CONFIRMED. Anthropic published Claude Code 2.1.284 on September 28 with a change that affects both capability and operating cost: Claude Sonnet 5.5 is now the default Sonnet model on the Anthropic API. The release combines that model switch with a long list of reliability and governance fixes, showing where agentic software is maturing: less focus on a flashy answer, more focus on controlling the session around it.
Sonnet 5.5 arrives with a 1M token context window and pricing of $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens, according to the release notes. For small and midsize companies, this can simplify the default choice for coding agents and tool-heavy workflows. It does not automatically reduce the bill. A larger context window can also make it easier to send more data than the task needs, while API availability and account entitlements may vary by surface.
The most immediately useful operational addition is /mcp reconnect all. When several MCP servers lose their connection, an operator can retry all failed or unauthenticated servers from the interactive terminal instead of reconnecting them one by one. The release also fixes resumed-session tool calls that previously failed while an MCP server was still connecting, now waiting up to ten seconds. That does not make a broken server healthy, but it turns a common interruption into a more recoverable event.
Spend control also becomes less abstract. Claude apps gateway users can see dollar amounts in /usage and the status line, while the rate-limit data gains used and limit values in US dollars and the billing period. For an agency or small product team, that makes budget drift visible before a month closes. The warning remains important: visibility is not a budget policy. Someone still needs to define limits, alerts, and what happens when a model approaches them.
The release tightens the edges around permissions and managed environments. Auto mode adds an option to allow one read outside the working directories while preserving the question for future reads. Managed policies now warn when their allowed-model list is empty or omits the model that Claude Code starts with. The gateway gains certificate-based client authentication and direct Google Cloud OTLP export support. Those changes matter most to teams moving from a personal terminal to a controlled internal workflow.
There is a broad reliability pass underneath the headline. Damaged response streams are retried or reported as interruptions instead of leaking raw errors. Overload and server errors after a thinking block can retry. Overlong prompts can compact again with less recent history. The Agent SDK handles malformed image input without poisoning later turns, plugin installation cleans up after dependency failures, and hooks now preserve useful error output in debug logs. None of this makes the model infallible, but it reduces the number of ways a useful run can die for reasons unrelated to the task.
My reading: this is worth testing in development now, especially for teams already using MCP or Anthropic gateways. The headline is Sonnet 5.5. The practical value is the combination of a new default, clearer spending signals, and better recovery. Update first, compare cost per completed task rather than token price alone, and keep a rollback path for prompts or plugins that behave differently.
Sources
Official release: https://github.com/anthropics/claude-code/releases/tag/v2.1.284
Official changelog: https://code.claude.com/docs/en/changelog
Independent version index: https://docs.rs/claude-codes/latest/claude_codes/io/struct.ThinkingBlock.html