AI developer tools · What shipped

Claude Code v2.1.295 released with multiple developer tool improvements. Added `onFailure: "block"` for command and HTTP hooks to block actions on failure, Program Status Protocol (OSC 7501) support for terminal status display, quoted text copying without markers in `/copy picker`, and timeout support for upstream TTFB on Claude apps gateway's cloud upstreams (Bedrock, Vertex, Foundry). New features include optional models list for each upstream, `forceLoginMethod: "gateway"` option, `upstream_request_id` in audit events, native notifications through `$.ui.notify`, and children support for mod buttons. Fixed numerous issues with MCP servers, file handling, plugin validation, hooks, and performance across multiple areas. Earlier v2.1.294 fixed prompt and agent hook instruction interpretation and improved SubagentStop hook judgment. v2.1.293 added Claude Haiku 5.5 as default Haiku model with expanded context (1M tokens) and updated pricing. v2.1.292 introduced `--marketplace` flag for plugin installation, agent effort parameter, prompt caching in `$.model.complete`, and workflow agent spawning in mod hooks. v2.1.291 fixed cloud session permission prompt regressions and message loss issues.

Moonshot AI released Kimi K2.7 Code, an open-source coding-focused agentic model showing substantial improvements over K2.6. On coding benchmarks, K2.7 Code achieved +21.8% on Kimi Code Bench v2 (62.0 vs 50.9), +11.0% on Program Bench (53.6 vs 48.3), and +31.5% on MLS Bench Lite (35.1 vs 26.7). For agentic task execution, it improved roughly 10% across benchmarks including Kimi Claw 24/7 Bench, MCP Atlas, and MCP Mark Verified. The model reduces thinking-token usage by approximately 30% compared with K2.6 while maintaining higher performance scores, optimized for long-horizon coding tasks and complex software engineering workflows. Kimi K2.7 Code uses a Mixture-of-Experts architecture with 1 trillion total parameters and 32 billion activated parameters per token, supports 256K context length, and is available through Kimi Code and Kimi API with usage-based pricing ($0.95 input per million tokens cache miss, $4.00 output).

Chutes AI and Harvard researchers released "A Year in LLM Serving," a public dataset of 6.12 billion LLM inference requests spanning 9,174 models collected over approximately one year from April 2025 through April 2026. The 91 GB Parquet file contains serving metadata including timestamps, token counts, cache hits, and anonymized user identifiers, but excludes prompts, responses, and tool calls. The dataset captures 35.8 trillion input tokens processed across 875,921 serving instances, providing researchers with ground-truth data on prefix caching behavior, request patterns, and infrastructure performance at a scale previously unavailable in public academic datasets. Led by Harvard assistant professor Juncheng Yang with co-authors from Harvard and University of Chicago, the release is positioned to support research on real-world LLM serving workload optimization, system design, and capacity planning, filling a gap that infrastructure teams have traditionally addressed only with proprietary internal telemetry.

AMD released detailed benchmark comparisons between the MI455X (Helios) and MI355X accelerators, showing up to 34x higher token throughput on DeepSeek-V4-Flash inference at high interactivity levels. The MI455X demonstrates 3.3x better MXFP4 GEMM compute performance, up to 3.8x improved memory bandwidth for MLA decode attention, and 3x higher HBM read throughput with the transition from HBM3E to HBM4. Networking benchmarks show up to 3.8x aggregate scale-up bandwidth and 27x single GPU-to-GPU bandwidth with the switched fabric topology, plus 2x scale-out bandwidth between nodes. These component improvements compound across inference phases—prefill, decode, and projections—explaining why end-to-end serving gains exceed individual component ratios.