Elyra 0.9.22: The One Where the Waiting Ends

Some releases are about code we wrote. This one is mostly about doors other people finally unlocked — and the quiet advantage of having built Elyra so that when a door opens anywhere, you're already standing in front of it. Version 0.9.22 is a model registry release, and it carries two stories worth telling.

Why this release

If you've followed along, you know the saga. Claude Fable 5 launched behind subscription-only surfaces, then vanished entirely under an export-control suspension ("Please use Opus 4.8"), then returned, then became API-reachable — a months-long on-again-off-again that we tracked release by release. And Sakana's Fugu Ultra launched with impressive claims and a hard wall for us: 403 · REGION RESTRICTED, not available in the EU/EEA while GDPR compliance was pending.

Two frontier models, two different kinds of "no." As of this week, both have a "yes" — and neither required us to write a feature. That's the story.

1. Claude Fable 5, for everyone, per token

claude-fable-5 is now in Elyra's model registry via the Anthropic API directly:

elyra --model claude-fable-5

or mid-session, like switching gears:

/model fable

What you're getting: adaptive thinking with a new max level (Elyra's thinking dial maps onto it automatically — Shift+Tab all the way up now reaches max), a 1M-token context window, text + image input, at $10/$50 per million tokens.

That price is also why we made a deliberate choice: the Anthropic default stays claude-sonnet-5 ($2/$10). Fable 5 is a 5x premium — the model you reach for, not the model you idle on. If the course taught anything (Chapter 13, "Watching the Meters"), it's that model choice is a per-task decision. Fable is for the architecture session, the gnarly debugging afternoon, the audit of your own instruction files. Sonnet remains the daily driver.

One practical note from the community's collective testing: Fable rewards being given real, verifiable work — it measures before it declares done. Point it at something hard.

2. Fugu Ultra: around the wall, not through it

Sakana's fugu-ultra — the multi-agent orchestration system wearing a single model API, 1M context, reasoning + vision — is still region-blocked for direct API access in the EU/EEA. That hasn't changed.

What changed: it's now served by OpenRouter and the Vercel AI Gateway, at $5/$30. And Elyra already speaks to both. So:

/login          # add an OpenRouter API key if you haven't
/model fugu

That's the whole migration guide.

This is worth pausing on, because it's the provider-agnostic argument in its purest form. We didn't negotiate access. We didn't write an integration. A model that was unreachable from this side of the Atlantic became reachable because aggregators exist, and Elyra treats aggregators as first-class providers like everyone else. A vendor-locked agent waits for its vendor. A provider-agnostic one inherits every door any provider opens.

(The direct Sakana API remains on our watchlist for when EU access opens — first-class provider support is a checklist item, not a rewrite.)

3. Housekeeping: defaults that follow the field

The registry refresh also surfaced newer defaults, applied where the target actually exists on each provider:

Provider Before Now azure-openai-responses gpt-5.4 gpt-5.6 github-copilot gpt-5.4 gpt-5.5 zai glm-5.1 glm-5.2 vercel-ai-gateway zai/glm-5.1 zai/glm-5.2

GLM 5.2 regulars will be pleased — it's now the out-of-the-box experience on both routes. (Google stays on gemini-3.1-pro-preview for now: the stable id isn't served by the Google provider yet, and we don't ship defaults that fail preflight.)

Plus new registry arrivals for the curious: Gemini 3.6 Flash, Poolside Laguna S 2.1, Tencent HY3, and Meituan LongCat 2.0. Run elyra doctor preflight after updating — it should come back all green, because we do.

The takeaway

Elyra's job isn't to have opinions about which lab is winning this month. It's to make sure that whoever is winning, you're one /model away. This release is what that looks like in practice: two long-blocked frontier models, zero new code, available this morning.

Get it

npm install -g @elyracode/coding-agent@latest
 - or simply use elyra update

Then try the two newcomers on something real:

/model fable    # the architecture session
/model fugu     # the multi-step marathon

Full details in the changelog. And if you're new here: the Elyra Course teaches when to reach for which model — Chapter 13 exists for exactly weeks like this one.