Claude Pro Default Model: Opus 5.5 in Claude Code (2026)

If you open Claude Code on a Pro plan today, you are no longer talking to Sonnet by default. Version 2.1.280 of the CLI shipped two lines that together changed the experience for millions of subscriptions: one added a new model, the other moved the goalposts under an existing plan tier. Anthropic published Claude Opus 5.5 on September 22, 2026, and the changelog entry is unusually blunt about who gets switched over — the default model on Pro and Team Standard moved from Sonnet to Opus, joining the tiers that were already on Opus.

This is the article for anyone who noticed the answers got longer, the session got slower, or the plan meter started draining differently and did not know why. We will cover what actually changed, what it means for cost and latency on a subscription, how to set the default model back with /model, /config, or settings.json, how to decide between Opus 5.5 and Sonnet for coding work, and where the "Sonnet 5.5 coming soon" claim actually comes from. At the end there is a section on the router-plus-DeepSeek route for people whose real problem is not model quality but quota — with an honest account of what you give up.

Quick answer

• Current default on Pro and Team Standard: Opus (Claude Code 2.1.280), matching Max, Team Premium, and Enterprise. Opus 5.5 is the model behind that alias.
• Switch for this session: type /model sonnet. Run /model with no argument to open the picker; pressing s there switches "for this session only".
• Set a permanent default: choose it in /config, or put "model": "sonnet" in ~/.claude/settings.json (user scope) / .claude/settings.json (project scope).
• Per-launch override: ANTHROPIC_MODEL=sonnet claude — the docs say these env settings "apply only to the session you launch with them".
• On API pricing, Opus 5.5 is exactly 2× Sonnet 5: $4/$20 vs $2/$10 per million input/output tokens.
• Sonnet 5.5: Anthropic's own launch post says "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks." No date.

What actually changed in Claude Code 2.1.280

The official changelog is two bullet points, and it is worth reading them exactly rather than through a summary:

## 2.1.280
- Added Claude Opus 5.5 (`claude-opus-5-5`), now the default Opus model —
  1M context, $4/$20 per Mtok with $0.20/Mtok cache reads
- Changed the default model on Pro and Team Standard plans from Sonnet to
  Opus, matching Max, Team Premium, and Enterprise

Two details in that text matter for precision. First, the default-plan line says "from Sonnet to Opus" — it names the alias, not the version. The version comes from the bullet above it: Opus 5.5 is "now the default Opus model", so the Opus that Pro lands on is Opus 5.5. Second, the changelog does not stamp a date on 2.1.280; the dated anchor we can verify is Anthropic's news post, which puts Claude Opus 5.5 at September 22, 2026, with "Input and output tokens are $4 and $20 per million."

The model-config docs confirm the scope from the other direction: Opus 5.5 is the default for Pro, Max, Team, Enterprise, the Anthropic API, AWS, Bedrock, and GCP. Microsoft Foundry is the exception and still defaults to Sonnet 4.5.

So if you are on Pro and never touched a model setting: you were on Sonnet before 2.1.280, and you are on Opus 5.5 after it — automatically, with no prompt asking you to opt in.

How to change the default model back

Four ways, from most temporary to most permanent. All of them are documented; pick the one matching how sticky you want the change to be.

1. /model — mid-session, fastest

/model sonnet

The interactive-mode docs describe /model as "Switch models without clearing your prompt", and the model-config page adds: "/model <alias|name> to switch immediately, or run /model with no argument to open the picker." In the picker, pressing s applies the switch "for this session only" instead of persisting it.

2. /config — persistent, via the UI

The costs page is explicit about the division of labour: "Use /model to switch models mid-session, or set a default in /config." If you would rather click than edit JSON, this is the intended path.

3. settings.json — persistent, per scope

The field is literally model, and the documented example value is an alias:

{
  "model": "sonnet"
}

Where you put that file decides who it affects. The documented scopes:

Aliases are stable across releases: opus resolves to the latest Opus, sonnet to the latest Sonnet, haiku to the fast/efficient option. That is the reason for pinning an alias rather than a dated model ID — when Sonnet 5.5 ships, "model": "sonnet" follows it, and "model": "claude-opus-5-5" does not move.

4. Environment variables — one launch at a time

ANTHROPIC_MODEL=sonnet claude

Documented as ANTHROPIC_MODEL=<alias|name>, with the caveat that these "apply only to the session you launch with them." Handy for scripts and CI, useless as your daily default. The model-config page also lists ANTHROPIC_DEFAULT_OPUS_MODEL and ANTHROPIC_DEFAULT_SONNET_MODEL, which change what those aliases resolve to — read the page before relying on them, since their exact semantics are beyond what we verified here.

Verifying which model you are actually on

Run /model with no argument — the picker shows the active selection. For the metering side, the costs page documents /usage ("The Session block in /usage shows API token usage and is intended for API users"), /usage-credits for subscription users — run it "after signing in with your claude.ai subscription through /login" — /context to "see what's consuming space", and /insights for a report on how you work.

Opus 5.5 vs Sonnet for coding: the honest comparison

Here is everything we could verify side by side, followed by the judgement calls, which are ours and clearly labelled as such.

The final row is Anthropic's own phrasing from the models overview, and it is the most useful signal in the table: Opus is framed around long-running agentic work; Sonnet around speed. Everything below is our reading, not the vendor's.

Where Opus 5.5 earns its keep. Long agentic sessions are where a plan default matters: multi-file refactors, migrations where the agent has to hold a mental model of the whole tree, code review that has to notice the thing that breaks two directories away. Those are exactly the "long-running agentic coding" jobs the model is described for, and a weak default there costs you more than tokens — it costs the debugging you do afterwards.

