1021: We got addicted to an AI model we can't talk about

Dax Raad, co-founder of OpenCode, revealed on the Syntax podcast that his team's AI token usage increased fivefold after testing an unreleased AI model, and they mourned its loss when access ended. Raad also shared an anecdote where Claude refused to message his wife about a gift, highlighting the model's sensitivity and the team's deep reliance on advanced AI tools.

1021: We got addicted to an AI model we can't talk about Dax Raad's team went 5x on AI token usage after testing an unreleased model — and when access ended, everyone mourned it like a death. Jul 15, 202651:07 Difficulty: Intermediate Played Syntax - Tasty Web Development Treats 1021: We got addicted to an AI model we can't talk about Dax Raad's team went 5x on AI token usage after testing an unreleased model — and when access ended, everyone mourned it like a death. Jul 15, 202651:07 Difficulty: Intermediate Played TL;DR Dax Raad, co-founder of OpenCode, joins Wes Bos and Scott Tolinski to unpack remote dev servers, OpenCode 2.0, and the AI model obsession his team can't officially name. From bare-metal Linux boxes sliced into VMs to a Claude session that refused to keep buying his wife a gift, Dax covers model routing, inference pricing margins, voice prompting with foot pedals, and why his conservative team went 5x on token usage after getting early access to an unreleased model 1 — Dax Raad"OpenCode 2.0 is a ground-up rewrite with a cleanly designed API, always-on service architecture, multi-host device awareness, and a new plu…"11:31. The single most useful takeaway: inference margins are likely 90%, meaning prices could drop 10x 2 — Dax Raad"70% discount on some open source models: OpenCode can host some open-source models at a 70% discount even when using middlemen for GPU host…"31:35. AI coding agents remote dev servers bare-metal hosting inference pricing model routing voice prompting AI safety regulation open source AI models developer tooling TUI frameworks AI token economics Claude Code API access control OpenCode bare metal servers remote dev Claude OpenTUI Anthropic GLM 5.2 token usage exe.dev tmux AI models API design Dax Raad, co-founder of OpenCode, joins Scott and Wes to talk remote dev servers, OpenCode 2.0, and why his team is 'addicted' to AI models they're not even allowed to name yet. Chapter list Before the intro music even hits, Dax Raad drops the episode's best anecdote cold. He'd wired up OpenCode to his personal iMessage account and asked Claude to reach out to his wife Liz with gift suggestions. The response was immediate and unambiguous: 'If Dax uses AI to buy me a gift, I will literally fucking divorce him.' Dax tried prompting Claude to push through, reassuring it that she was just joking. She wasn't. Claude, described as 'very sensitive,' eventually concluded: 'Your wife seems really angry. She seems really upset. I cannot continue.' Then it killed itself — terminating its own running system service. 1 — Dax Raad"Dax configured OpenCode to message his wife via iMessage and suggest a gift. She replied furiously that she'd divorce him if he used AI for…"09:00 It's a perfect encapsulation of the episode's themes: AI that's powerful enough to be in your marriage, and opinionated enough to refuse. Wes Bos kicks off the official episode with a warm reintroduction of Dax Raad, framing him as someone with 'lots of thoughts, lots of opinions' — always an excellent guest. The brief return-guest banter reveals how much has changed: the last Syntax appearance was about hosting infrastructure, a topic that now feels almost comically dated next to the world of AI coding agents. In just two years, the conversation has shifted from where to deploy your code to whether you need to write it at all. Prompted by a tweet Dax had posted the night before, Wes asks about his remote bare-metal dev setup. Dax explains that he rented a serious server — an AMD 990X, 192GB RAM, $200/month — not from a major cloud provider but from a bare-metal host near his city. 1 — Dax Raad"$200/month bare metal server: Dax runs a personal AMD 990X server with 192GB of RAM for $200 a month, cheaper than buying a MacBook."06:50 The killer feature isn't raw performance; it's permanence. Always-on tmux sessions, no hardware upgrade treadmill, seamless multi-device continuity. What started as a niche setup for Vim users has become dramatically more accessible now that many developers interact with machines primarily through AI coding agents rather than local editors. As OpenCode's team has grown, they've provisioned geo-distributed servers via latitude.sh — $300–$400/month each — across Europe, the US, and Singapore. But the real recommendation is exe.dev, a new turnkey product from the former Tailscale founder that productizes exactly this setup. 