TL;DR: Claude Code now defaults to Claude Opus 5 on Max, Team Premium, Enterprise, and pay-as-you-go API accounts — only Pro and Team Standard still default to Sonnet 5. On the API, Opus 5 costs exactly 2.5x Sonnet 5 per token ($5/$25 vs $2/$10 per million), roughly $1.63 vs $0.65 for a heavy agentic session. Pinning Sonnet back is one command.

	Stay on Opus 5 (default)	Pin Sonnet 5	Split with opusplan
Best for	Max/premium subscribers; hard multi-file work	Pay-as-you-go API users doing routine coding	API users who want Opus judgment, Sonnet execution
Cost per heavy session (200K in / 25K out)	~$1.63 API list	~$0.65 API list	between the two
Monthly at 5 sessions/day	~$179	~$72	between the two
SWE-bench Verified	96.0%	85.2%	n/a
The catch	2.5x the token bill; thinking-on-by-default inflates output	~11-point benchmark gap shows on gnarly refactors	Plan/execute boundary isn’t always where you’d draw it

Honest take: If you pay per token, pin Sonnet 5 today — /model sonnet, done — and step up to Opus 5 manually for the 20% of sessions that are genuinely hard. If you’re on Max or another seat plan, leave the Opus 5 default alone; you already paid for it, and the only thing downgrading buys you is slower usage-cap burn.

When did Claude Code switch its default model to Opus 5?

In three steps between June and August 2026, each tied to a specific Claude Code release (dates below are the npm publish timestamps for @anthropic-ai/claude-code, so they’re exact):

Version	Published	What changed
2.1.197	Jun 30, 2026	Sonnet 5 launches and becomes the default model in Claude Code
2.1.219	Jul 24, 2026	Opus 5 ships as “the default Opus model” — from this point, plans that default to the Opus family (Max, Team Premium, pay-as-you-go API) resolve to Opus 5
2.1.251	Aug 28, 2026	”Changed the default model for seat-based Enterprise subscriptions to Opus 5, matching other premium plans” (changelog wording)
2.1.257	Sep 1, 2026	Fable 5.1 arrives as the new fable/best alias target; /model gains a session-only switch

So the September chatter about Claude Code “now defaulting to Opus 5” is really the tail end of a rollout that finished on August 28 with seat-based Enterprise. There was no blog post for that last step — it shipped as one line in the CHANGELOG.md of the anthropics/claude-code repository. If your enterprise org’s token bill jumped in late August without anyone changing config, this is why.

The current default by plan, verified September 13, 2026 against Anthropic’s Claude Code model-config docs:

Plan / provider	Default model
Max, Team Premium, Enterprise, pay-as-you-go Claude API	Opus 5
Pro, Team Standard	Sonnet 5
Bedrock, Google Cloud Agent Platform, Claude Platform on AWS	Opus 5
Microsoft Foundry	Sonnet 4.5

What does Opus 5 cost vs Sonnet 5 in a real Claude Code session?

2.5x, at every tier of the pricing table. Verified against Anthropic’s official pricing page on September 13, 2026:

Per million tokens	Claude Opus 5	Claude Sonnet 5	Ratio
Input	$5.00	$2.00	2.5x
Output	$25.00	$10.00	2.5x
Cache read	$0.50	$0.20	2.5x
Batch input / output	$2.50 / $12.50	$1.00 / $5.00	2.5x

One pricing footnote matters here: Sonnet 5’s $2/$10 was announced at launch as introductory pricing through August 31, 2026, with a scheduled increase to $3/$15 on September 1. That increase was canceled. Anthropic’s pricing docs now state $2/$10 is the standard price, so the 2.5x gap is the durable number, not a promo artifact.

For a heavy agentic session — 200K input tokens, 25K output, the same session shape we’ve used in our other cost breakdowns — at list prices before caching:

• Opus 5: $1.00 input + $0.63 output ≈ $1.63 per session
• Sonnet 5: $0.40 input + $0.25 output ≈ $0.65 per session

At 5 sessions a day, 22 working days a month, that’s roughly $179/month on Opus 5 vs $72/month on Sonnet 5 — a $107/month gap for one developer’s default-model setting. Prompt caching shrinks both bills substantially (Claude Code caches aggressively, and cache reads are 10% of input price on both models), but since the multiplier is identical, the 2.5x ratio survives caching intact.

