Claude Opus 5.5 launches at 40% lower cost, plugins directory opens to all developers, Claude Tag gets personal connectors

Anthropic released Claude Opus 5.5 on September 22 as the first model in the new Claude 5.5 family. API pricing is $4 per million input tokens and $20 per million output tokens — a 20% reduction from Opus 5's $5/$25 — while cache reads fall 60% from $0.50 to $0.20 per MTok. Anthropic says typical workloads cost about 40% less overall because the model also uses fewer tokens to complete tasks. On benchmarks, Opus 5.5 matches Fable 5.1 on most work and outperforms it on several agentic coding tasks, making it the most cost-efficient self-serve model Anthropic has shipped. It runs over 30% faster than Opus 5. Context window: 1M tokens. Max output: 128K synchronous, 300K on Batch API. Adaptive thinking is always on with effort defaulting to medium. Knowledge cutoff: June 2026. Model ID: claude-opus-5-5. Available on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Also on September 22, five-hour usage limits on Pro, Max, and Team plans rose 20%, and every subscriber received a one-off rate-limit reset token.

Opus 5.5 has four breaking changes from Opus 5: (1) effort defaults to medium, not high — the 40% cost saving Anthropic cites is measured at medium effort; switching to high restores Opus 5-level thinking depth but narrows the cost gap. (2) The thinking.display parameter now only accepts 'omitted' and 'updates' — 'full' returns a 400 error. (3) Cache write costs change: 5-minute TTL falls from $6.25 to $5 per MTok, 1-hour TTL falls from $10 to $8. (4) The model produces shorter completions by default in agentic contexts — a deliberate efficiency calibration. Audit any agent harness that passes effort: 'high' explicitly or relies on full thinking display before migrating.

Anthropic confirmed that Claude Sonnet 5.5 and Claude Haiku 5.5 will follow Opus 5.5 in the coming weeks, completing the 5.5 family across all three tiers. No pricing or benchmark details were provided. The 5.5 family positioning suggests an across-the-board efficiency improvement over the 5.0 generation — Opus 5.5 is both cheaper and faster than Opus 5, suggesting Sonnet 5.5 and Haiku 5.5 will follow the same pattern.

Anthropic opened its plugin directory to any developer on a paid Claude plan on September 25, with a new submission portal at claude.ai/directory/manage. Plugins package MCP connectors, Agent Skills, or both as the main way to build third-party extensions for Claude. Developers can submit a single MCP connector (a remote HTTPS server) or a plugin bundle (MCP servers plus skills, hosted on GitHub). In Claude Code, plugins can also include LSP servers, commands, hooks, and agents. The portal provides auto-validation, step-by-step review tracking, publishing controls, and post-launch usage analytics. Being an Anthropic partner is not required. The directory now lists over 2,000 plugins and connectors. MCP usage has grown approximately 110x over the past year. Plugins support MCP 2.0, MCP Apps, and Enterprise Managed Auth.

Claude Tag expanded on September 24 to let users bring their own personal connectors into Slack channels — not just the shared admin-attached tools. Users can now ask Claude to pull from their personal Google Drive, calendar, CRM, or any connected personal tool when @mentioning Claude in a channel, without exposing that data to other channel members. Review or auto mode for posting was also added — in review mode Claude drafts a response and waits for approval before posting. Stronger Enterprise admin governance controls give workspace admins visibility into which personal connectors are in use across channels. Rolling out to Team plan users now; Enterprise support coming soon.

Anthropic's Threat Intelligence team published its second malicious use report, covering eight months of operations since its first 2025 report. The team identified and disrupted operations where threat actors used Claude for influence operations, targeted fraud, spear-phishing campaigns, and attempts to generate offensive security content. Malicious use has grown more sophisticated — moving from bulk low-effort prompting to targeted multi-turn campaigns that probe safety classifiers. Anthropic blocked tens of thousands of accounts across the period and shares anonymized case studies to help other AI providers recognize similar patterns.

World health organizations are using Claude to accelerate their response to a rare Ebola strain with no confirmed vaccine spreading through eastern Democratic Republic of Congo. Claude is being used for literature review, protocol analysis, and supply chain modeling across multiple organizations simultaneously coordinating the response. Anthropic described the use case as Claude for Science's most time-sensitive application to date.

Claude Code set claude-opus-5-5 as the new default Opus model from September 22. The same release adds broader gateway, MCP, plugin, workflow, and /doctor audit controls, plus faster startup and latency improvements, richer list and terminal navigation, and fixes across sessions, models, sandboxing, Windows, VS Code, cloud, and Claude Tag. Skill folders and command files in the anthropic-skills or claude-ai namespace no longer load — a plugin so named still loads but yields only name-matched commands. The /doctor audit now surfaces these namespace changes. A proxy/gateway regression from 2.1.275 causing HTTP 400 on every request containing the advisor_20260301 tag was also fixed.