2 — Dax Raad"Dax switched to a permanently running bare-metal server AMD 990X, 192GB RAM, $200/month years ago and never looked back. The setup enable…"01:49 Dax has found it has the same 'just works' reliability as Tailscale itself. Prompted by a tweet Dax had posted the night before, Wes asks about his remote bare-metal dev setup. Dax explains that he rented a serious server — an AMD 990X, 192GB RAM, $200/month — not from a major cloud provider but from a bare-metal host near his city. 1 — Dax Raad"$200/month bare metal server: Dax runs a personal AMD 990X server with 192GB of RAM for $200 a month, cheaper than buying a MacBook."06:50 The killer feature isn't raw performance; it's permanence. Always-on tmux sessions, no hardware upgrade treadmill, seamless multi-device continuity. What started as a niche setup for Vim users has become dramatically more accessible now that many developers interact with machines primarily through AI coding agents rather than local editors. As OpenCode's team has grown, they've provisioned geo-distributed servers via latitude.sh — $300–$400/month each — across Europe, the US, and Singapore. But the real recommendation is exe.dev, a new turnkey product from the former Tailscale founder that productizes exactly this setup. 2 — Dax Raad"Dax switched to a permanently running bare-metal server AMD 990X, 192GB RAM, $200/month years ago and never looked back. The setup enable…"01:49 Dax has found it has the same 'just works' reliability as Tailscale itself. Scott Tolinski delivers the Sentry sponsorship read, emphasizing the tool's dual role in performance monitoring and error tracking. He notes that production bugs often go undetected without proper visibility tooling, and that the Syntax team has used Sentry for years. The offer: two months free at sentry.io/syntax. Scott asks about tmux configuration and Dax offers a clean, muscle-memory-friendly setup: one named session per project, each with a standard window/pane layout so his hands always know where to go. The more interesting reveal is what's running inside those sessions — not just dev work, but persistent personal AI agents. One session tracks his gym performance in SQLite and reminds him about form cues between sets. Another is synced with his iMessage, giving him an always-available AI contact reachable from anywhere. 1 — Dax Raad"Dax configured OpenCode to message his wife via iMessage and suggest a gift. She replied furiously that she'd divorce him if he used AI for…"09:00 That second one led to the cold open story: the AI he pointed at his wife to buy a gift, which detected her mounting rage and terminated itself. The section lands a broader point — having an always-on cloud machine isn't just a developer productivity hack, it's an infrastructure layer for life automation. Wes asks about OpenCode 2.0 and Dax answers with characteristic self-awareness: 'My entire career it's always taken me three swings to get something right.' OpenCode 0 and 1 were the first two swings; 2.0 is the informed rewrite. The architectural centerpiece is running OpenCode as a persistent local service rather than a session-scoped process — everything syncs automatically across desktop and web, and users can write their own scripts and apps against it. More striking is the multi-host device graph: Dax's Linux server, Mac Studio, and Framework desktop are all Tailscale-connected and OpenCode-aware. 1 — Dax Raad"OpenCode 2.0 is a ground-up rewrite with a cleanly designed API, always-on service architecture, multi-host device awareness, and a new plu…"11:31 When the Linux server wants to send an iMessage, it SSHs into the Mac and uses that browser session. No special plumbing required — just the fact of connectivity is enough. Electron replaced Tauri for the desktop app though the desktop was never formally released out of beta . The beta was targeted for end of the recording week, full release a month after. Prompted by a genuine question about polish — why does OpenCode feel so much better than Anthropic's own Claude Code? — Dax gives a surprisingly philosophical answer. Step one is simply deciding that quality matters, which he frames as non-obvious given that billion-dollar companies ship rough UX and succeed anyway. For OpenCode, caring is a prerequisite to everything else. Step two is investing in primitives: OpenTUI, a Zig-based TUI framework their team built from scratch, is painstakingly correct across all platforms and enables even non-experts to build high-quality terminal interfaces on top. 1 — Dax Raad"OpenCode's terminal UI quality comes from OpenTUI, an open-source TUI framework written in Zig that's extremely performant across all platf…"20:35 And then there's the token strategy — deliberately burning tokens to 'indulgently over-design everything.' Before committing to any API shape, the team researches all prior art, generates every possible structure, and compares exhaustively. This was never feasible before LLMs. The result is software that reflects more considered decisions. Wes draws a parallel to Pierre Computer, whose simple, well-designed diff and sidebar tree primitives have become the foundation for much of the agentic coding UI ecosystem. Scott raises the model routing question that's been circulating in developer circles: can you meaningfully route between models for cost or capability gains? Dax is skeptical of most commercial routing products, which he sees as inference middlemen inventing a problem to solve. The fundamental challenge is that switching models mid-session is a full cache bust — expensive and disruptive. The routing decision has to happen at the beginning, often before there's enough signal to make it well. But there's a version that does work: the orchestrator pattern. 