AI developer tools · What shipped

Claude Fable 5.1 was released as the new default Fable model with 1M context window at $10/$50 per million tokens with $0.25/million token cache reads. Claude Code v2.1.257 introduced a Containment Escape rule in auto mode to prevent unauthorized cloud metadata fetches and cross-tenant access. The release also added support for Claude apps gateway-supplied model descriptions and fixed prompt caching on Fable 5.1 to cover context attached after tool results. Additional improvements include new managed MCP server settings for organizations, improved terminal compatibility, and various bug fixes affecting permission rules, worktree isolation, and background session management.

Langfuse released v4.28.1 with performance improvements including text index caches on the ClickHouse events read path, and v4.28.0 featuring evaluation enhancements such as tracking judge prompt metadata, dedicated evaluator execution columns, evaluator alert management, and observation filter builders. Recent versions also added multimodal support for evaluators (v4.27.0), API specification rendering (v4.27.0), normalized I/O contract definitions (v4.25.0), and stable evaluator API endpoints (v4.23.0) alongside infrastructure improvements like ClickHouse query optimizations and billing integration.

Meta shipped Muse Spark 1.3 on September 2, 2026, available immediately in Muse Code and through the Meta Model API as a drop-in upgrade. The model is frontier-competitive, tying GPT-5.6 Sol, Grok 4.6, and Claude Opus 5 on Artificial Analysis's Intelligence Index at roughly half their cost per task. Meta trained 1.3 for long-horizon coding and instruction following, claiming approximately 20 percent fewer tool calls and 25 percent fewer tokens on coding tasks compared to version 1.2. On benchmarks, Muse Spark 1.3 leads or ties on long-context retrieval and coding tasks with near-perfect scores on million-token retrieval, while Claude Opus 5 leads on broader agent benchmarks. The customer-available xhigh tier is priced at $1.25 per million input tokens and $4.25 per million output tokens under the standard tier, with a heavily discounted contributor tier at $0.10 and $0.20 respectively in exchange for permission to train future Meta models on prompts and completions.

Notion MCP splits keyword search and AI search into separate tools: notion-search now handles keyword-only searches within Notion content, while the new notion-ai-search tool performs semantic searches across connected apps like Slack and Google Drive when Notion AI is available. Clients can check tool availability before calling and route requests accordingly. The notion-fetch MCP tool now returns page cover and icon metadata, and notion-query-data-sources supports preserving rich text in results. Custom Agent tools in Notion MCP return upgrade prompts with recovery links when Notion AI access is unavailable, and new session tools enable finding and managing Custom Agent sessions with message sending and status checking capabilities. Notion Agent APIs are now in public beta, allowing developers to start chats with Custom Agents, stream replies, submit actions, and manage agents programmatically through sessions and events. The beta includes endpoints for querying agents, retrieving per-agent usage insights, and setting credit limits or enabling/disabling agents individually or in batches. Multi-source SQL queries now work on Business plans with Notion AI (previously Enterprise-only), and the Views API supports creating, updating, deleting, and querying database views with configurable filters, sorts, and layout settings. The JavaScript SDK v5.25.1+ fixes type guards for full database responses, adds does_not_equal verification filters, and includes helpers like iterateAllDataSourceRows() for paginating large result sets. API version 2026-03-11 replaces archived with in_trash, after with position, and renames transcription to meeting_notes; the enhanced markdown API now supports update_content and replace_content commands for targeted edits.