Model Guide
Agnes 3.0 Flash: The New Free Preview Model, on OpenClaw and Hermes Agent
Agnes 3.0 Flash is the newest text model from Agnes AI (Sapiens AI), aimed at real-world agent work — stable execution, reliable tool use, and holding context across long tasks. It is live on the Agnes API today and currently costs nothing to call. You can select it on both OpenClaw and Hermes Agent instances, and this guide covers exactly how, plus the one caveat worth understanding first.
What Is Agnes 3.0 Flash?
Agnes 3.0 Flash (agnes-3.0-flash) is the 3.0-generation successor to Agnes 2.5 Flash. Agnes describes it as a new-generation text model focused on stable execution, reliable tool use, long-task context retention, and high-quality delivery — in other words, it is tuned for agents that run multi-step jobs rather than for one-shot chat.
It accepts text and image-URL input, supports function calling and multi-step tool orchestration, and exposes an optional thinking mode. It reaches your agent over the same OpenAI-compatible Agnes gateway (https://apihub.agnes-ai.com/v1) as every other Agnes model, so if you already have an Agnes API key saved, there is nothing new to set up.
Measured limits, not copied from a spec sheet
Agnes has not published a context window for 3.0 Flash yet, so we measured it against the live endpoint on 7 September 2026 rather than guess:
- Context window: 524,288 tokens. A 520,354-token prompt is accepted. A 600,354-token prompt is rejected, and the API states the limit itself: “longer than the model’s context length (524288 tokens)”.
- Output ceiling we use: 65,536 tokens. A request capped at 65,536 is accepted; 131,072 returns an error. That fixes a value known to work rather than pinning the model’s exact upper bound, so 65,536 is what our OpenClaw provider block declares. Self-hosted Hermes sets no Agnes output cap, so yours applies.
- It is a reasoning model. Every response we tested came back with internal thinking alongside the visible answer, and that thinking counts as output tokens. Agnes also documents an optional thinking mode you can control per request.
That puts it in the same 512K class Agnes documents for Agnes 2.5 Flash.
Is It Free?
Yes — calling agnes-3.0-flash today does not add anything to your Agnes balance. We confirmed that against the live billing meter, and against a paid Agnes model as a control: the paid model moved the meter on an identical request while 3.0 Flash left it unchanged.
There is one honest caveat, and it is the reason we treat this model as a preview. Agnes’s own documentation for 3.0 Flash says “pricing will be announced separately” — there is no published price at all. Compare that with Agnes 2.5 Flash and 2.0 Flash, which publish list prices of $0.03 per million input tokens and $0.15 per million output tokens, both with a current price of $0. Those are promotional zeroes against numbers you can watch. For 3.0 Flash there is no published number yet — Agnes says pricing will be announced separately.
Nothing about that stops you using it — it is free right now and you can select it in one click. It is simply why 3.0 Flash is offered as an explicit choice rather than switched on for everybody by default.
How It Behaves on OpenClaw Launch
On managed OpenClaw Launch instances, Agnes 3.0 Flash appears in the model picker as Agnes 3.0 Flash (Preview) with a $0 price, sitting alongside the other Agnes models. Three things are deliberate:
- It is not the suggested default. Agnes 2.5 Flash remains the suggested pick, because it has a published price and a published context window. No existing instance was moved onto 3.0 Flash.
- It is opt-in per instance. Selecting it declares it for that instance only, and moving that instance to another model takes the declaration back out again (a model still pinned as an agent’s primary is kept). You can select it again at any time.
- It is labelled Preview, not Paid. It genuinely costs nothing today, and the picker shows $0 rather than hiding a price we cannot quote.
How to Use Agnes 3.0 Flash (Managed)
The managed steps are identical on OpenClaw and Hermes Agent — both instance types use the same model selector, so nothing below changes between frameworks.
- Get a free Agnes API key from Agnes AI. Full walkthrough: Agnes API key setup guide.
- Save it on the API Keys page under the Agnes card. One key covers text, image, and video Agnes models.
- Open your dashboard, open the model picker on the instance you want, choose the Agnes provider, and pick Agnes 3.0 Flash (Preview).
- That is the whole flow — choosing the model applies it right away. Running instances pick the change up without a redeploy.
Self-Hosted Setup
Running your own gateway instead? The two frameworks configure this differently, so here is each one in full. They are not interchangeable.
OpenClaw (openclaw.json)
OpenClaw declares Agnes as a provider and takes a provider/model string as the primary:
{
"models": {
"providers": {
"agnes": {
"baseUrl": "https://apihub.agnes-ai.com/v1",
"apiKey": "YOUR_AGNES_KEY",
"api": "openai-completions",
"models": [
{
"id": "agnes-3.0-flash",
"name": "Agnes 3.0 Flash (Preview)",
"input": ["text", "image"],
"contextWindow": 524288,
"maxTokens": 65536
}
]
}
}
},
"agents": {
"defaults": {
"model": { "primary": "agnes/agnes-3.0-flash" }
}
}
}Note the agnes/ prefix on the primary: it is what points OpenClaw at the Agnes provider block you just declared, rather than leaving the model to resolve against whatever default provider is configured. Keep the prefix.