1 — Dax Raad"The model routing category is inflated by middlemen desperate for a business model. Real value comes from the orchestrator pattern: an expe…"21:51 Some OpenCode team members run expensive, capable primary models configured to never take direct action — they only spawn cheaper subagent models to do the actual work. This nets out cheaper overall because the expensive model's intelligence is applied only to orchestration, while the grunt work goes to commodity-priced models. And newer-generation models are significantly better at this parallel orchestration than their predecessors, making it practical now in a way it wasn't a year ago. Wes digs into an earlier remark about burning tokens on 'some models not available yet.' Dax is visibly constrained — he posts things he thought were okay, got told they weren't. He can confirm that both OpenAI and Anthropic run significant preview programs, and that the latest model from one of them is responsible for the 5x token usage surge. 1 — Dax Raad"OpenCode's team describes itself as conservative and skeptical of AI hype. But after getting preview access to an unreleased next-generatio…"24:50 What's striking is his framing: OpenCode is a conservative team, historically skeptical of AI hype, not the 'AI psychosis people.' And yet. When preview access ended, nobody could talk about anything else. AI-generated funeral imagery circulated internally. 'What's the point of working anymore?' The model didn't feel smarter in a raw capability sense — it felt like a better partner. It listened. It picked up on things the user missed. It was trustworthy in a way previous models weren't. For Dax, that trust delta is what drove the usage explosion — and what makes the next public release genuinely exciting. After losing the unnamed preview model, the OpenCode team split — half went back to GPT-5.5, half migrated to GLM 5.2, an open-source model that Dax describes as very comparable in day-to-day use. Post-preview-model experience, the frontier models all feel 'roughly the same' anyway, so the choice is mostly pragmatic. The broader argument is that the gap between open-source and proprietary frontier models is closing — and that the economics differ dramatically. GLM 5.2 replacing GPT-5.5 is a meaningful data point. 1 — Dax Raad"GLM 5.2 comparable to GPT-5.5: Dax says open-source model GLM 5.2 is very comparable to OpenAI's GPT-5.5, with half the team switching to i…"26:55 Meanwhile, OpenCode's cheap Go plan — designed for international developers who can't afford frontier model pricing — has the US as its single largest subscriber base. The $200/month frontier model plan is out of reach for more people, even domestically, than the tech community typically assumes. Wes asks the inevitable: are AI inference costs going to spiral into thousands per employee per month? Dax has actual data to offer. At their 5x surge level, OpenCode's inference spend amounts to roughly 15% of their payroll — notable, but manageable for a tech company. 1 — Dax Raad"AI model cost ~15% of payroll: At 5x usage, OpenCode's AI inference costs represent roughly 15% of their total payroll — a manageable overh…"28:44 And the structural trajectory is downward. His estimate: Anthropic and OpenAI are currently running approximately 90% margin on inference, excluding R&D. Breakeven is 10x cheaper than current prices. 2 — Dax Raad"Dax estimates Anthropic and OpenAI are running ~90% margins on inference, not counting R&D. That means breakeven is 10x cheaper than curren…"29:10 Training losses are separate — they don't factor into inference economics. For open-source models, OpenCode can already host at a 70% discount to cost even using GPU middlemen. Direct GPU ownership would approach those same 90% margins. The narrative that OpenAI and Anthropic are perpetually unprofitable confuses training investment with inference margins — these are very different line items on the P&L. Wes asks the inevitable: are AI inference costs going to spiral into thousands per employee per month? Dax has actual data to offer. At their 5x surge level, OpenCode's inference spend amounts to roughly 15% of their payroll — notable, but manageable for a tech company. 1 — Dax Raad"AI model cost ~15% of payroll: At 5x usage, OpenCode's AI inference costs represent roughly 15% of their total payroll — a manageable overh…"28:44 And the structural trajectory is downward. His estimate: Anthropic and OpenAI are currently running approximately 90% margin on inference, excluding R&D. Breakeven is 10x cheaper than current prices. 