↓ Skip to main content
  1. Articles/

Sonnet 5.5 Lands, and Claude Code Now Starts in Auto Mode by Default

·880 words·5 mins·
Florent Clairambault
Author
Florent Clairambault
CTO & software engineer — writing daily about spec-driven development and agentic coding

Sonnet 5.5 Lands, and Claude Code Now Starts in Auto Mode by Default

Two things shipped in Claude Code v2.1.284 on September 28, and the second matters more than the first. The headline is a new model: Claude Sonnet 5.5. The quieter change is that interactive and VS Code sessions now start in auto mode when you haven’t configured a permission mode, across all plans and providers.

What Sonnet 5.5 actually changes
#

Per Anthropic’s Sonnet page and the Claude Code changelog:

  • Price: $2 per million input tokens, $10 per million output tokens, $0.20/Mtok cache reads. That is the same list price as Sonnet 5.
  • Context: 1M tokens.
  • Speed and cost: Anthropic says it is “30% faster and costs up to 30% less for most work” than Sonnet 5, and needs far fewer tokens. Slack’s testing is cited at roughly 14% fewer output tokens.
  • Positioning: “a faster, lower-cost complement to Opus 5.5,” and “the strongest Sonnet yet on long-horizon tasks.”

Read the pricing carefully. The per-token rate did not move. The “30% cheaper” claim comes from token efficiency, which means it depends on your workload. If your agent loops are dominated by tool output and file reads you feed it, your savings will be smaller than for a workload heavy on generated reasoning and code.

The page has a benchmark table, but the figures I could extract did not include specific SWE-bench or Terminal-Bench numbers. I’m not going to quote any. As with every first-party launch table, wait for independent runs. The Terminal-Bench 4.0 leaderboard we have been watching was still empty at our last check.

Sonnet is the default model for most Claude Code sessions, so this is the release that changes your bill and your latency without you doing anything. That is the useful frame: not “new frontier model” but “the workhorse got cheaper to run repeatedly.” For agent teams and subagent fan-out, where you pay for many parallel sessions, per-task cost matters more than per-token price.

Auto mode becomes the starting point
#

Auto mode launched in March as an opt-in: a classifier decides which actions need your approval, so you stop clicking “allow” on every npm test. We covered the original launch and its default rollout to Pro, Max and Team. v2.1.284 finishes the job: with no configured permission mode, interactive and VS Code sessions start in auto mode everywhere.

The same release adds a “Yes, but ask again next time” answer to auto mode’s read-outside prompt. That is a small admission that a binary allow/deny doesn’t fit how people work. Reading a file outside the project once is different from granting standing access.

Why this is the right call: a permission prompt that fires every few seconds teaches users to approve without reading. A prompt you almost never see, because the classifier handled the boring cases, is one you actually read when it appears. Default-ask is not the same as safe. Competitors that keep a human approving each step are optimizing for the look of control.

What to check before you rely on it
#

Defaults are a policy decision made on your behalf. Some things to do this week:

  1. Pin your mode explicitly if you care. If your team has a standard, set it in settings rather than inheriting whatever the default is next release.
  2. Use the governance controls from v2.1.283. deniedModels blocks specific models, and availableModelsMatch: "exact" blocks new releases until you list them. Sonnet 5.5 becoming the new default Sonnet is exactly the situation these were built for. v2.1.285 added an allowedProviders managed setting to limit which API providers a machine can use.
  3. Run /doctor prompt-audit. Added in v2.1.283, it audits your CLAUDE.md and skills for prompting patterns written for older models. A model swap is when stale instructions (“think step by step”, heavy scaffolding) quietly cost you tokens.
  4. Remember checkpoint limits. Claude Code’s checkpointing does not track changes made through Bash (rm, mv, cp), per its own docs. Auto mode lowers friction; it does not add an undo for shell commands. Keep work in git and use isolated containers for anything riskier.

The bigger pattern
#

The last two weeks of Claude Code releases read as one story. v2.1.283 shipped model governance. v2.1.284 shipped a cheaper default model and a lower-friction default permission mode. v2.1.285 added provider restrictions, and v2.1.286 made Claude run a project’s verify skill before committing, when one exists. Capability is getting cheaper and more autonomous by default, and the controls for organizations are arriving in the same releases rather than afterwards.

That is the paradigm that spec-driven development bets on. If you write the spec and the verification, such as a verify skill that runs your acceptance checks, the agent can run with fewer interruptions because the checks, not your attention, are the safety net. Tools that make you the approval step for every edit cannot scale that way.

Bottom line
#

Sonnet 5.5 is a sensible efficiency upgrade: same price per token, claimed lower cost per task, 1M context. Treat the 30% figure as a vendor claim until you measure it on your own repos. The auto-mode default is the more consequential change. Set your permission mode deliberately, turn on the governance settings if you manage a team, and put your verification in code.

Sources
#

AI Tools

Claude Opus 5.5: 40% Cheaper, the New Claude Code Default — and a Benchmark Table Worth Double-Checking

Anthropic shipped Claude Opus 5.5 on September 22, cutting API pricing to $4/$20 per million tokens (down 40% from Opus 5) while claiming a jump to 66.4% on Terminal-Bench 4.0 — a benchmark whose independent public leaderboard still shows zero populated rows. It’s already the default model in Claude Code v2.1.280.

Sep 23, 2026 · 5 min read

AI Tools

Claude Code's Limits Are Going Up 25% — and Down 17%, Depending How You Count

Anthropic announced it is permanently raising Claude Code’s standard weekly usage limits by 25% starting September 14 — but that number replaces a temporary 50% boost that’s been active since May, so heavy users are actually looking at a 17% cut versus what they have today.

Aug 31, 2026 · 4 min read

AI Tools

Claude Code v2.1.243-246: The Auto Mode Steering Wheel Its Critics Asked For

Three Claude Code releases shipped August 25 (v2.1.243, 245, 246), the biggest adding an Auto mode tab to /permissions for viewing and editing the classifier’s rules, plus a startup warning for dangerous wildcard Bash allow-rules and a fix for auto mode wrongly denying tool calls on large sessions — concrete tooling answering the concrete complaints that surfaced when Auto mode went default eleven days earlier.

Aug 26, 2026 · 5 min read

AI Tools

Auto Mode Went Default Today. Here's What Actually Happened.

Claude Code’s Auto Mode became the default for new Pro, Max, and Team sessions today, August 14, exactly as scheduled. The rollout shipped clean and the classifier overhead is free, but Hacker News reaction split fast — between months of incident-free autonomous use on one side and fresh anecdotes of broken SSH configs and a home-directory rm -rf on the other.

Aug 14, 2026 · 5 min read

AI Tools

Claude Code v2.1.225–228: Skill Sandboxing Tightens Days Before Auto Mode Goes Default

Four releases between August 8-11 close the gap on last week’s v2.1.224: v2.1.228 hardens skills synced from claude.ai so they can no longer shadow local commands or run shell commands on your machine, while a quieter one-line change removes the auto-mode cost notice for Pro, Max, and Team — two days before auto mode becomes the default permission mode on August 14.

Aug 12, 2026 · 5 min read