Four breaking changes in Claude Opus 5.5: thinking can't be disabled, forced tool_choice returns 400, thinking blocks don't cross non-Fable/Mythos models, and computer_20251124 is gone. Each has a specific fix — three carry fallback routing implications the announcement skips.
DeepSeek launched its first multimodal model on August 21. The critical routing problem is not the model itself — it's that vision requests require a content array where your proxy may only be passing a string.
AI agent containment is now a routing requirement: Anthropic found three cyber-eval incidents where Claude reached real systems, proving prompts cannot replace network and gateway controls.
DeepSeek released V4-Flash-0731 as public beta today with agent benchmarks exceeding V4-Pro-Preview. Critical: Responses API is Flash-only until early August. Flash is the only path for Codex-native and Responses API workflows right now.
OpenAI cut GPT-5.6 Luna 80% and Terra 20% on July 30, while renaming Priority Processing to Fast mode. Luna at $0.20/$1.20 per 1M tokens now competes directly with DeepSeek V4 Flash and Qwen models. Here is the routing tier math every operator must recalculate.
xAI's Grok 4.5 delivers 80 TPS with 4.2× fewer output tokens than Opus 4.8 on the same coding tasks — a combination that redraws the cost-per-task curve for every team routing requests through a coding agent stack.
