One New Star and a Steady Compass: Elyra 0.9.34

This is a small release with exactly one headline, and the headline earns it: Claude Fable 5.1 is here, and Elyra has it on day one. But the more interesting story might be what we didn't change — so let's talk about both.

Why releases like this exist

Elyra ships its model registry inside the package. No phone-home, no remote config, no surprise model swaps behind your back — the models your Elyra knows are the models it shipped with. That's a deliberate choice: it keeps offline mode honest and your setup reproducible.

The tradeoff is that when something new lands upstream, we cut a release. So when a new Anthropic flagship appears, you get a version bump whose entire job is: update, and it's in your /model picker. This is that release.

Claude Fable 5.1

Anthropic's newest flagship, available in Elyra everywhere it exists upstream:

/model fable
  claude-fable-5-1        Claude Fable 5.1     anthropic     $10/$50
  anthropic.claude-fable-5-1                   bedrock (us)
  global.anthropic.claude-fable-5-1            bedrock (global)
  anthropic/claude-fable-5.1                   gateways

The numbers worth knowing: 1M token context, 128K output, and adaptive thinking with a new max effort level above xhigh — the model decides when to think, and you can now tell it to really lean in. Cache reads at $0.25 per million make long agentic sessions far cheaper than the sticker price suggests.

At $10/$50 this is a premium instrument. Which brings us to the decision part.

Why the default stays Sonnet

Every time a shiny new flagship lands, there's a temptation to crown it the default. We didn't. Elyra's Anthropic default remains claude-sonnet-5 — and that's a philosophy, not an oversight.

A default model is what runs when you haven't expressed an opinion: quick questions, routine edits, the everyday middle of the work. That job description calls for the best value, not the best ceiling. Sonnet 5 at $3/$15 does everyday agentic coding brilliantly. Fable 5.1 at $10/$50 is what you reach for when the problem deserves it.

The good news: with smart routing you don't have to choose. Pin Fable 5.1 to your deep tier and let the router escalate only when a turn actually warrants it:

/route
Next turn: tier=deep (architecture question spanning 14 files)
  cheap    → gemini-3.7-flash
  standard → claude-sonnet-5
  deep     → claude-fable-5-1

The expensive model works the hard problems; the sensible model handles the rest. Your invoice reflects the difference.

Also in the registry

GLM-5.3 on HuggingFace — the new GLM now reachable through HF inference too
A wave of OpenAI :batch variants on the gateways — half-price siblings for gpt-4o through gpt-5.6. If you run budgeted /goal loops with the goalBatchModels setting, your unattended work just got a wider discount aisle
New arrivals: Qwen 3.8 Flash Next, Inception's Mercury 2.5 preview, IBM Granite 4.2, and a MiMo "ultraspeed" variant

Get it

elyra update

Then /model fable if today's problem deserves the good stuff — or change nothing, and enjoy knowing your defaults were chosen on purpose.