Claude Sonnet 5 Is the New Default (And Your Bill Just Went Down)

Some releases add buttons. This one changes a single line of configuration — the default model — and that quiet change is the most consequential thing we've shipped in a while. Here's the story of v0.9.12: why we moved the default to Claude Sonnet 5, how the smart router now uses it, and what it means for your day-to-day work.

The news: Sonnet 5 landed

Anthropic released Claude Sonnet 5 on June 29. We verified it the boring, trustworthy way — straight against Anthropic's own /v1/models endpoint and a live "say pong" round-trip — before touching anything. It's real, it's claude-sonnet-5, and as of this release it's in Elyra's registry.

The specs that matter:

Old default (Opus 4.8) New default (Sonnet 5) Context window 200K 1,000,000 Thinking adaptive adaptive Input text + image text + image Price (per 1M, in/out) premium Opus tier $2 / $10

For comparison, the previous Sonnet (4.5) was $3/$15. Sonnet 5 is cheaper than the Sonnet it replaces while being Anthropic's newest model — and it carries a full million tokens of context. That combination is exactly why it deserves to be the default.

Why change the default at all?

A default model is a quiet promise: "when you don't tell me otherwise, this is what runs your work." For a long time that promise was Opus — the heavyweight. Opus is wonderful, but it's expensive, and most of what a coding agent does all day isn't heavyweight: reading files, running greps, writing a test, applying a straightforward edit. Paying Opus prices for grep is like taking a limousine to the mailbox.

Sonnet 5 changes the math. It's strong enough to be the everyday default, priced so that an afternoon of agent work doesn't make you wince, and roomy enough (1M context) to hold a real chunk of your codebase in one session. So we moved the default — for both the Anthropic API and Amazon Bedrock (us.anthropic.claude-sonnet-5).

You don't have to do anything. Update, and your next session quietly runs on a newer, cheaper, longer-context model.

How: the smart router does the interesting part

Here's the piece I'm most happy about. Changing the default is half the story; the other half is when to reach for more power.

Elyra's smart router classifies each turn into a tier — fast, balanced, or powerful — based on what's actually happening (read-only tools, file edits, errors, long conversations, complex keywords). In v0.9.12, the powerful tier is now pinned to claude-sonnet-5 by default.

So the flow becomes:

Mechanical turns (reading, searching) → a cheap fast-tier model.
Ordinary work → balanced.
The genuinely hard turns — a refactor, a tricky bug, an architecture decision → Sonnet 5, with its big context and adaptive thinking.

You can preview exactly what it'll do:

/route

Next turn: powerful — "complex keywords in user message" fast: z-ai/glm-5.2 balanced: claude-sonnet-5 powerful: claude-sonnet-5 (pinned)

And watch the effect on spend:

/cost
... burn rate: $0.40/hr

It's a default, not a cage

This is the part that keeps it friendly. The pin is a default, and it bends to you:

Your settings win. If you set your own smartRoutingModels, your choice overrides the default, per tier:
```
"smartRoutingModels": {  "powerful": "claude-opus-4-8"   // prefer Opus for the hard stuff? done}
```
It fails safe. The pin only takes effect if claude-sonnet-5 is actually available to you. If you don't have Anthropic configured, the router quietly falls back to its normal heuristic ranking among the models you do have. No broken sessions, no "model not found" surprises. We resolve the pin against your real model registry, so an absent model simply isn't pinned.

That's the design rule we hold to: smart routing decides the tier, you decide the model — and sensible defaults fill the gap when you haven't.

A small note on what didn't land

While we were in the model registry, we also re-checked Claude Fable 5. Still export-control blocked for us (404: not available, use Opus 4.8) — it's listed in Anthropic's catalog but gated. Sonnet 5 is not affected by that restriction, which is part of why it's a safe model to build a default around. We'll keep watching Fable; the moment access opens, it already resolves from our registry with no code change.

The takeaway

v0.9.12 has an almost-empty feature list and an outsized effect. The default model is now newer, cheaper, and roomier; the smart router reaches for it precisely when the work is hard and rations cheaper models the rest of the time. The net result is the nicest kind of upgrade — you do nothing, your sessions get better, and your invoice gets smaller.

npm install -g @elyracode/coding-agent@0.9.12
elyra            # already running on claude-sonnet-5
/route           # see the tiers

That's the whole release: one default, chosen carefully, doing a lot of quiet good.