Hey Forge developers! 
We have an exciting announcement: the Forge LLMs API is now available in Preview and can be enabled on production environments today.
The Forge LLMs API lets you call Claude models (Haiku, Sonnet, and Opus) directly inside your Forge app. No external API keys, no model infrastructure setup, no data leaving the Atlassian platform. That means your LLM-powered features can still qualify for the Runs on Atlassian badge.
Getting started takes just a few lines of code:
import { chat } from '@forge/llm';
const response = await chat({
model: 'claude-sonnet-4-5-20250929',
messages: [{ role: 'user', content: 'Summarize this Jira issue thread.' }],
});
A few things worth knowing about the Preview release:
-
Billing starts June 1, 2026. LLM usage is charged to the app developer based on token credits. There’s no free quota, so check the pricing docs before you ship.
-
Three Claude models available. Haiku (fast & lightweight), Sonnet (balanced), and Opus (powerful reasoning). Mix and match per request based on your cost/latency needs.
-
Examples and tutorials are ready to go. From a Confluence AI assistant to agentic web triggers with function calling.
We’re really excited to see what you build with this. If you ship something cool, share it here in the community 
Read the full announcement and get started →
Sounds great!
The llm module is required to enable Forge LLMs. When you add the llm module to an app’s manifest.yml, it triggers a major version upgrade and requires administrators of existing installations to review and approve the update.
Will Rolling Releases be compatible with introducing Forge LLM? Since the llm module isn’t a permission, I’m wondering if we could introduce Forge LLM without having version fragmentation, even when using Rolling Releases.
Great news!
I haven’t read up on the documentation yet, and at this time I’m not sure if we are going to need it, but will the LLM’s also be available to be invoked through a Remote? Like accessing the Forge Secrets KVS I would see a reason that customers would like us to use Atlassian’s models/LLM interface instead of us relying on outside LLMs when interfacing through remotes.
Will Rolling Releases be compatible with introducing Forge LLM?
Rolling Releases will honour the Admin consent screen, where we show admins that the new major version includes LLM usage.
The current iteration of Rolling Releases does not support Forge LLMs, so will currently block the upgrade.
Under the future rolling release we aim to remove all blocked upgrades, and as stated before, we intend to honour the admin consent screen.
Any plans to keep pace with model releases going forward? Already two versions behind on Opus.
And also when other providers?
Hey Nathan, yes we’ll be releasing support for Opus 4.7 and Opus 4.8 soon.
Also happy to take feedback on which family of models people would like to see next. Would people like to see even cheaper open weight models? Or different capabilities like multimodal or embeddings? Please let us know.
@AdamMoore, for our use-case in our new app we are yet unsure about throughput. We tested with all 3 current models and tend to use the mid-tier model (Sonnet) as the best compromise of cost and quality.
Ask: currently the API’s list() returns name and status for each model. We’d also like something like tier (high, mid, low), so that in future we can dynamically select from the available models as we see fit, instead of hardcoding a specific model, which I understand is not future-proof because available models might change intermittently (or?)
You probably need to use something like OpenRouter and simply let developers choose.
One model provider with weeks/months of delays from version release to Forge LLM availability is a recipe for perpetually uncompetitive marketplace apps.
You can see that by the time Atlassian adds a new model it’s already extremely outdated - you’re currently two versions behind yet again!
The internal bureaucracy will never keep pace with manually approving new model releases. Nor do any models in 2026 need to be reviewed before making them available to developers.