> For the complete documentation index, see [llms.txt](https://doc.lockerprotocol.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://doc.lockerprotocol.com/hyper-agentic-trader/models-and-costs.md).

# Models and costs

The model included with your plan, where to get one of your own and its key, what a day costs, the prices lpa knows, the caps, and the bill.

Two different things think in Hyper Agentic Trader: the **assistant**, which answers you in plain words in the session and in the application's chat, and the **pilot**, which decides trades. Both use a model, and the same one. Since version 2.6.0 one is included with your plan; you can also bring your own, or run one on your own machine.

## Getting a model, from nothing

Four ways, the ones `lpa setup` and the application explain at their model step, the first one first:

1. **The model included with your plan.** Nothing to paste: your sign-in is its key (`lpa hyperkeel login`, or the application's). Every plan has `hat-fast`, the Elite plan `hat-advanced` too. It is counted in credits, 50 a month on the free plan, and a day of the daily pilot costs about 0.3 credit. It is the one recommended as soon as the terminal is connected. See [Plans and credits](https://doc.lockerprotocol.com/hyper-agentic-trader/plans-and-credits.md).
2. **A provider you pay yourself.** Open an account at [Anthropic](https://platform.claude.com/settings/keys), [OpenAI](https://platform.openai.com/api-keys) or [Google](https://aistudio.google.com/apikey), add a few dollars of credit, and create an **API key**: a long password that lets the program use the model in your name. Paste it on the settings page; it stays on this computer, in a file only you can read, and is never shown whole again.
3. **A model on your own computer**, free: [LM Studio](https://lmstudio.ai) or [Ollama](https://ollama.com). Nothing you write leaves the machine; it needs a recent computer and answers more slowly.
4. **Later**: setup goes on without one, and the pilot waits for it.

**Among the providers you pay yourself, the recommended model is `claude-sonnet-5`**, at Anthropic: the one the pilot's figures were measured with. A day of the daily pilot costs about 9 cents of model when it deliberates, and under one cent when no market is on the table (measured on 5 October 2026). It is the model recommended while the terminal is not connected to your account.

```sh
lpa config
```

The settings page opens in your browser, on `127.0.0.1`. It is typed by you: an agent's shell is refused.

## The included model

The first provider of the settings page is **Hyper Agentic Trader**, the model included with your plan. The page refuses a key pasted for it, since there is none, and counts in credits. Each answer of the assistant ends with what the call cost and what is left, the application shows the same under each answer, and `lpa hyperkeel status` says the plan and the credits left until the end of the month. The record keeps each call in dollars of reference, 100 credits to the dollar, so that `lpa store tokens` and `lpa status` read the same whatever the provider.

What leaves your computer for it: the question and the context of the workshop, to `api.hyperagentictrader.com`, under a pseudonym; never your trading key nor your address. A call it refuses for the account's own reasons, no credits left this month, the free budget of the day used or the account suspended, is `NO_CREDITS`, never tried again: the assistant says so, and the pilot pauses until midnight UTC (its day ends `no-credits`). The way out is a bigger plan or a top-up on [hyperagentictrader.com](https://hyperagentictrader.com), or a personal key or a local model on the settings page.

## The providers

| Provider                        | Its key                                                             | Where it runs                                                                                                                      |
| ------------------------------- | ------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------- |
| Hyper Agentic Trader (included) | your sign-in, `lpa hyperkeel login`                                 | api.hyperagentictrader.com. `hat-fast` for every plan, `hat-advanced` for Elite; counted in credits, 100 credits = $1 of reference |
| Anthropic (Claude)              | `ANTHROPIC_API_KEY` or `ANTHROPIC_AUTH_TOKEN`, or typed on the page | api.anthropic.com. Default model `claude-sonnet-5`                                                                                 |
| OpenAI                          | `OPENAI_API_KEY`, or typed                                          | api.openai.com                                                                                                                     |
| Google Gemini                   | `GEMINI_API_KEY`, or typed                                          | generativelanguage.googleapis.com                                                                                                  |
| OpenRouter                      | `OPENROUTER_API_KEY`, or typed                                      | openrouter.ai                                                                                                                      |
| LM Studio                       | `LM_STUDIO_API_KEY` if it asks for one                              | **your machine**, port 1234                                                                                                        |
| Ollama                          | none                                                                | **your machine**, port 11434                                                                                                       |
| Other                           | as that server asks                                                 | any OpenAI-compatible address, over https or on your own network                                                                   |

A key in the environment wins over a typed one and is never copied to disk. `ANTHROPIC_BASE_URL` points Anthropic elsewhere, and is refused unless it is https or on your own machine. A key typed on the page is kept in `~/.lpa/assistant-keys.json`, readable by your user only, and is never shown whole again. `LPA_ASSISTANT_MODEL` overrides the model. `LPA_HAT_API` moves the included model's gateway, for tests.