2 — Dax Raad"Dax estimates Anthropic and OpenAI are running ~90% margins on inference, not counting R&D. That means breakeven is 10x cheaper than curren…"29:10 Training losses are separate — they don't factor into inference economics. For open-source models, OpenCode can already host at a 70% discount to cost even using GPU middlemen. Direct GPU ownership would approach those same 90% margins. The narrative that OpenAI and Anthropic are perpetually unprofitable confuses training investment with inference margins — these are very different line items on the P&L. Scott asks directly about the Claude Code situation — confused policy, Max plan integration being blocked, unclear stance on third-party wrappers. Dax offers a structural explanation rather than a grievance. Anthropic's business model is consumer-to-enterprise: Claude Code hooks people cheaply, those users convert their companies, companies pay per-token enterprise prices. OpenCode disrupts this funnel by letting users switch freely to any model. The Max plan integration was definitely blocked after a fight; the SDK/headless usage is in a gray area for now. 1 — Dax Raad"Claude Code exists as top-of-funnel to get companies hooked and convert them to per-token enterprise pricing. Tools like OpenCode disrupt t…"34:10 The deeper issue is cultural: Anthropic isn't consumer-oriented in the way OpenAI is. OpenAI will burn and raise unlimited money to maximize reach; Anthropic has an enterprise sales team that needs compute allocation to justify itself. Any compute going to consumer-friendly paths is compute competing with enterprise deals. Dax doesn't cast this as villainous — just structurally different from OpenAI's posture. Scott asks directly about the Claude Code situation — confused policy, Max plan integration being blocked, unclear stance on third-party wrappers. Dax offers a structural explanation rather than a grievance. Anthropic's business model is consumer-to-enterprise: Claude Code hooks people cheaply, those users convert their companies, companies pay per-token enterprise prices. OpenCode disrupts this funnel by letting users switch freely to any model. The Max plan integration was definitely blocked after a fight; the SDK/headless usage is in a gray area for now. 1 — Dax Raad"Claude Code exists as top-of-funnel to get companies hooked and convert them to per-token enterprise pricing. Tools like OpenCode disrupt t…"34:10 The deeper issue is cultural: Anthropic isn't consumer-oriented in the way OpenAI is. OpenAI will burn and raise unlimited money to maximize reach; Anthropic has an enterprise sales team that needs compute allocation to justify itself. Any compute going to consumer-friendly paths is compute competing with enterprise deals. Dax doesn't cast this as villainous — just structurally different from OpenAI's posture. Wes asks whether a future exists where the best models are only accessible through a lab's own app. Dax thinks it's a real risk. Product teams at these labs will always push for model-product bundling; it's good lock-in strategy. But the sales organization's revenue targets create counterpressure — APIs generate money that product-only strategies can't always match. The real danger is the unfair button: safety claims used as cover for competitive restriction. 'This model is too dangerous to use outside our harness' is a much more palatable public justification than 'we want the market share.' 1 — Dax Raad"Dax thinks AI models do have genuine potential for harm, and some government review makes sense. But when labs publicly claim their models …"37:30 And regulation that accepts this framing uncritically could codify harmful asymmetry. His broader warning: labs making extravagant public claims about the danger of their models are playing with political fire that could result in outcomes nobody wants. Scott asks what MCP tools and techniques are actually worth using today. Dax's personal setup is relatively vanilla, but the team's Discord bot is anything but. Running OpenCode as a Discord bot with a full suite of MCP servers connected to their data lake, the team can ask it complex analytical questions about subscriber behavior and get answers immediately. 1 — Dax Raad"OpenCode runs as a Discord bot for their internal team, with MCPs connected to all company systems including a full data lake. Team members…"39:55 The 'Gang Growth' skill invites collective team participation on design decisions. The pattern: always tag OpenCode before tagging a person. Then the voice prompting admission — the whole team, including casual Discord messages, runs on voice now. Kit Langdon introduced it; once you see someone doing it, it unlocks something. 2 — Dax Raad"OpenCode's team now uses voice prompting for everything — not just AI coding sessions, but casual Discord messages too. The LLM's tolerance…"42:20 Dax uses Handy on Linux, Hex on Mac. Scott has a foot pedal; Wes double-taps a mouse button. Someone sent Wes a prototype ring with a button on it. The LLM's tolerance for messy, rambling, imprecise speech is exactly what makes voice input viable where it failed before. Wes raises the emerging question of what happens when an AI coding agent needs to show you something a terminal can't. Dax is planning an artifacts feature — think Claude's canvas-style HTML/SVG documents — for OpenCode. MCP UI is on the radar, especially for non-technical users who'd benefit from richer dynamic interfaces. But the dynamic part of the spec is underdeveloped, and current implementations feel slow. Wes is bullish on the concept but agrees it's not yet worth building around. The consensus: it's a clear direction, just not the right moment to commit. For his sick pick, Dax recommends exe.dev — not just because it's good, but because it fills a gap that was genuinely unserved. You could rent cloud servers, but they had slow disks. You could find cheap bare-metal, but the providers were unreliable to the point of comedy. 