Opus 5 sets the bar, OpenAI cuts prices, MCP goes stateless, and DeepSeek puts V4-Flash under MIT.
July belonged to Opus 5, moving frontier prices, and an MCP that throws its state away. As always, every link goes to the primary source.
A new state of the art for coding and knowledge work at unchanged pricing of $5/$25 per million tokens, with an effort dial and an optional fast mode. It became the default model on Claude Max and in Claude Code on day one.
OpenAI delivers GPT-5.6 as Sol, Terra, and Luna. On July 30 Luna then fell 80 percent to $0.20/$1.20 per million tokens and Terra to $2/$12, which reopens the math on high-volume agent subtasks.
A 284B mixture-of-experts model with a 1M token context, released as open weights under MIT. That makes Flash a serious option for self-hosted coding agents; the big V4-Pro stayed in preview for now.
Google's Flash line moves up. One note for API users: temperature, top_p, and top_k are deprecated on the new Flash models.
Two embodied-reasoning endpoints add spatial reasoning, multi-step tool orchestration, and video moment finding. The previous gemini-robotics-er-1.6-preview shuts down on August 31, so dependent projects have to migrate.
The protocol goes stateless: the initialize handshake and session IDs are gone, version and client identity travel in _meta on every request, and multi round-trip requests replace server-initiated streams. Every MCP server author has to migrate; all four Tier 1 SDKs are updated.
Interpretability work identifying an internal workspace holding thoughts that never surface in output, plus a technique for reading it. Useful background when a model's stated reasoning diverges from its behavior.
Moves agent memory into the process and pushes latency down to around 100 microseconds, making memory access per step instead of per turn practical. A cheap architectural idea worth stealing.
215 attack scenarios across 24 enterprise integrations show success rates of 32 to 81 percent without guardrails; the released AgentRedGuard cuts that by around 75 points. A concrete eval for anyone wiring agents to Gmail or Jira class tools; a revised version landed in July.
Frontier Red Team results on model-driven discovery of cryptographic vulnerabilities. Evidence that offensive security capability is now an ordinary model property.
Free access to Sol-Pro-class models, starting with 10,000 researchers and scaling to 100,000 through 2027. It shows where subsidized frontier compute is flowing.