{% hint style="info" %}
**What leaves your machine.** With a provider, the included one included, the model receives your positions, the marks and the account address (the included model, under a pseudonym). With a local model (LM Studio, Ollama), nothing leaves the machine at all. The pilot says, before it starts, what will leave and where it goes.
{% endhint %}

## The settings page

| Panel      | What it sets                                                                                                                                                                                          |
| ---------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Assistant  | the provider (the included one first), its key where one is needed, the model (the recommended one first, each with its price), and a **Try it** button that tries the model shown                    |
| Cost       | the price the model is counted at and where it comes from, your own price if your provider bills another, and the most the assistant may spend in a day; in credits when the included model is in use |
| Limits     | the policy, the same as `lpa policy set`                                                                                                                                                              |
| Simulation | simulation mode on or off, the simulation budget, start over                                                                                                                                          |
| hyperkeel  | connect with Google, disconnect                                                                                                                                                                       |
| Account    | how the account and the agent are set up                                                                                                                                                              |

```
Settings page: http://127.0.0.1:53817/
Saved: assistant, policy.
```

## The prices are known

`lpa` holds the public price lists of Anthropic, OpenAI and Gemini, and the reference price of the included model (read on 20 October 2026, read again at each version): for a model on them there is nothing to type, and the Cost panel says the price it counts with and the day the list was read. Your own price always wins: type it on the Cost panel when your provider bills you another one (a negotiated rate, another region), and remove it to go back to the public one. The Cost panel takes five figures per million tokens, input, output, the cache's read price and its two write prices, for five minutes and for an hour; left empty, the hour's write price is taken as the five minutes' one.

A model on no list (OpenRouter, a server of your own) is yours to price. Prices are not decoration: the daily pilot weighs the model's own cost against the gain it aims at, and plays no day with a model it cannot count:

```
anthropic/claude-of-tomorrow has no price set, and it is not on the price lists lpa holds: its cost cannot be counted. `lpa config`, Cost, takes the price your provider bills.
```

A model on your machine costs nothing to count. The same figures make `lpa store tokens` show dollars instead of "no price".

## What the account must hold

The charter keeps the model's cost under a tenth of what a trade aims at (`modelShare`, 0.1), and the first step of its ladder must be an order Hyperliquid takes ($10). With the default charter the daily pilot therefore opens nothing below $120 of equity, and with `claude-sonnet-5` it trades from about $324; a day without a trade is carried by the next trade, so an account at that figure trades when it trades every day and waits longer otherwise. A cheaper model lowers the figure (`claude-haiku-4-5`: about $180), a dearer one raises it (`claude-opus-5`: about $830). `lpa setup` says these figures at its funds step, and `lpa pilot schedule --on` before you say yes.

`lpa pilot charter` says both floors for the strategy in use, and only the first when that strategy asks for no model:

```
The charter is the measured one. The smallest account it opens anything on: $120; with claude-sonnet-5, about $324.
```

## A day the code decides costs nothing

A strategy may be decided by the code alone (`decider: "code"`): `donchian` ships that way, and its one sentence in `lpa pilot strategy list` opens on *No model*. On a day like that no model is asked, no brain is read, nothing is spent, and nothing the pilot reads leaves the computer for a provider:

```
The code decides each day by the strategy's rule: no model is asked, nothing the pilot reads leaves this computer for one, and a day costs nothing but the fees of its orders.
```

`lpa pilot run-now`, `lpa pilot replay` and `lpa pilot schedule --on` then need no model set at all, and `lpa pilot charter` asks the account for no share of a model. A day with such a strategy in use reads *with the code's rule, no model* in its first line.

## The two short calls beside a day

| What                                   | When it is called                                                                                                                                                                                                                 | Where its dollars land                                                                                                        |
| -------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------- |
| **The choice of the day**              | before the day, when the charter names two to five strategies and `--by model` chooses among them: one brief call that sees each market's regime and each strategy in one sentence, and can only answer with a name it is allowed | the day's own tokens, counted against the charter's `apiBudgetPerDayUsd`, like a deliberation                                 |
| **A strategy drafted from a sentence** | when you run `lpa pilot strategy draft --message "..."`, and only then                                                                                                                                                            | an activity of its own in the record, which `lpa store tokens` counts. It is not a day, so the day's budget does not bound it |

A choice `--by table` reads the regime of the leading market on the same candles as the dossier and calls no model at all. A day whose choice can only fall on strategies the code decides needs no model either; one that may fall on a strategy a model decides asks for a model up front, whichever is chosen in the end.

## The prompt cache

On Anthropic, `lpa` asks the provider to cache what it sends, so that the next call pays full price only for what is new.

* **In a session**, every request asks for an hour. The instructions, the tools and the earlier exchanges are then read from the cache at a fraction of the input price, a tenth or less depending on the model.
* **The daily pilot's trader** asks for five minutes when a day holds more than one deliberation, because those calls follow each other within seconds, and a write for an hour costs twice the input where five minutes cost 1.25 times. Measured on 2 October 2026: each call after the first read the whole prefix from the cache.
* **The pilot's reader**, called once a day, asks for no cache at all: nothing would read it back.