Where it is overkill. Single-file edits, renaming, writing a test for a function you already understand, boilerplate, doc comments, "explain this stack trace". On a subscription, an expensive model doing a cheap task is not a quality gain; it is latency. Anthropic's model docs put 128K max output on both, and the same 1M context on both, so you are paying for reasoning depth you may not be using.

The lever most people skip: /effort. Documented on the costs page as a way to "reduce costs by lowering the effort level" for "simpler tasks where deep reasoning isn't needed." If your complaint about the new default is slowness rather than quota, effort is often the better knob than switching models outright.

What we could not verify, and will not pretend otherwise: how a Pro subscription meters Opus minutes against Sonnet minutes. Anthropic publishes no per-model consumption multiplier in the docs we checked, so any blog that tells you "Opus burns your Pro limit 3× faster" is making it up. Check your own meter with /usage-credits and /usage over a week and decide from data.

What about Sonnet 5.5?

Anthropic's Opus 5.5 launch post contains the sentence everyone is quoting: "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks."

That is the entire claim. It is official, it is recent, and it is dated September 22, 2026 — so "the coming weeks" nominally runs into late October 2026. What it is not is a date, a price, or a spec sheet. There is no published benchmark, no confirmed context window, and no statement about whether Sonnet 5.5 becomes the Pro default when it lands or whether Opus 5.5 stays.

The practical consequence for you: if you pin "model": "sonnet" in settings.json, an alias pin means you inherit Sonnet 5.5 automatically the day it ships — which is either exactly what you want or a surprise mid-sprint, depending on how much you like your defaults moving. If you want reproducibility, pin the dated model ID instead (claude-opus-5-5 is the form shown in the changelog) and change it deliberately.

Quota anxiety is a different problem

Read this section only if your actual complaint is that the plan runs out. Switching Pro's default back to Sonnet usually helps, because the same work costs fewer plan units — but it is still a flat meter that resets on someone else's schedule.

The structural alternative is to stop paying for the meter and pay for tokens instead. Two documented shapes:

1. A local router. Keep Claude Code as the harness and point it at a router so you can swap models per task without editing env vars repeatedly. We wrote the full walkthrough: Claude Code router tutorial.
2. A different harness. If the terminal agent itself is negotiable, the same DeepSeek endpoint drops into other tools — see DeepSeek harness setup and, for editor users, using DeepSeek in Cursor.

DeepSeek's Anthropic-compatible endpoint is metered per token at published API rates — deepseek-flash from $0.15 per million input tokens off-peak, with the pricing table and the two-environment-variable wiring spelled out in our Codex usage dashboard piece.

The honest trade. This is not "free and just as good." A router hop adds a failure surface; the DeepSeek models are different models, and on long-horizon agentic refactors a frontier Opus-class model is doing something they may not. What you are buying is a meter you control at a price that makes a heavy session cost cents instead of a plan upgrade, and the ability to send a rename job to a cheap model without touching your default. If your work is mostly short, well-scoped edits, that trade tends to be favourable. If it is 40-minute autonomous migrations, stay on Opus and pay for it. Our Claude Code project examples post is a decent proxy for which category you are in.

And if you would rather not configure anything at all: the DeepSeek App Builder runs the same models in a browser with no setup — one prompt to a working React app, with the source visible.

FAQ

What is Claude Code's default model on Pro in 2026? Opus — specifically Claude Opus 5.5 (claude-opus-5-5) — as of CLI version 2.1.280, which changed the Pro and Team Standard default from Sonnet to Opus so it matches Max, Team Premium, and Enterprise. Microsoft Foundry still defaults to Sonnet 4.5.

How do I change Claude Code's default model? /model sonnet switches mid-session; /config sets a default through the UI; "model": "sonnet" in ~/.claude/settings.json makes it persistent for all your projects, or in .claude/settings.json for one repo. ANTHROPIC_MODEL=sonnet claude overrides a single launch. The changelog line itself does not name 5.5 — it names the Opus alias, and the adjacent bullet makes Opus 5.5 the default Opus.

Is the 2.1.280 change permanent for me? It is the new default until you change it, and there is no announced revert. The default-plan assignment is documented in Claude Code's model-config docs, so treat it as the vendor's chosen configuration rather than a temporary rollout.

Does Opus 5.5 cost more than Sonnet on a Pro subscription? On the API, unambiguously: $4/$20 versus $2/$10 per million tokens, exactly double. On a Pro subscription, Anthropic does not publish a per-model consumption multiplier in the docs we checked, so nobody can honestly quote you a ratio. Watch your own meter via /usage and /usage-credits.

When is Claude Sonnet 5.5 released? Not announced. Anthropic's Opus 5.5 post from September 22, 2026 says "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks" — a window, not a date, and with no specs attached. Pin the sonnet alias if you want to inherit it automatically; pin a dated model ID if you want your runs to be reproducible.

Which should I use for coding, Opus 5.5 or Sonnet? Anthropic's own descriptions are the best starting line: Opus 5.5 is "for long-running agentic coding and knowledge work", Sonnet 5 is "the best combination of speed and intelligence". Multi-file refactors, migrations and review where the agent must reason across the tree favour Opus; single-file edits, tests for code you already understand, and anything where you are waiting on the answer favour Sonnet, or Opus at a lower /effort.

Related reading

• Claude Code router tutorial — swap models per task through a local router instead of editing env vars
• DeepSeek harness setup — wire a terminal agent to a per-token endpoint end to end
• Using DeepSeek in Cursor — the same idea inside an editor
• Codex usage dashboard — the other agent's answer to "how much have I burned", plus DeepSeek pricing table
• Claude Code project examples — the kinds of jobs where an Opus-class default pays for itself
• DeepSeek App Builder — zero-setup, browser-side, source visible