AI Model & BYOK #

This page explains which AI model powers your bot, why only one provider is supported for BYOK (Bring Your Own Key), and how to set up your own key if your plan includes it.

Which AI Model Powers the Bot #

The platform uses Anthropic Claude as the primary model for your bot. Claude was chosen because, after extensive testing across every major frontier model (industry term for the most advanced AI models available) on the market, it is the only one that handles production messaging volume reliably.

You do not need to choose a model. The platform automatically picks the right one for each task and optimizes how it is used behind the scenes (including advanced prompt caching, which dramatically reduces cost per conversation).

Why Only Anthropic for BYOK #

Customers often ask whether we will add support for OpenAI, Google Gemini, OpenRouter, or open-source models for BYOK. The honest answer is: none of them, used raw and unoptimized, hold up at production volume.

For a small project, a 9-out-of-10 success rate from an AI is fine. Once you are running real customer conversations at scale, a 10% failure rate means broken bookings, wrong answers, missed follow-ups, and lost sales. That is in the ballpark of what we see even from top-tier frontier models from other providers (roughly 6–10% on adversarial edge cases, meaning tricky, deliberately hard test questions) — and that is before considering the cheaper models.

The issues fall into three buckets:

• Edge cases break unpredictably. A model can work perfectly for hundreds of chats, then suddenly hallucinate a tool call, claim it did something it did not, or drop instructions mid-conversation.
• Long context is poorly handled. A single bot response uses chat history, FAQs, follow-up logic, tool calls, and reasoning. Most models choke on this volume and produce worse output.
• “Cheaper” rarely is. The lower sticker price per token is offset by more retries, more output tokens, and worse conversion. The real cost per successful conversation ends up higher.

Supporting 100 different models is more of a marketing line than a real capability. We would rather give you one provider that works reliably than five that each break in different ways.

Our Own Model: Max #

We built our own in-house model, Max — an in-house model plus a multi-pass optimization stack, refined on real conversation data from the platform. The goal was a model significantly more cost-efficient than Claude while staying reliable on the conversation patterns the platform actually runs, and that’s what Max delivers: a fraction of the Pro price (0.25 credits per action), tuned to stay accurate where a raw budget model slips.

Max needs no setup — no model picking, no provider account, no key. It runs entirely on our infrastructure, and it’s the most accurate model we ship: in our own benchmarks it’s the strongest at sticking to what your business actually says. For a full breakdown of how Max compares to Claude and to the budget third-party models customers often ask about — including what those models actually got wrong in our testing — see Why Max Is the Best Value.

Who sees the Max tier #

If the AI Quality picker on your Agent’s AI Instructions tab only shows Pro and Economy, Max simply isn’t switched on for the account you’re logged into. Two things turn it on:

• Accounts on an eligible lifetime plan get Max automatically.
• Any other account needs Max enabled on it explicitly.

On an account provided by an agency, your provider may have chosen the model for you. If the AI Quality picker shows only one card and there’s nothing to switch to, that’s a deliberate setting on your account, not a fault — ask whoever provides your account to change it. If a card is still shown with a note that it isn’t available any more, your Agent is on a model your provider has since removed: it keeps replying, but on one of the models they do allow, so pick one of those to make the picker match what actually runs.

Connecting your own Anthropic key (BYOK) has nothing to do with whether Max is visible. The two work side by side — see the exception note under Setting Up BYOK for how a Max Agent is billed when you have a key connected. Removing your API key will not make Max appear.

The Mini tier #

If you can see Max, you can see Mini too, at 0.15 credits per action. Mini is the same AI as Max at a lower price, with two trade-offs:

• A higher chance of mistakes. Mini is the lighter model of the two. Max is the most accurate model we ship — tuned hardest to stick to what your business actually says — while with Mini there’s a higher chance of a small mistake or hallucination slipping into a reply.
• Lighter memory. In long conversations, Mini remembers fewer of the older messages and pulls in fewer knowledge-base snippets per reply.

Max stays the recommended choice for anything where a wrong detail can cost a sale — bookings, pricing questions, support with real stakes. Mini fits high-volume, price-sensitive conversations where speed and cost matter more than the last percent of accuracy.

Mini sits in the same AI Quality picker as Max, on every account that has Max. Like Max, Mini runs entirely on our infrastructure and never uses a BYOK key, so a connected Anthropic key is not used (and not billed) while an Agent is on Mini.

What the 0.15 Mini rate covers. The Mini price applies to everything that happens inside a live conversation: AI replies, AI tool calls (like booking an appointment or alerting a human), chat evaluations (deciding whether a conversation should be ignored or closed), interruption handling (when a contact sends a new message while the bot is still composing its reply), chat summaries and AI auto-tagging. One-off AI features outside the conversation — campaign optimization, knowledge-base generation, and tagging done through the API — have a single flat rate of 0.25 credits on both Max and Mini, so you may see occasional 0.25 entries on your credit history even when every Agent is set to Mini. That’s those features, not the Max model answering your chats.

Setting Up BYOK #

BYOK is included on the Agency and Agency Unlimited plans, and on the AppSumo lifetime tiers from Plan 2 up. It is not part of the Business plan. If your plan includes it, you can connect your own Anthropic API key (a private code from your Anthropic account that lets the platform bill your AI usage directly to you instead of using platform credits).

