#14 | GPT-Live + production benchmark + Grok 4.5
GPT-Live gives agents voice, Databricks benchmarks coding agents on real production code, and Grok 4.5 ships at $2/M tokens.
Greetings! Your daily Agent Builders update is here:
━━ ◆ ━━
On this day in AI: On July 9, 1922, Kathleen Booth was born in Stourbridge, England. She invented assembly language and co-designed three early computers (ARC, SEC, and APE(X)C) alongside her husband Andrew Booth, working with John von Neumann at Princeton on stored-program architecture. She wrote one of the first programming textbooks in 1958, co-founded Birkbeck's computing department, and later conducted research on neural networks for pattern recognition and machine translation. Her work on bridging human intent and machine execution through symbolic languages remains the foundation of every programming language and AI system running today.
━━ ◆ ━━
What's happening:
🎙 OpenAI launches GPT-Live with full-duplex voice and agentic delegation
🏭 Databricks benchmarks coding agents against a multi-million line production codebase
🚀 SpaceXAI ships Grok 4.5 at $2/M input tokens with agentic RL training
+ 💬 Community updates
+ 🛠 One thing to try today
━━ ◆ ━━
Hand-picked news:
🎙 OpenAI Launches GPT-Live, a Full-Duplex Voice Model That Delegates Agentic Work → OpenAI Blog
GPT-Live uses a continuous full-duplex architecture that listens and speaks simultaneously, then delegates complex tasks like search and reasoning to GPT-5.5 in the background while maintaining natural conversation flow.
This decouples real-time interaction (sub-second decisions on whether to speak, pause, interrupt, or invoke a tool) from agentic delegation (background call to a frontier model) without forcing a turn-based pattern.
For agent builders, this creates a new interface paradigm: lightweight voice interaction layer on top of heavyweight agent orchestration, rolling out to ChatGPT users today across iOS, Android, and web.
🏭 Databricks Benchmarks Coding Agents on a Multi-Million Line Production Codebase → Databricks Blog
Databricks published an internal benchmark built from real merged PRs across Python, Go, TypeScript, Scala, and Rust, testing coding agents against their own multi-million line production codebase with the Pi harness.
The results show three capability tiers with GLM 5.2 statistically tied with Opus 4.8 on quality at $1.28/task vs $1.94/task, and the Pi harness sending 3x less context per turn than Claude Code or Codex while maintaining equal quality.
Per-token price is a poor proxy for per-task cost: harness efficiency in context management per turn matters more than model choice, and open models like GLM 5.2 now compete at the top tier for real-world coding.
🚀 SpaceXAI Launches Grok 4.5 at $2/$6 per Million Tokens, Trained with Agentic RL → xAI News
Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs with RL over hundreds of thousands of multi-step software engineering tasks, achieving 80 tokens per second throughput.
The model resolves SWE-Bench Pro tasks with 4.2x fewer output tokens than Opus 4.8 (15,954 vs 67,020 on average), thanks to per-token intelligence optimization from agentic training.
Available today in Cursor, Grok Build, and the xAI API, this puts significant downward pricing pressure on the entire frontier model market and enables new economics for agent products.
━━ ◆ ━━
What builders are talking about:
💬 Budget-constrained benchmarks as an alternative to cost-per-task -- A lively HN thread on OpenAI's SWE-Bench Pro analysis proposed flipping the evaluation model: give an agent a fixed $100 API budget and measure how much it accomplishes, rather than reporting cost-per-task after the fact. Commenters noted this would surface interesting tactics like self-verification and iterative refinement with cheaper models. Source: [Hacker News]
💬 Rethinking "AI agents" as just better human tools -- The Flint HN discussion pushed back on the "for AI agents" marketing label, with several commenters arguing that a well-designed intermediate language with clear semantics and sensible defaults benefits humans and agents equally. The thread highlighted a growing fatigue with agent-specific tooling that reinvents existing solutions with a buzzword wrapper. Source: [Hacker News]
💬 pilotfish: multi-model orchestration layer for Claude Code -- A new open-source tool (215 stars, Jul 8) lets a frontier model plan while cheaper models execute with verification guards, installable via one prompt. It signals the emerging "routing" pattern where cost optimization happens at the harness level rather than the model level, similar to what Databricks' Pi harness demonstrated. Source: [GitHub]
━━ ◆ ━━
One thing to try today
Open ChatGPT on iOS, Android, or web and try GPT-Live's voice interface for an agentic task. Ask it something multi-step -- "research the latest MCP server releases, summarize the top three, and draft a short memo comparing their API surfaces." Watch how the full-duplex interaction layer handles interruptions and clarifications while the backend delegates actual reasoning to GPT-5.5. The architectural pattern of a lightweight interaction layer decoupled from heavy agentic execution is worth experiencing firsthand, even if you are not building voice interfaces today.
━━ ◆ ━━
AI Dad Joke of the Day
I told my AI “I’m going to tell you a joke.”
It said “I already know it. You’re going to say something witty about AI, then pause for comedic effect, then...”
I said “Can you let me finish?”
It said “I already did. 0.2 seconds ago.”
━━ ◆ ━━
Agent Builders -- The daily insight for AI agent builders: what's new, what's working, and what to try next.
Found this useful? Forward it to someone building agents.