Two cost traps that make real Opus 5 sessions run past the naive math:

1. Thinking is on by default. Opus 5 runs adaptive thinking unless you disable it, and thinking tokens bill as output tokens — the $25/M rate. A workload migrated from Opus 4.8 (where thinking was off unless requested) produces more output tokens per request at the same per-token price. Anthropic documents this explicitly and tells you to re-baseline cost.
2. You can’t disable thinking at high effort. thinking: disabled above effort high returns a 400 error on Opus 5. The sanctioned cost lever is lowering effort, not switching thinking off.

Is Opus 5 actually better at coding than Sonnet 5 — and by how much?

Yes, by about 11 points on SWE-bench Verified and 16 on the harder Pro variant — a real gap, not a rounding error. The numbers, from Anthropic’s published benchmarks (we verified these when reviewing Opus 5 as a coding backend in July):

Benchmark	Opus 5	Sonnet 5	Fable 5
SWE-bench Verified	96.0%	85.2%	95.0%
SWE-bench Pro	79.2%	63.2%	80.3%

Where the gap actually shows up in Claude Code: long-horizon agent runs that touch many files, refactors where the model has to hold an architecture in its head across dozens of tool calls, and end-to-end feature work. Anthropic’s own Opus 5 notes call out agentic coding and long-horizon tasks as the step-change areas, along with the 1M-token context window that is both default and maximum on Opus 5.

Where it doesn’t show up: the routine 70-80% of daily work — writing tests, fixing a known bug, wiring a CRUD endpoint, renaming things across a module. Sonnet 5 at 85.2% Verified is a stronger model than anything that existed at any price a year ago. Paying 2.5x for the delta on a task Sonnet finishes correctly is pure waste, which is exactly why the default matters: defaults are what you pay when you’re not thinking about it.

How do you pin Claude Code to Sonnet 5?

The one-liner: run /model sonnet inside any session. It switches immediately and saves as your default by writing the model field to ~/.claude/settings.json. On v2.1.257+ you can press s in the /model picker to switch for the current session only, without saving.

Every other supported method, in Claude Code’s priority order (highest wins), verified against the official model-config docs September 13, 2026:

# 1. In-session, persists as default
/model sonnet

# 2. Per-invocation flag (this session only)
claude --model sonnet

# 3. Environment variable (this shell only)
export ANTHROPIC_MODEL=sonnet

# 4. Settings file — ~/.claude/settings.json (global)
#    or .claude/settings.json (per project)
{
  "model": "sonnet"
}

# 5. Default for new sessions, lowest priority (v2.1.236+)
export ANTHROPIC_DEFAULT_MODEL=sonnet

Use the sonnet alias rather than a dated model ID unless you have a reason to pin: on the Anthropic API the alias resolves to Sonnet 5 and will track future Sonnets. /model default clears your override and reverts to the account-type default (which, on premium plans, is Opus 5 again).

Two aliases worth knowing beyond the obvious ones:

• opusplan — Opus in plan mode, Sonnet in execution mode. You get Opus 5’s judgment for the architecture and task breakdown, then Sonnet 5’s rates for the long tool-call grind where most tokens are spent. For API users this is the best default most people never set.
• Per-project settings — a .claude/settings.json with "model": "sonnet" in a repo pins the cheap model for that codebase while your global default stays Opus. Good for the side project that doesn’t deserve Opus rates.

If you route subagents, CLAUDE_CODE_SUBAGENT_MODEL sets the default subagent model separately — we covered the full routing pattern in the subagent model-routing cost guide, and it composes with everything above: Opus orchestrator, Sonnet or Haiku 4.5 ($1/$5) workers remains the cheapest way to keep Opus-level supervision.

Does the Opus 5 default change your bill on Pro or Max?