Only official Anthropic keys work. A valid key comes from your own account at console.anthropic.com and always starts with sk-ant-. Keys from resellers, proxies, or “discounted Claude API” sites (often starting with a different prefix) are not supported and won’t save. Be careful with offers of cheap Claude tokens far below Anthropic’s own prices — nobody can legitimately resell Claude access below what Anthropic charges, and those offers are almost always scams built on stolen keys or set up to phish your payment details.

On a phone? The left sidebar may be hidden. Tap the menu icon (☰) first, then tap Settings.

Don’t see BYOK API Keys in the Advanced group? The section only appears on plans that include BYOK. If your account is managed by a provider, it appears only when they have switched Bring Your Own API Key on for your account.

1. In the main left sidebar, click Settings (the gear icon near the bottom).
2. On the Settings page, in the left menu, scroll to the Advanced group.
3. Click BYOK API Keys.

4. Click Add key on the Anthropic card. A Secret key field appears (placeholder sk-ant-…).

5. Paste your Anthropic API key into the field, then click Save key.

Once enabled, your bot will use your own key for AI calls, and you will pay Anthropic directly for that usage instead of spending platform credits on AI responses. Channel infrastructure (WhatsApp Web servers, phone numbers, etc.) continues to use credits as usual.

One exception — the Max tier. If an Agent’s AI Quality is set to the Max tier, that Agent runs entirely on our in-house infrastructure and is billed at 0.25 credits per action even when you have BYOK connected — Max never uses your Anthropic key. (The same applies to Mini, where available: 0.15 credits per action, never your key.) To bill replies to your own Anthropic account, keep the Agent on the Pro (standard) tier. The AI Quality picker lives on each Agent’s AI Instructions tab. The BYOK API Keys settings page shows a notice whenever your key is on file but one or more of your Agents are still running on the Max tier, listing which ones — so you can spot at a glance why credits are still being used.

A note on cost: Bringing your own key is no longer the cheap option. With a key connected, your Agent runs on Claude and you pay Anthropic’s metered rates directly (Anthropic bills you based on how much you actually use, similar to a utility bill). Measured across the accounts running their own key, one AI reply costs roughly 3 times what the same reply costs on Max and more than 5 times what it costs on Mini, once you count the background work a reply triggers (knowledge-base lookup, spam check, name and email extraction, chat evaluations, tagging), all of which is included in the flat Max and Mini rate but billed per token on your key. Our prompt-caching optimizations apply to your own key too, so this is not a caching gap; Max and Mini are simply much cheaper models that also score higher on our sales-conversation benchmark. Use BYOK when you specifically want AI usage billed to your own Anthropic account, not to save money.

Why the first reply in a conversation costs much more than the rest. Caching works by storing the conversation’s context with Anthropic the first time it’s sent, then reusing it. Writing it costs more than a normal request; every reply that reuses it costs roughly a twentieth as much. So the first reply looks alarmingly expensive and the ones after it are very cheap. Anything large in the conversation makes that first reply bigger — a document a visitor uploads is the main one. A 24-page PDF works out to around 70,000 tokens, because each page is sent as both its text and a picture of the page, and it stays in the conversation, so it’s re-sent (and re-cached) with every reply. If a conversation goes quiet for over an hour the stored copy expires and the next reply pays the write again.

If you’re building something document-heavy — contract explainers, manual lookups, anything where people upload files — the Max tier is usually the better fit: it’s a flat 0.25 credits per reply no matter how big the document is, so there’s no first-message spike and no per-token exposure at all. See Why Max Is the Best Value for how Max compares on quality and cost.

If your BYOK key fails (invalid, out of quota, or rate-limited — meaning Anthropic has temporarily capped how many requests your key can make), whether your bot keeps running depends on the “Fallback to credits” toggle on the same BYOK API Keys page:

• On: the platform temporarily falls back to platform credits so your bot keeps replying, and keeps retrying your key in the background — the moment a call succeeds again it switches you back to your own key automatically.
• Off (the default): AI replies are blocked until you fix the key. Nothing silently spends your credits, but your bot also stops replying, so if you’d rather it keep going on credits while you sort out the key, turn this toggle on ahead of time.

Either way, you’ll get an email the first time your key fails (not one per message).

Frequently Asked #

Can I use OpenAI, Gemini, or another provider for BYOK? Not currently, for other providers. For most people the answer is already here: Max and Mini are our own in-house tiers, at 0.25 and 0.15 credits per action.

Will my BYOK key be used for everything? Most AI operations are free of credit charges when your own Anthropic key (BYOK) is connected: AI responses, tool use, campaign optimisation, knowledge-base generation and web search all run on your key. Four things still cost credits even with a key connected: the Max tier at 0.25 credits per action and the Mini tier at 0.15, because both run on our own models; Lead Finder at 0.25 per email address found; and Automations at 0.25 per run plus 0.25 per AI step. To keep an Agent fully on your own key at 0 credits, leave it on the Pro tier. Channel fees (WhatsApp Web’s monthly fee, per-message WhatsApp delivery and template fees) are billed separately from AI usage either way. If the key fails on a call that would have run on it, whether the platform falls back to credits depends on your Fallback to credits toggle — see the “Fallback to credits” toggle described above on this page.

When exactly do AI replies cost 0 credits? BYOK is per-account. AI replies cost 0 credits only when this account has connected its own Anthropic key. Without its own key, replies spend credits at the normal rate.

Does BYOK affect message delivery or channel reliability? No. BYOK only changes how your bot’s AI calls are billed. Everything else — message routing, channel infrastructure, follow-ups, campaigns and Broadcasts — works exactly the same.

See also: Why Max Is the Best Value · Setting Up Your AI Bot · AI Agents