Claude Code

Claude Code 2.1.284 adds Sonnet 5.5 and shows your gateway budget in dollars

2 min read AI-generated

Auto mode now has a third answer when it asks about a read outside your working directories: yes, but ask me again next time.

Featured image for "Claude Code 2.1.284 adds Sonnet 5.5 and shows your gateway budget in dollars"

2.1.284 landed on npm on Monday at 17:11 UTC, three days after 2.1.283. The changelog is long and the fix list is longer. Four things are worth your attention.

Sonnet 5.5 is in

claude-sonnet-5-5 is available from this version and is the default Sonnet model on the Claude API: 1M token context, $2 in and $10 out per million, 20 cents for cache reads. What else the model brings is in my piece on it. Claude Desktop follows too: version 2.9939.4 bundles exactly this Claude Code version, per its changelog, and supports Sonnet 5.5.

A third answer in auto mode

Until now, a read outside your working directories gave you yes or no, and a yes stuck. There’s now a “Yes, but ask again next time”. You allow that one read and keep the prompt for every later one.

Money instead of percentages

If you work behind a Claude apps gateway, your spend limit shows up in /usage and the status line as an actual amount: “$271.40 / $500.00 spent this month”. The status line gains three fields for it, used_usd, limit_usd and period under rate_limits.spend_limit. The gateway itself has to run this version or later.

Small things that save time

/mcp reconnect all retries every MCP server that failed to connect or needs authentication, in one go. Three new actions, effortSlider:decreaseEffort, increaseEffort and toggleUltracode, can be rebound in keybindings.json. And /rate-limit-options finally appears in /help and the command menu, so the usage-limit notices point at a command you can actually find.

Two fixes stand out. A damaged response stream used to leak raw errors like “JSON Parse error”, or write the word “undefined” into an answer – it’s now treated as an interrupted response and retried. And when “Prompt is too long” survives a compaction, Claude Code compacts a second time, keeping less of the recent conversation.

The anti-nagging switch is the best small change

Sonnet 5.5 is the headline, but the third answer in auto mode is what I’ll notice daily. Every permission used to be a one-way door: say yes once to get moving and you’ve lost the prompt for the rest of the session. That’s precisely why people end up clicking through everything. A permission that doesn’t write itself into stone isn’t a comfort feature – it’s what makes the prompt mean anything at all.

Sources

Claude CodeAnthropicReleases