AI developer tools · What shipped

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, with Fable 5.1 available on all platforms and Mythos 5.1 available through trusted access programs for cybersecurity and life sciences professionals. Fable 5.1 achieves approximately 25% lower costs than Fable 5 for typical workloads and up to 45% savings for highly agentic work, driven by 75% reduced cache read pricing ($0.25 per million tokens). Input pricing remains $10 per million tokens and output $50 per million tokens. Fable 5.1 demonstrates significant performance gains across multiple benchmarks: Terminal-Bench-Science 0.1 reaches 52.6% (vs. Fable 5's 24.7%), Terminal-Bench 4.0 scores 60.9% (vs. 42.0%), Humanity's Last Exam achieves 65.0% with tools (vs. 63.8%), CursorBench 3.2.0 reaches 73.4% (vs. 70.5%), and OSWorld 2.0 strict mode reaches 41.7% (vs. 36.1%). The model features improved safeguards with 60% fewer false positives in cybersecurity tasks, now permitting vulnerability discovery work, and an updated biology safeguard that fires 85% less often on benign elementary biology and medical queries while restricting research applications to verified professionals through the Life Sciences Verification Program developed with the US government.

Cline v4.1.17 SDK and related releases (September 2-3) delivered significant updates to the AI developer tool. Desktop v0.0.23 refined Agent Plugins discovery through a shared Hub with improved settings management and fixed MCP server shutdown leaks. v4.1.17 and SDK v0.0.82 both refreshed the model catalog adding ten providers (Bothub, OpenReason, SenseNova, TokenGo, TokenRouter, Vancine, Volcengine Ark variants, above.dev, klokintegration.se) and changed default model resolution for 57 providers—most notably Anthropic switching from Claude Opus 5 to Claude Fable 5.1, with Amazon Bedrock, Vertex, OpenRouter, and others following. Both releases fixed tool calling silently disabled for gateway models (Dify, SAP AI Core, opencode, Codex CLI) where empty capability lists were misinterpreted as capability denials, and resolved image input drops from file reads on models with empty capability declarations. Langfuse tracing now functions in minified release builds after detection switched from class-name-based to structural identification. Desktop v0.0.22 added session import from Claude Code, Codex, and opencode with transactional, idempotent translation to Cline's native format, grouped scheduled runs, and macOS voice input support.

The website content is a references section from an academic article consisting entirely of bibliographic citations for papers and technical reports related to code generation models and benchmarks. It contains no actual news, announcements, releases with version numbers, API updates, or recent developments in AI developer tools. It is reference material only, not a news article or announcement.