xAI launched Grok Voice Think Fast 2.0 at $0.08 per audio minute. It scored 82.9% on the Artificial Analysis speech-to-speech quality index, ahead of GPT-Realtime-2.1 High at 79.1% and Gemini 3.1 Flash High at 69.5%, with 0.70s to first audio. (xAI)

Luna dropped from $1.00/$6.00 to $0.20/$1.20 per million input/output tokens. Terra fell from $2.50/$15.00 to $2.00/$12.00. Flagship Sol unchanged at $5.00/$30.00. (OpenAI pricing) OpenAI attributed the cuts to inference efficiency improvements Sol developed by optimizing its own inference stack. If you priced a Luna workload at launch, re-run your unit economics this week.

CTGT distilled DeepSeek V4 Flash into GPT-OSS-120B on a finance-only corpus and judged 304 prompts with four AI judges: DeepSeek's +45.45-point selective-avoidance gap shrank to under 1.4 points once distilled. That finding cuts directly against a core rationale for restricting Chinese open-weight AI exports.

GitHub shipped agent SKILL.md files and read-only MCP server connections to GA on July 29 across every paid Copilot tier (Pro, Pro+, Business, Enterprise). A file under .github/skills/ now feeds the review agent your internal coding standards - useful, and also a new supply-chain surface to vet before it reaches production. GitHub and Playwright MCP servers are on by default.

The final 2026-07-28 MCP revision drops the stateful initialize handshake, so MCP servers run on plain serverless behind round-robin load balancers with no sticky sessions. Adds multi-round-trip requests, header-based routing, client-side tool-list caching, and OAuth 2.1 hardening. The TypeScript and Python SDKs have each crossed 1 billion total downloads.

GitHub just made SKILL.md GA, so an agent file in .github/skills/ now steers every code review. This is the security review your team runs before one reaches production - vendor vetting, scope limits, secret exposure checks, and rollback plan.

With the spec now stateless and SDK downloads past a billion, the governance question is which servers you trust and how you scope them. Vendor-vetting checklist, scope policy templates, and a monitoring runbook for enterprise MCP deployments.

A working GitHub Actions example that fails a merge on a dropped eval score but keeps the pipeline moving for everything else. Covers eval harness setup, score thresholds, and a bypass path for hotfixes.

How to move from a single latency number to full span-level traces using the OpenTelemetry GenAI semantic conventions, step by step. Covers instrumentation, the collector config, and reading the resulting spans in your existing observability stack.