Elyra · · 3 min read

When a Model Vanishes: Meet Mistral Medium 3.5

A few days ago we shipped same-day support for Claude Fable 5. This morning it returns a 404. That's the whole argument for the boring feature — switching models in one command — and a good excuse to talk about a workhorse worth switching to.

When a Model Vanishes: Meet Mistral Medium 3.5

A few days ago we shipped same-day support for Claude Fable 5 — Anthropic's new Mythos-class flagship, a million tokens of context, the whole event. We wrote a warm post about it. People tried it.

This morning it returns a 404.

The model still appears in Anthropic's catalog, but actual requests now come back with a polite redirect: "Claude Fable 5 is not available. Please use Opus 4.8." Access has been pulled. No drama, no announcement in your terminal — just a model that worked yesterday and doesn't today.

It's a useful reminder of something easy to forget: you don't own the models you build on. You rent them, and the landlord can change the locks. Which is exactly why the boring, unglamorous feature — being able to switch models in one command — turns out to be the important one. So let's talk about a model worth switching to.

Mistral Medium 3.5

While the flagship drama played out, Mistral quietly has a model sitting in Elyra's registry that deserves more attention than it gets. Here's what it actually is:

  • 256k token context — not a million, but more than enough for real codebases and long documents.

  • Reasoning built in — it thinks through hard problems, and Elyra's thinking levels map onto it.

  • Vision — text and images, so screenshots and diagrams work.

  • Genuinely affordable — about $1.50 per million input tokens and $7.50 per million output. For a capable reasoning-and-vision model, that's a comfortable middle.

It's the kind of model that doesn't top a leaderboard but quietly handles most of what you throw at it without emptying your wallet — a workhorse, not a show pony.

Using it takes two lines

export MISTRAL_API_KEY=...
elyra

Then inside the session:

/model

Search for Mistral Medium 3.5, select it, keep going. There are two IDs for the same model — mistral-medium-3.5 (a stable alias) and mistral-medium-2604 (date-stamped) — use the alias and you won't have to think about it.

Prefer headless? It's a one-liner:

elyra -p "Audit src/billing for edge cases" \
  --provider mistral --model mistral-medium-3.5

The part that matters today: switch without flinching

Here's where the morning's 404 becomes a teaching moment instead of an outage. If Fable 5 had been your only plan, today would hurt. In Elyra, it's a shrug:

/model

…pick Mistral Medium 3.5, and your session keeps going — same conversation, same files, same context. Or make it permanent for a repo by pinning it as your balanced tier in .elyra/settings.json:

{
  "smartRouting": true,
  "smartRoutingModels": {
    "balanced": "mistral/mistral-medium-3.5"
  }
}

Now smart routing reaches for Mistral on the everyday turns, and you've insulated that project from any single provider's decisions. Want to confirm a model is actually reachable before you rely on it? That's what the new probe is for:

elyra doctor models --provider mistral --model medium

It makes a real call and tells you the truth — which, as today proves, is not always what the catalog claims.

The quiet lesson

We're not here to dunk on Anthropic — restricting a powerful new model is a defensible call, and Opus 4.8 is excellent. The point is bigger than any one model: in a world where capabilities appear and disappear on someone else's schedule, the durable advantage isn't picking the "best" model. It's being able to change your mind in five seconds when the best model changes on you.

Mistral Medium 3.5 is a great answer to "what now?" — affordable, capable, multimodal, and available today. But the real answer is the muscle Elyra gives you: thirty-plus providers, one /model, and never being held hostage by a single 404.

Models come and go. Your agent shouldn't have to.

Happy building.