step's $14,000 a year of expected outage loss. The bigger change is that four of six steps keep running if the frontier model disappears tomorrow.

How we use this at Beam

In Beam, the model is a setting on each step, so the extraction step and the deck step do not have to share a provider. Each step defaults from the workspace's preferred models, carries its own evaluation criteria, so a model change shows up as a score, and has a fallback model for the day the first one fails. There is no hidden router choosing a model at run time; the choice is made per step, by the people who built it, and the criteria tell them whether it still holds.

Keep the frontier for the hard work, where a 10 percent gain saves hours of review. Train an open model on its logs and run your processes on it. Start with the process whose stop would cost the most. For choosing among frontier models, read The AI Model Landscape for Building Agents; for assigning a model per step, How to Assign AI Models to Your Agents, Per Step; for checking each step, How Quality Gates Make AI Agents Reliable in Production.

Open source versus frontier is the wrong fight. Rent the frontier for the work people send back, and own the model that runs the line.

FAQ

Should enterprise AI agents use open source or frontier models?

Both, by step. Frontier models earn their price on work people send back, such as decks and hard reasoning, where a 10-point accuracy gain removes rounds of review. Extraction, classification and rule checks reach the same accuracy on an open model once the prompt is written properly, and that is the work a business depends on daily.

Is it cheaper to self-host an open model than to use a frontier API?

Above about 135,000 tasks a month on one step, yes: dedicated GPUs at roughly $5,050 a month beat a frontier API at $37.50 per 1,000 tasks in our model. Below that, pay-per-use hosting of the same open weights costs about $1.30 per 1,000 tasks and keeps the weights yours.

Can you fine-tune an open model on frontier model outputs?

Technically yes, and cheaply: Stanford's Alpaca was trained on 52,000 generated examples for under $600 in 2023. Legally, check the provider's terms first. OpenAI's API terms bar using Output to develop models that compete with OpenAI, so read the clause with counsel before training on logs.

Are open-weight models a security risk for enterprises?

They move one job to you. A closed provider can refuse a malicious request centrally; open weights in your environment have no central switch, so you run your own filters and logs. That is ordinary security work, and it is also what a regulator will ask to see. The rest of the "unsafe category" claim is largely competitive.

Figures as of 15 November 2026. Modelled values are labelled as such; vendor-reported figures are marked as vendor claims.

Empieza hoy

Empezar a crear agentes de IA para automatizar procesos

Únase a nuestra plataforma y empiece a crear agentes de IA para diversos tipos de automatizaciones.

Empieza hoy

Empezar a crear agentes de IA para automatizar procesos

Únase a nuestra plataforma y empiece a crear agentes de IA para diversos tipos de automatizaciones.