Not in dollars — subscription plans are flat-rate — but it changes how fast you hit your usage cap. On seat plans the meter is usage, and Opus 5 tokens draw the cap down faster than Sonnet 5 tokens. If you’re on a Max plan and keep hitting limits mid-afternoon, /model sonnet (or opusplan) is the difference between rationing your best model and rationing your workday. Pro users are unaffected by this change entirely: Pro still defaults to Sonnet 5, and Opus 5 on Pro is a manual selection, not the baseline.

Enterprise seat admins got the least warning here. The August 28 change (v2.1.251) moved seat-based Enterprise defaults to Opus 5 via a changelog line, so every seat that never set a model preference silently moved to the more capable, faster-cap-burning model. If your org wants Sonnet as the fleet default, push "model": "sonnet" through managed settings rather than asking every developer to run /model.

And if the answer to “what’s my API budget” is “zero,” the escape hatch is local: Claude Code runs against Ollama-served open-weight models on your own GPU — see the VRAM-based model guide on runaihome.com for what your card can serve, and aifoss.dev’s Ollama review for the server itself. You give up a lot of capability against either Claude model; you also give up the meter.

Who should actually stay on Opus 5?

Anyone whose sessions regularly fail on Sonnet 5 — that’s the whole test. In practice that means: multi-file refactors across large codebases (the 1M default context is doing real work there), long-running autonomous sessions where a mid-task reasoning slip costs you the whole run, architecture and design work, and code review where Opus 5’s bug-catch rate per pass is documented as a step-change area. At $1.63 a session, one avoided hour of untangling a botched agent run pays for a week of the Opus premium. Developers billing $100+/hour should not be optimizing away $107/month if Opus finishes tasks Sonnet fumbles.

The people who should move are the ones for whom the default was never a decision: pay-as-you-go API users running routine sessions on Opus 5 rates simply because July 24 changed what opus means and their plan defaults to Opus. For them, 2.5x is a tax on not having run one command. That’s a verdict with a loser, and the loser is inertia: pin sonnet or opusplan on API accounts; keep Opus 5 on seat plans.

FAQ

Was the switch to Opus 5 announced or silent? Half and half. Opus 5 becoming the default Opus model shipped with the July 24 launch (v2.1.219) and was part of the launch coverage. The August 28 seat-based Enterprise flip (v2.1.251) appeared only as a changelog line — no blog post.

Did Sonnet 5 pricing go up on September 1, 2026? No. The scheduled increase to $3/$15 per million tokens was canceled; Anthropic’s pricing page now lists $2/$10 as Sonnet 5’s standard price. Any cost math you saw quoting $3/$15 “from Sep 1” is outdated.

Does /model sonnet affect other terminals or just this one? It saves as your default for new sessions (it writes ~/.claude/settings.json). For a one-session switch, press s in the /model picker (v2.1.257+) or launch with claude --model sonnet.

What about Fable 5 — isn’t that the best model? Fable 5 ($10/$50, and Fable 5.1 since September 1) is the fable/best alias target, at double Opus 5’s price for a 1.1-point SWE-bench Pro edge. It was never made the plan default, and its credit-metered access makes it a specialty tool, not a daily driver.

Is Opus 5’s thinking-by-default why my output token counts jumped? Very likely, if you migrated from Sonnet 5 or Opus 4.8 defaults. Thinking tokens bill as output on Opus 5 and adaptive thinking is on unless disabled — and disabling it requires effort high or below, otherwise the API returns a 400 error.

Sources

• Claude API pricing — Anthropic Platform docs
• Claude Code model configuration — official docs
• What’s new in Claude Opus 5 — Anthropic Platform docs
• Claude Code CHANGELOG — anthropics/claude-code on GitHub
• @anthropic-ai/claude-code publish history — npm registry
• Introducing Claude Opus 5 — Anthropic announcement

Last updated September 13, 2026. Pricing verified against Anthropic’s official pricing page on September 13, 2026; model defaults verified against the Claude Code model-config docs the same day. Pricing and defaults change frequently — verify before committing a team.

Know which coding tool is worth paying for

Hands-on comparisons of AI coding assistants and what each one costs to run — including the local-model path. Sent only when something changes. Unsubscribe anytime.

Email address Subscribe