Summary

The day's six signals cluster around one question dominating enterprise AI infrastructure: how to make agents affordable and reliable enough for production. Three vendors attacked cost and speed head-on. Google launched Gemini 3.7 Flash, a coding-and-agent model at an introductory $0.75/$3.75 per million input/output tokens; Writer shipped Palmyra X6 plus harness upgrades it claims cut agent-workflow cost 52% and run time 48%; and OpenAI previewed an Ultrafast API tier running GPT-5.6 Sol up to 14x faster (~750 tokens/sec) on Cerebras wafer-scale hardware. Meanwhile the plumbing for durable, interoperable agents kept filling in: Oracle added asynchronous Background Mode to the OCI Responses API, GitHub made write-once Agent Plugins 1.0 generally available across its Copilot surfaces, and 6sense exposed its B2B buying-intelligence data to agents via an MCP server. Together they mark a competitive front that has shifted from peak capability to cost, latency, and portability per unit of agent work.

Key themes

  • The model-cost race intensified: Google (Gemini 3.7 Flash), Writer (Palmyra X6), and OpenAI (tiered inference) all pressed on agent economics in the same window, competing on total cost of ownership and cost per workflow rather than benchmark peaks.
  • Inference speed is becoming its own paid product axis: OpenAI's Cerebras-backed Ultrafast tier puts a frontier model into a low-latency lane, letting latency-sensitive agentic products keep top-tier reasoning without dropping to a smaller model.
  • Durable, connection-free execution is converging into a baseline for agent runtimes: Oracle's OCI Background Mode mirrors the async, long-running-task pattern spreading across OpenAI's Responses API and MCP.
  • Interoperability and portability standards keep spreading: GitHub's write-once Agent Plugins 1.0 and 6sense's MCP server both push toward build-once, run-anywhere tooling — GitHub within its Copilot footprint, 6sense via the cross-vendor MCP standard now reaching data providers well beyond DevOps.

Notable items

  • Google releases Gemini 3.7 Flash (high impact): a 1M-token coding-and-agent model with introductory pricing of $0.75/$3.75 per million input/output tokens through end of 2026, reporting coding gains including FrontierCode 43.6% and DeepSWE 65.3%.
  • OpenAI previews Ultrafast (high impact): a limited-preview API tier serving GPT-5.6 Sol at up to ~750 output tokens/sec (up to 14x standard) on Cerebras hardware, priced above the existing Fast tier.
  • Writer launches Palmyra X6 ($2/$8 per million tokens) plus Writer Agent harness and token-spend governance upgrades, claiming 52% lower cost and 48% faster multi-step agent execution, with unattended runs up to eight hours.
  • GitHub Copilot Agent Plugins 1.0 reaches GA — a portable plugin format across VS Code, Copilot CLI, the Copilot SDK, and the Copilot app — alongside CLI /tasks and queueing, a /rewind undo, JetBrains persistent memory plus Ollama support, and VS Code mid-session model switching.
  • Oracle adds Background Mode to the OCI Responses API, running long-running inference and agent tasks asynchronously without holding an open connection.
  • 6sense ships an MCP server exposing account insights, predictive buying stages, 6QA status, and keyword intent to agents in Claude, ChatGPT, Writer, and Salesforce Agentforce — extending MCP's reach into revenue intelligence.

Source coverage

Source rows used: 6