KDCube
← AI Industry News

GPT-6.1 Terra Hits Bedrock, Step Functions Become Agent Tools

GPT-6.1 Terra is now generally available on Amazon Bedrock as a cheaper, faster computer-use model, and AWS Step Functions can now be handed to agents as deterministic tools via AgentCore Gateway over MCP. Meanwhile Android Studio's Bring Your Own Agent brings 40+ coding agents under one protocol, and OpenAI's no-code ...

Highlights

  • OpenAI's GPT-6.1 Terra is now generally available on Amazon Bedrock — AWS pegs it at ~15 points higher than GPT-5.6 Terra, roughly 3x faster task completion, up to 1M input tokens (text + image), and at least 50% cheaper tokens than GPT-6.1 Sol. (AWS)
  • AWS Step Functions can now be handed to an agent as a tool via Bedrock AgentCore Gateway over MCP — existing workflows become callable without rewrites, and each step still runs deterministically. (AWS)
  • Android Studio Rabbit 2 ships "Bring Your Own Agent" as stable — 40+ coding agents (Claude Agent, Codex, Antigravity, Junie) plug in through the open Agent Client Protocol (ACP). (Android Developers)
  • OpenAI's no-code Agent Builder sunsets Nov 30 — teams are pushed to the self-hosted Agents SDK or ChatGPT Workspace Agents. (OpenAI)

Key Signals

  1. AWS Weekly Roundup · Nov 2, 2026 GPT-6.1 Terra lands on Bedrock as a cheap "computer-use" flagship

    Terra targets high-volume agentic and computer-use work, not frontier reasoning: AWS cites a big low-effort accuracy jump over GPT-5.6 Terra, 3x faster completion, and ≥50% lower token cost versus GPT-6.1 Sol, reachable via Bedrock APIs, ChatGPT Work, and Codex. (AWS, BleepingComputer) For operators running fleets of agents, the economics of per-call model choice just shifted again.

  2. AWS Weekly Roundup · Nov 2, 2026 Deterministic workflows become first-class agent tools

    Bedrock AgentCore Gateway now exposes Standard and Express Step Functions to agents over MCP, so a non-deterministic planner can invoke a deterministic, audited workflow that "runs each step exactly as you defined it." (AWS) This is the same deterministic-execution thesis Sema4.ai pushed last week, now baked into a hyperscaler runtime.

  3. Android Studio Rabbit 2 One protocol, 40 agents: ACP goes stable in the IDE

    BYOA pairs your chosen agent with IDE-native context — full project graph, build diagnostics, emulator controls — passed in to cut token spend, with local agents running as subprocesses over JSON-RPC via ACP. (Android Developers) One integration covering 40 agents is the clearest sign yet that agent↔tool protocols (MCP, A2A, now ACP) are collapsing into shared plumbing.

Why It Matters / What To Watch

  1. Model-ops teams get a cheaper default for grunt-work agents.
    • Re-benchmark your computer-use and bulk-automation agents on gpt-6.1-terra before committing budget to pricier tiers; the 1M-token window changes context-engineering tradeoffs. (AWS)
    • Watch whether Terra's low-effort accuracy holds on your evals, not OpenAI's internal benchmark. (BleepingComputer)
  2. "Workflow as a tool" is becoming the governance story.
    • If you already run Step Functions, test wrapping them as AgentCore Gateway tools to keep auditable, deterministic execution under an agent's control. (AWS)
  3. Another no-code builder sunset — plan the migration now.
    • Agent Builder stops Nov 30; export canvases to the Agents SDK (self-hosted) or ChatGPT Workspace Agents, and note ChatKit front-ends keep working. (OpenAI) Pair this with the Vertex AI Extensions Nov 26 cutoff: two managed-agent lifecycles end within days of each other.

Quick Links