Some releases you plan for weeks. This one happened in an afternoon. NVIDIA's neighbors over at Anthropic shipped a brand-new model generation, and a few hours later Elyra could already drive it. That's the whole story of 0.9.6 — small, fast, and right on time.
Anthropic introduced Claude Fable 5, the first of their new "Mythos-class" models, calling it the most capable model they've ever made generally available. It's a genuine generational step, and 0.9.6 wires it straight into Elyra on the direct anthropic provider.
What you're getting:
A one-million-token context window. New territory for Claude — you can hand it an enormous amount of code or documents in a single pass.
128k max output. Long refactors and big files don't get cut short.
Adaptive thinking, all five gears. Fable 5 decides how hard to think, and Elyra's thinking levels map cleanly across its full effort range — minimal → low, low → medium, medium → high, high → xhigh, xhigh → max.
Vision built in. Text and images, so screenshots and diagrams are fair game.
Using it is the usual one step. Inside a session:
/model
Search for Claude Fable 5, select it, keep going. Same conversation, new generation of brain.
Here's the unglamorous engineering bit worth a sentence, because it's the difference between "listed" and "works." Fable 5 is adaptive-only — it doesn't accept the older budget-based thinking that earlier Claude models used. Elyra's logic for "does this model use adaptive thinking?" only recognized the Opus and Sonnet families, so Fable 5 would have tried the wrong path and failed the moment you turned thinking on.
0.9.6 teaches that check about the Mythos class. We verified it end to end: a real reasoning request through Elyra's provider, thinking active, correct answer, no errors. Invisible to you — which is exactly how a fix should feel.
Now the honest conversation, because Fable 5 is powerful and premium, and you should know that going in.
Model Input / 1M Output / 1M Claude Opus 4.8 $5 $25 Claude Fable 5 $10 $50
Fable 5 costs roughly twice as much as Opus 4.8 on both ends. Paired with a 1M-token context, the bill can climb fast if you're careless — fill that window and a single turn is a lot of input tokens at $10 per million.
The good news is that Elyra is built to keep this in check, and there are three habits that make Fable 5 affordable:
1. Reserve it for the hard part. You don't have to run the whole session on it. Use /model to bring Fable 5 in for the genuinely difficult reasoning, then switch back to something cheap — Haiku, MiniMax M3, Nemotron — for the mechanical edits and file reads. The expensive brain for the expensive decisions; nothing more.
2. Lean on caching. Fable 5's cached input reads are $1 per million — a tenth of fresh input. When you keep working over the same large context, those reads are cheap. Pin the files that matter with /pin so the stable context stays put and gets reused instead of re-sent.
3. Watch the meter. /cost shows your token usage and an estimated spend at any time. With a model in this price class, glancing at it now and then turns "surprise" into "decision."
The short version: Fable 5 is a scalpel, not a default. When the task is big and hard and needs a million tokens of context held in one mind, it earns its price. For everything else, Elyra makes it trivial to reach for something lighter.
npm install -g @elyracode/coding-agent
Or, from inside a running session:
/update
Then /model, pick Claude Fable 5, and point it at something worthy of it.
Added: Claude Fable 5 on the direct anthropic provider — Mythos-class, 1M context, 128k output, adaptive thinking, vision.
Fixed: adaptive-thinking detection now covers Mythos-class models, so Fable 5 works correctly the moment thinking is on.
Cost note: it's premium ($10 / $50 per 1M, cached reads $1). Use /model, /pin, and /cost to wield it deliberately.
The flagship gets the headlines. Knowing when not to use it is what keeps the bill sane. Elyra gives you both — the new generation, and the dial to spend it wisely.
Happy building.