Hermes Agent (~/.hermes/config.yaml)
Hermes has no native Agnes provider. Agnes rides the generic custom OpenAI-compatible slot instead, with the model name bare and the gateway supplied as base_url:
model:
default: agnes-3.0-flash
provider: custom
base_url: https://apihub.agnes-ai.com/v1
api_key: YOUR_AGNES_KEYTwo things people get wrong here. The model name is bare — agnes-3.0-flash, not agnes/agnes-3.0-flash; that prefixed form is OpenClaw’s, not Hermes’s. And once base_url is set it takes precedence, so the endpoint is what actually decides where the request goes. Apply it with hermes gateway restart — note gateway is part of the command. New sessions pick a model change up without any restart at all.
Agnes 3.0 Flash vs 2.5 Flash: Which Should You Use?
Both are free to call today and both give you a 524,288-token window, so this is not a capability-versus-cost trade. The difference is maturity:
- Stay on Agnes 2.5 Flash for anything you depend on. It has a published price, a published context window, and months of production use. It is still our suggested default for a reason.
- Try Agnes 3.0 Flash for agent work that leans on tool calling and long multi-step tasks, which is what Agnes built this generation around. It is the better fit for a bot that runs jobs rather than answers one-off questions.
- Move off Agnes 2.0 Flash when convenient. Agnes has deprecated it and recommends 2.5 Flash for new integrations. It still works, we still ship it, and Agnes has not announced a shutdown date — but it is the only model here that upstream has formally deprecated.
What About the Paid Agnes Models?
Agnes 2.5 Pro Alpha’s hosted API has been deprecated; Agnes points callers at Agnes 2.5 Pro instead, which we now offer as the paid Agnes option. It is a reasoning model with text and image input, priced by Agnes at $0.45 per million input tokens and $0.90 per million output tokens, billed to your own Agnes balance.
Paid Agnes models are never selected for you and never applied automatically. They appear in the picker marked (Paid) and only take effect when you choose one deliberately.
Troubleshooting
I don’t see Agnes 3.0 Flash in my model picker
The Agnes section only appears once an Agnes API key is saved. Add or refresh your key on the API Keys page, then reopen the picker. 3.0 Flash sits below the two free Flash models, above the paid option.
Model not found (404)
Check the model id is exactly agnes-3.0-flash. On Hermes, also check you have not written the OpenClaw-style agnes/agnes-3.0-flash — Hermes wants the bare name and the gateway in base_url.
401 Unauthorized
The key is missing, wrong, or was pasted with whitespace. Re-save it on the API Keys page. The same Agnes key works across text, image, and video Agnes models, so you only need one.
Replies come back empty or get cut off
Agnes 3.0 Flash is a reasoning model, so it spends output tokens on internal thinking before the visible answer. If your max-output setting is very small, the whole budget can go to thinking and leave nothing to show. Give it room — our OpenClaw config allows up to 65,536 output tokens.
Context length errors on long jobs
The hard ceiling is 524,288 tokens including the prompt. Past that the API rejects the request outright rather than truncating it. If you are hitting it, compaction or a shorter working set is the fix.
Frequently Asked Questions
What is Agnes 3.0 Flash?
Agnes 3.0 Flash is Agnes AI’s newest text model, built for real-world agent tasks — stable execution, reliable tool use, and long-task context retention. It takes text and image-URL input, supports function calling, and is served on the Agnes OpenAI-compatible API.
Is Agnes 3.0 Flash free?
It costs nothing to call today — verified against Agnes’s own billing meter. Agnes has not published a price for it, stating only that pricing will be announced separately, so unlike Agnes 2.5 Flash there is no published rate to watch. Account entitlements and rate limits still apply.
Does Agnes 3.0 Flash work with Hermes Agent?
Yes. The same Agnes key and the same model work on both OpenClaw and Hermes Agent instances, and on managed hosting the selection steps are identical because both use the same model picker. On self-hosted Hermes, Agnes has no native provider, so it is configured through the custom OpenAI-compatible slot with base_url set to the Agnes gateway.
What is the Agnes 3.0 Flash context window?
The context window is 524,288 tokens. For output, our OpenClaw provider block declares a 65,536-token cap — a value we tested as working, not a published model limit. Agnes documents neither figure for 3.0 Flash; both come from testing the live API, and the endpoint states the 524,288 limit itself when a larger prompt is rejected. On self-hosted Hermes no output cap is set for Agnes at all, so whatever your own config specifies applies.
Should I switch from Agnes 2.5 Flash to 3.0 Flash?
Not automatically. Both are free today with the same context window, so keep 2.5 Flash for anything you rely on — it has a published price and a longer track record. Choose 3.0 Flash when your agent leans on tool calling and long multi-step tasks.
Why is Agnes 3.0 Flash labelled Preview instead of being the default?
Because Agnes publishes no price and no context window for it. It is free right now, but Agnes has not announced what it will cost, so it is offered as a deliberate per-instance choice rather than applied to everyone. Your existing model selection is never changed for you.
Is Agnes 2.0 Flash still usable?
Yes. Agnes has deprecated it and recommends 2.5 Flash for new integrations, but it is still served and still selectable, and no instance running it was moved. Plan a move to 2.5 Flash when convenient.