Sonnet 5.5 matches Opus 5.5 on office work at a fifth of the price, Claude for Government goes FedRAMP High GA, Claude Mods ship for Code

Claude Sonnet 5.5 shipped on September 28 at 18:03 UTC, six days after Opus 5.5. Pricing is unchanged from Sonnet 5: $2 per million input tokens, $10 per million output tokens, cache reads $0.20/MTok, Batch API at 50% discount. The model is 30%+ faster and cuts per-task cost by up to 30% through fewer tokens and fewer tool calls. On Anthropic's office knowledge work benchmark, Sonnet 5.5 scores 1,844 Elo — essentially matching Opus 5.5's 1,846 — meaning for document creation, report generation, and well-scoped routine tasks, the two models are statistically tied while Opus costs 5x more on input and 2x more on output. Gaps remain in hard coding problems (46.2% vs Opus's 54.4%), deep academic knowledge (64.5% vs 67.7%), and complex open-ended judgment work. Sonnet 5.5 leads on computer use from screenshots (80.1%), chart reading (61.6%), and Terminal-Bench agent tasks (70.6%). Context: 1M tokens, max output 128K (300K on Batch API beta), knowledge cutoff June 2026, model ID claude-sonnet-5-5. Anthropic's guidance: start on Sonnet 5.5 at high effort; escalate to Opus 5.5 only where your tests fail.

Sonnet 5.5 has five breaking changes: (1) thinking is disabled via between_tools rather than disabled — update any explicit thinking-control logic; (2) forced tool use syntax now returns errors — remove it from prompts; (3) thinking blocks are account-bound and only work in the account that created them — cross-account thinking block reuse fails; (4) the older computer_use_20241022 tool is rejected on Claude API and Google Cloud — update to computer_use_20260814; (5) some advisor pairings are refused — check your advisor configurations. The tokenizer is unchanged from Sonnet 5, so token counts are the same for identical text. Audit agent harnesses that control thinking explicitly or rely on forced tool use before migrating.

Claude for Government left beta and became generally available on September 30 under FedRAMP High authorization — the US government's top cloud security tier. The offering covers the full Claude surface: desktop chat with file handling, skills, plugins, projects, Claude Code, and Claude Cowork. Agencies authenticate via their own identity providers with SCIM group mappings for user sync; conversation history stays on agency-managed devices. Two-person approval is required for sensitive operations, and admins receive audit logs. Compliance exports contain usage data only, not conversation content. Pricing has no per-seat fees — agencies purchase usage in prepaid blocks with hard spending caps configurable by dollar amount and model type. Claude Code CLI and Claude for Microsoft 365 entered early access the same day. Available directly from Anthropic or through GSA OneGov — the GSA extended offer deadline is October 31, 2026.

Anthropic committed $100 million to the Claude Frontier Academy — a 12-week residency program aimed at producing 10,000 certified Frontier Deployed Engineers by end of 2027. The model follows medical training: an in-person course with Anthropic engineers covering simulated enterprise deployment scenarios and a graded practical exam, followed by a 12-week residency where engineers lead real Claude projects at their own organizations, then a final assessment to earn the FDE credential. First cohorts launched with eight organizations: Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk. The program runs in San Francisco, New York, and London. Participation is invitation-only — organizations nominate engineers through their Anthropic account team or Partner Account Manager; each nominee arrives with a pre-identified Claude project. First FDE badges are expected in early 2027.

Claude Code 2.1.287 introduced Mods — a new plugin capability that lets teams write TypeScript functions to modify deeper Claude Code behavior. Mods can intercept and replace built-in tools, add new UI panels, surface custom information in the interface, and override default behaviors. The release ships a built-in 'You should know' mod for enhanced session oversight. Write a mod yourself or ask Claude Code to scaffold one; configure existing plugin mods via `claude plugin configure <plugin>`. Also in 2.1.287: n:<text> filter in agents view for matching session names and tasks, URL prompts from MCP servers for OAuth authentication flows, improved Remote Control reconnection with 30-second timeout, better /model picker for Fable versions, and a Windows startup warning when Bash tool denial affects PowerShell.

Four Claude Code releases shipped this week. v2.1.285 (Sep 28): new `claude --desktop` command opens the Claude app on the current directory; `CLAUDE_CODE_DISABLE_WEB_FETCH` env var; plugin configuration via `claude plugin configure <plugin>`; `allowedProviders` managed setting to restrict API providers; restored dangerous rm command safeguard. v2.1.286 (Sep 30): permission prompt counters showing '2 of 5' format; mouse support for list navigation in fullscreen; fixed simultaneous login processes on credential refresh; fixed session resumption losing parallel tool calls. v2.1.287 (Oct 1): Mods, n:<text> agent filter, MCP OAuth URL prompts, Remote Control reconnection improvements. v2.1.288 (Oct 2): `$.ui.selection()` for mods retrieving selected text; Ctrl+C prompt recovery restores draft on Up arrow; re-authentication prompts when MCP servers request additional OAuth scope; `/code-review --max-findings` for customizable finding limits; Ctrl+F session search and Alt+↑/↓ group navigation in agents view; fixed mid-response API timeouts and long conversation failures.