Claude Code 2.1.284 ships Sonnet 5.5 at the leaked price

Claude Code 2.1.284 came out on September 28, the day Anthropic released Claude Sonnet 5.5. It makes the new model the default Sonnet on the Anthropic API. It also starts interactive terminal and VS Code sessions in auto mode for anyone who never set a permission mode.

Those two changes sit at the top of the v2.1.284 release notes on GitHub, which run to roughly a hundred entries. Most of the rest falls into three groups: admin controls for the Claude apps gateway, a set of fixes for turns that used to die on transient errors, and a pass over how Claude Code handles permissions and trust. One figure in the notes has since gone out of date. The section on what we checked covers it.

Sonnet 5.5 at Sonnet 5’s price

The release note lists claude-sonnet-5-5 with a 1M context window, $2 per million input tokens, $10 per million output tokens and $0.20 per million for cache reads. The 1M context bills at the standard rate, with no long-context premium. Maximum output is 128,000 tokens, and fast mode isn’t offered for this model.

Those per-token rates are the same as Sonnet 5’s, cache writes at $2.50 per million included. The comparison that matters more for model choice is Opus 5.5, released on September 22, six days before Sonnet 5.5. Opus 5.5 costs $4 per million input and $20 per million output, so Sonnet 5.5 comes in at half on input, output and cache writes. At launch, cache reads were the exception: both models listed them at $0.20.

With the per-token price unchanged, Anthropic makes its cost argument on token counts. Sonnet 5.5 “typically needs far fewer tokens to do the same work,” the company’s announcement says, and “in our testing, it costs up to 30% less per task than its predecessor.” That 30% is per task, from fewer tokens and fewer tool calls, and the rate card doesn’t change. Customer figures point the same way: Box reported 12% fewer total tokens, and one customer needed roughly 121,000 tokens for a job that took close to 497,000 on Sonnet 5. Haiku 5.5 was due to follow within weeks of the Sonnet release.

Two smaller entries matter if you run Sonnet 5.5 through a gateway or proxy. Claude Desktop behind a Claude apps gateway used to offer no 1M context option at all. The gateway now marks each 1M-capable model for Desktop automatically. And if a session runs a model ID Claude Code doesn’t recognize, such as a custom model behind a proxy, the Explore subagent now inherits that model. Before, it switched to Opus on the Claude API.

Auto mode as the starting point

The permissions change reaches more users than anything else in the release. Interactive terminal and VS Code sessions now start in auto mode when no permission mode is configured, on every plan and every provider. If you want another default, set permissions.defaultMode. It still overrides the new behaviour.

Auto mode still asks before a read outside the working directories, and that prompt gains a third answer, “Yes, but ask again next time.” It allows that single read and keeps asking about later ones, where the existing yes would stop the questions.

A group of entries tightens what reaches the model or gets approved without you:

• Auto-memory loading now neutralizes invisible characters and tags that imitate Claude Code’s own markup in MEMORY.md and recalled memory notes before Claude sees them.
• Under managed allowManagedPermissionRulesOnly, plugins from marketplaces, claude.ai and npm used to be able to pre-approve their own tools via allowed-tools. Now only plugins from an official Anthropic source, or a source that managed settings vouch for, keep that pre-approval.
• Rules symlinked into .claude/rules from outside the project used to be skipped without the external-imports approval prompt ever appearing. A .claude directory symlinked from outside the project now asks for the same approval.
• {"decision":"block"} from Elicitation and ElicitationResult hooks was being ignored. It now declines the MCP elicitation, as exit code 2 does.
• claude remote-control in a folder you haven’t trusted yet now asks for workspace trust in the terminal instead of exiting.

Safety-triggered model switches change as well. In sessions that pin an Opus model with ANTHROPIC_DEFAULT_OPUS_MODEL or modelOverrides, the Anthropic API now picks the model to switch to for each kind of flag, and the pinned model no longer decides it. When a Sonnet model’s safeguards flag a message, the notice now explains why it happened and offers to edit and retry.

What gateway admins get

Organizations running the Claude apps gateway get the longest run of new controls. Spend limits in /usage and the status line now show dollar amounts, for example ”$271.40 / $500.00 spent this month”, provided the gateway itself runs this version or later. Status line scripts gain used_usd, limit_usd and period under rate_limits.spend_limit.

On the configuration side:

• The gateway warns at startup when a managed policy’s availableModels is empty, or leaves out the model Claude Code starts on without setting model or enforceAvailableModels.
• telemetry.forward_to destinations accept auth: { google: {} }, so telemetry can go straight to Google Cloud’s OTLP endpoint using the gateway’s own Google Cloud credentials.
• The connection to the identity provider supports certificate client authentication (private_key_jwt), for providers that issue certificate credentials instead of client secrets.
• Sign-ins whose identity provider lists many groups used to get 431 Request Header Fields Too Large on every request. The gateway now accepts request headers up to 256 KiB.

Other managed-settings entries fix commands that failed without telling you. claude mcp add used to report success when managed settings restrict MCP servers to plugins, then save a server that never loaded. It now refuses and says what to do. When managed settings require a sign-in through forceLoginMethod or forceLoginOrgUUID and an API key, token or apiKeyHelper is configured, the startup refusal now names the credential in use, where it is set, and how to remove it. On Foundry, a value of ANTHROPIC_FOUNDRY_RESOURCE that isn’t a plain resource name is refused; it used to go into the endpoint host unvalidated.

Fewer turns lost to errors

Several fixes target turns that used to end on a transient failure. A damaged response stream could show raw errors such as “JSON Parse error” or “undefined is not an object”, or write the word “undefined” into an answer. It is now retried or reported as an interrupted response. An overloaded or server error arriving right after a thinking block is now retried instead of ending the turn.

“Prompt is too long” errors that persisted after compacting get a second attempt. If the compacted request is still too long, Claude Code compacts once more and keeps less of the recent conversation. Retries pull the other way. After a dropped connection mid-response, they now share one budget with the rest of the request’s retries, so a request that keeps failing gives up sooner.

MCP gets similar treatment. In a resumed session, tool calls used to fail with “No such tool available” while their server was still connecting. They now wait up to 10 seconds for it. /mcp reconnect all retries every server that failed to connect or needs authentication in one go. The non-interactive first turn waits up to 2 seconds for connecting servers named by --allowedTools or an mcp_tool hook, even when CLAUDE_CODE_MCP_STARTUP_WAIT_MS is 0. Usage views (/usage, /extra-usage and the IDE’s) now back off when the plan-usage endpoint rate-limits or rejects your login, instead of asking again and again.

Ultracode is now a separate toggle in /effort, switched with Tab or /effort ultracode [on|off]. It no longer forces xhigh effort and stays on at any effort level. The slider’s keys can be rebound through the new effortSlider:decreaseEffort, increaseEffort and toggleUltracode actions in keybindings.json. VS Code gets the same change as an on/off switch under the Effort slider. Users already on the highest Max plan no longer see usage-limit warnings that suggest /upgrade. The warnings, and /upgrade itself, now point to /usage-credits when it is available.

The rest is fixes to vim mode, fullscreen rendering, plugins, the VS Code extension, Claude Tag in Slack and Code Review.