The reasoning is asked for as a summary, so the record and the pilot's log can keep it.

{% hint style="warning" %}
**One deliberation writes no cache.** By steps, the pilot runs one deliberation a day, so nothing reads the trader's prefix back and `lpa` does not write it to the cache. Measured on 3 October 2026: 32 066 tokens written for five moments and none ever read, a tenth of that run's cost.
{% endhint %}

## Caps on what the model may spend

| Cap                                             | Where                                                           | Default                                                           |
| ----------------------------------------------- | --------------------------------------------------------------- | ----------------------------------------------------------------- |
| The credits of your plan, on the included model | the plan: its credits a month, and the free plan's daily budget | the free plan's 50 a month                                        |
| Dollars a day on the model                      | the charter's `apiBudgetPerDayUsd`                              | 0.5, up to 50                                                     |
| Dollars a day on the assistant                  | the Cost panel's "Most a day"                                   | none. Reached, every call answers `SPEND_CEILING` until 00:00 UTC |
| Dollars a day, for the continuous pilot         | `lpa pilot start --max-spend-a-day`                             | none. It needs a price set for the model                          |
| Deliberations a day                             | `lpa pilot start --max-thoughts-a-day`                          | 120, 1 to 2000                                                    |
| Dollars for one day played now                  | `lpa pilot run-now --budget`                                    | the charter's                                                     |
| Dollars for a whole replay                      | `lpa pilot replay --budget`                                     | 2                                                                 |

A pilot that uses up its day's dollars or its day's deliberations pauses itself, says so, and goes on after midnight UTC. A pilot the included model refuses for credits pauses the same way (`credits`), and its day ends `no-credits`. `lpa pilot deliberate --estimate` says what one deliberation costs before it is asked, from the last ten made here.

## Reading the bill

```sh
lpa store tokens
lpa store tokens --by day --since 7d
lpa store tokens --by provider
lpa store tokens --run 42
```

```
MODEL                      CALLS  INPUT   OUTPUT  CACHE   REASONING  COST
anthropic claude-opus-5    12     48 210  3 904   31 000  0          $0.4312
lmstudio qwen/qwen3.8-27b  3      27 115  18 902  0       17 650     no price
total                      15     75 325  22 806  31 000  17 650     $0.4312*
```

| Column    | What it counts                                                                   |
| --------- | -------------------------------------------------------------------------------- |
| CALLS     | how many times a model was asked                                                 |
| INPUT     | the tokens sent, as the provider counted them                                    |
| OUTPUT    | the tokens written back                                                          |
| CACHE     | what the provider read from its cache **and** wrote to it                        |
| REASONING | the part of the output the provider counts as reasoning, when it counts it apart |
| COST      | the dollars, from the prices you set                                             |

The star on the total means some calls had no price set for their model, so their dollars are not counted. A call that failed before any usage, a timeout or an overload, is neither spent nor unpriced: the model's price never refuses a trade because of a call that never happened.

`--by` groups the table by `run` (one activity), `provider`, `model` or `day`.

## Every call is kept

Each call to a model is a line of the record (`~/.lpa/activity.db`), carrying the activity, the turn and the step, the provider and the model, the tokens, the dollars, how long it took, how it ended, and the request, the reasoning and the answer in full.

```sh
lpa store export --what calls --out ~/calls.jsonl
lpa store prune --calls-before 30d
```

Nothing of that register is ever removed on its own. `lpa store prune --calls-before <date>` is the one hand that takes from it, and `lpa doctor` says when that register weighs more than 2 GiB.

The live page shows the same figures in its Tokens block, on each activity's card, and opens each call with its request, its reasoning and its answer.

## What a day actually costs

The pilot's log prints a line per call with its tokens, its seconds and its dollars, and the day's sum at the end:

```sh
tail -f ~/.lpa/logs/pilot.log
```

A replay prints the same kind of sum for the whole run:

```
Tokens: 14 calls; input 12940, cache read 0, cache written 30227 for 5 min and 0 for 1 h, output 9265; cache reuse 0 %; $0.1941
```

{% hint style="info" %}
**The model is a trading cost.** The daily pilot amortizes it over the trade and refuses a trade where the model would eat more than `modelShare` of the gain aimed at, 10 % by default. A model too expensive for the stake is a refusal, not a surprise on the bill.
{% endhint %}

## Learn more

* [Plans and credits](https://doc.lockerprotocol.com/hyper-agentic-trader/plans-and-credits.md)
* [The charter](https://doc.lockerprotocol.com/hyper-agentic-trader/charter.md)
* [The brain and the lessons](https://doc.lockerprotocol.com/hyper-agentic-trader/brain-and-lessons.md)
* [lpa config](https://doc.lockerprotocol.com/hyper-agentic-trader/commands/config.md) and [lpa store](https://doc.lockerprotocol.com/hyper-agentic-trader/commands/store.md)
