AI developer tools · What shipped

I've reviewed the Google Cloud blog content against your user intent for "AI developer tools · What shipped" with focus on AI inference optimization benchmarks, machine learning framework releases, and developer tools performance comparisons. The content primarily covers Google Cloud service announcements and events rather than specific AI developer tool releases with benchmarks or framework performance data. While there are mentions of model releases (Claude Opus 5.5, Claude Opus 5, Claude Sonnet 5, Gemini 3.1 Pro, Gemini 3.1 Flash-Lite, Grok 4.6), these lack concrete benchmarks or performance comparisons relevant to inference optimization or framework evaluation. The only content with quantifiable performance data is: GCSFS 2026.8.0 achieved 5x single-file throughput improvement and scaled to 21 GiB/s with Rapid Bucket, reducing training wait times in PyTorch ecosystems. Run:ai Model Streamer for TPUs showed 2x faster model loading while cutting peak host memory usage by half for 480B parameter models. Cloud Run sandboxes entered public preview for executing AI-generated code safely. However, these are isolated infrastructure optimization notes rather than comprehensive benchmarks or framework comparisons meeting the "real changes in AI developer tools" standard described in your intent.

Three frontier-class models launched within 48 hours in late September 2026, each repricing their tier: SpaceXAI's Grok 4.7 ($2/$6 per million tokens), Anthropic's Claude Opus 5.5 ($4/$20 with cache reads at $0.20), and OpenAI's GPT-6 family including Sol ($2/$10) and Luna ($0.10/$0.50). Vendor-reported benchmarks showed significant disagreements, with Opus 5.5 leading on Terminal-Bench 4.0 (66.4%), Grok 4.7 scoring 71.0% on DeepSWE, and performance splits across CursorBench and GDPval metrics. On cached agent loops (90% input cached), GPT-6 Sol costs $0.276 per task versus Opus 5.5 at $0.516, while uncached tasks show Sol at $0.60 competing with Grok 4.7 at $0.52 and Opus 5.5 at $1.20. GPT-6 Luna became the cheapest frontier model ever shipped, undercutting DeepSeek V4.1-Flash ($0.15/$0.60) on all axes and moving the cost floor from open-weights to closed providers. The article emphasizes running independent golden-set evaluations before platform migration, as vendor benchmark tables disagreed by double digits on the same test names.