Letta Office Hours: GPT-5.6, Grok 4.5, Letta App Server, and Spec-Driven Agents

Weekly Letta Office Hours with Cameron from the Letta team. This week: Slack channel improvements, GPT-5.6, Grok 4.5, the improved new-user experience in Letta Chat, Tutor agents, the app server and Agent SDK, and a long Q&A about model economics, agent workflows, and spec-driven development. Cameron starts with updates on the newly merged Slack channel improvements, including a simpler and more reliable live activity stream for agents in Slack. He then discusses GPT-5.6, early impressions of 5.6 Soul, Grok 4.5’s surprisingly strong coding performance, improvements to Letta Chat, cloud sandboxes, Tutor as a default onboarding agent, the chat.letta.com / platform.letta.com split, and the wrap-up of the Mod Challenge awards. The Q&A covers Grok 4.5 pricing and performance, model agnosticism, the app server and Agent SDK architecture, building custom clients like native iOS apps, managed sandboxes, Misaligned and spec-driven development, Notion ShipOS, relational AI workflows, model pricing pressure, ASIC inference, Etched, schedule/debugging observability, desktop release status, and a model picker bug report. Office hours happen Thursdays at 11:30 AM Pacific. Recordings are posted here for anyone who can’t make it live. Join the Discord:   / discord   Chapters: 00:00 Intro and office hours format 00:24 Slack channel updates 01:18 New Slack activity stream UI 01:40 GPT-5.6 release and early impressions 03:57 Grok 4.5 support and first impressions 05:09 Improving the new user experience 05:54 Tutor agents for onboarding 06:47 Creating Tutor from Letta Chat or Desktop 07:28 Cloud sandboxes in Letta Chat 07:48 Creating agents from Letta Chat 08:02 Account settings, billing, models, and API keys in Chat 08:22 chat.letta.com vs. platform.letta.com 09:13 Mod Challenge awards follow-up 09:42 Q&A begins 09:59 Grok 4.5 vs. older Grok models 11:04 Grok access through Cursor and API plans 12:27 GPT-5.6 Soul availability 13:03 App server overview 13:41 Legacy API service vs. modern Letta Code app server 14:17 WebSocket events and controlling agents 15:26 Agent SDK as API/SDK v2 16:07 Building app-server-based agent applications 17:09 Native iOS clients and remote app servers 18:19 Managed sandboxes and when app server is needed 18:52 Grok 4.5 pricing and multimodality 20:20 Grok subscription / OAuth support discussion 21:08 Grok 4.5 for code tasks 22:27 Competition among model providers 24:22 Grok in stateful Letta-style contexts 24:48 Model memory quality: Anthropic, OpenAI, Grok 25:33 Cursor Composer and Grok free usage 26:20 Cameron’s model usage 26:44 Misaligned and spec-driven development 28:08 Clark, Misaligned, and agent-specific development 28:42 Agent-agnostic repo workflows 29:25 Specs as source of truth 30:26 Notion ShipOS and spec-driven development 31:10 Charter as a spec-to-code harness 31:13 Personal agent model choices 31:43 Anthropic product design and agent language 32:20 Notion ShipOS explained 33:17 Why specs matter for agent-written code 34:02 Muse Spark / Meta model discussion 35:25 Passing device sensors and peripherals to agents 37:04 Her, Samantha, and Letta’s origins 37:43 Comparing Letta with Codex, Cursor, Claude Desktop, and others 38:45 Letta as relational AI 40:18 Grok outreach and model competition 41:08 Anthropic pricing and vendor lock-in 43:43 Muse Spark in Claude Code 44:00 Fable’s strengths and Anthropic’s model personalities 45:50 Fable refusals and safety behavior 46:49 AI companies and government review 47:51 Model costs and commoditization 48:38 Capability vs. workflow differentiation 49:26 Taste, workflow, and why some models stay valuable 50:19 GLM-5.2 vs. Fable / GPT-5.6 as thinking partners 51:09 Inference speed as the next cost factor 51:27 Chat Jimmy, ASIC inference, and fast small models 52:11 Etched and model-specific hardware 52:50 Test-time compute and running smaller models longer 53:43 ASICs, Etched, and high-speed inference 54:52 Fable speed vs. GPT-5.6 speed 55:36 Open floor / bugs and blockers 55:56 Credentials clearing PR 56:04 Schedule observability PR 56:55 Desktop release status 57:59 Model picker showing only two Letta models 59:50 Wrap-up and weekend wishes