Claude Haiku 5.5 in a Discord Ticket Bot: What Changed

A model change is usually invisible to the people running a server. This one is worth five minutes, because it changed a number most admins care about: how many tickets the AI can answer before the month's allowance runs out.

What changed on October 9?

The model that writes your AI's replies. Release v2.5.16 made Anthropic's Claude Haiku 5.5 the default for every server on our AI. The previous default, Claude Haiku 4.5, was retired from the model picker the same day.

If your server had picked a Sonnet model by hand, it stayed on that model. If it connects its own AI provider, nothing changed for it at all.

Do you need to do anything?

No. The move happened on our side. The bot was not re-invited, no command had to be run, and no setting was reset.

Two places show which model your server is on:

Both are also on the dashboard, on your server's AI page. Changing the model needs Manage Server.

Why does the AI answer more tickets now?

Because each ticket costs fewer tokens. Every plan comes with a monthly allowance of AI tokens. A token is the unit the AI provider bills in, and a ticket uses some for every reply. The allowance did not change. What a ticket costs did.

We measure the cost of an average ticket on each model from real conversations. On Haiku 5.5 it is a small fraction of what it was on the old default, so the same allowance stretches much further.

Read those as estimates. A server with long conversations uses more per ticket and a server with short ones uses less. The numbers move as the measurement updates, so the pricing page prints the current figures and this article gives floors.

It is worth being exact about what the figure is. It is not a cap on tickets. A member can always open a ticket and your staff can always answer it, on every plan. The figure is how many tickets the AI answers before it has to stop for the month.

Are the answers better, or only cheaper?

Better too, and we will not put a number on it. In our release note we said the new model understands support tickets better and uses fewer tokens for each reply. That is our own view from running it before the release. It is not a published benchmark, and you should judge it on your own tickets.

What decides answer quality has not changed, and it is still not the model:

If you have not taught the AI since setup, training it will do more for your answers than any model change.

When is a Sonnet model still worth it?

When tickets need judgment and you have allowance to spare. The picker still offers Claude Sonnet 5.5, Claude Sonnet 5 and Claude Sonnet 4.6. They reason further on tickets where several rules have to be weighed together. They also cost ten times as much per ticket as the default, or more.

The gap between the two is wider than it was. On the old default a Sonnet model cost a few times more. Now it is ten times or more, because the default got so much lighter. A server that moved to Sonnet earlier in the year should look at the picker again. The confirm screen shows tickets a month before and after any switch, and choosing a model covers the decision in full.

What did not change?

Everything around the reply.

What this does not mean

The honest list.

When the allowance does run out, tickets keep working for your staff and the AI starts again when it resets. A token pack tops it up without a plan change. Servers on Pro can also connect their own AI provider, which takes replies off the plan's allowance altogether.