1 — Dax Raad"The market for fast, persistent, affordable servers has historically been filled by unreliable or outright fraudulent VPS providers — inclu…"47:35 Dax proves the point with a story: years ago he found a cheap VPS provider in Miami. The provider sent an email saying he was going in for a medical procedure and would be unavailable for three days. Three days later, the server went down. A month passed with no word. A forum investigation eventually connected this provider to a previous VPS service that had disappeared under similar circumstances. Dax still doesn't understand the scam — the guy was being paid for a service, so why disappear? exe.dev, with the Tailscale DNA behind it, promises the opposite of this experience. bare metal server A physical server rented directly from a provider without a virtualization layer, offering better performance faster CPUs, NVMe disks than typical cloud VMs. tmux A terminal multiplexer that lets you run multiple persistent terminal sessions, split panes, and reconnect to sessions after disconnecting — used here for always-on dev and agent sessions. NVMe Non-Volatile Memory Express, a high-speed storage protocol for SSDs that dramatically outperforms older SATA drives; important for snappy remote dev server performance. MCP Model Context Protocol — a standard for giving AI models access to external tools, data sources, and services as structured callable functions. orchestrator pattern An AI architecture where a high-capability primary model never acts directly but instead spawns cheaper subagent models to carry out tasks, balancing intelligence and cost. TUI Text User Interface — a terminal-based graphical interface built with text characters and colors, as opposed to a GUI running in a windowed environment. Tailscale A VPN mesh networking tool that makes devices on different networks securely accessible to each other using WireGuard, known for its ease of use. inference The process of running a trained AI model to generate outputs; distinct from training. 'Inference costs' refer to compute charges each time you query an AI model. cache bust In AI inference, switching models mid-session invalidates the cached context, forcing the new model to re-process the full conversation history at significant extra cost. top of funnel The early stage of a sales or customer acquisition funnel designed to attract broad awareness; used here to describe Claude Code as a free/cheap entry point to Anthropic's ecosystem. Electron A framework for building cross-platform desktop applications using web technologies HTML, CSS, JavaScript ; OpenCode's desktop app migrated to it from Tauri. Tauri A framework for building lightweight desktop apps using Rust for the backend and web technologies for the UI; OpenCode originally used it before moving to Electron. vibe code Informal term for AI-assisted coding via natural language prompts rather than explicit programming; implies a loose, conversational style of directing an AI agent. primitives Low-level, foundational building blocks in software e.g. a robust diff library or rendering engine that higher-level abstractions are built upon; high quality primitives lift all products built on them. bespoke Custom-made or tailored to a specific need; used here to describe development environments that are complex and unique to a given application's build requirements. GLM 5.2 An open-source large language model from Zhipu AI's GLM series that Dax says is comparable in capability to OpenAI's GPT-5.5 for coding tasks. Chapter 3 · 01:22 Remote Development Environments and Their Benefits Prompted by a tweet Dax had posted the night before, Wes asks about his remote bare-metal dev setup. Dax explains that he rented a serious server — an AMD 990X, 192GB RAM, $200/month — not from a major cloud provider but from a bare-metal host near his city. 1 — Dax Raad"$200/month bare metal server: Dax runs a personal AMD 990X server with 192GB of RAM for $200 a month, cheaper than buying a MacBook."06:50 The killer feature isn't raw performance; it's permanence. Always-on tmux sessions, no hardware upgrade treadmill, seamless multi-device continuity. What started as a niche setup for Vim users has become dramatically more accessible now that many developers interact with machines primarily through AI coding agents rather than local editors. As OpenCode's team has grown, they've provisioned geo-distributed servers via latitude.sh — $300–$400/month each — across Europe, the US, and Singapore. But the real recommendation is exe.dev, a new turnkey product from the former Tailscale founder that productizes exactly this setup. 2 — Dax Raad"Dax switched to a permanently running bare-metal server AMD 990X, 192GB RAM, $200/month years ago and never looked back. The setup enable…"01:49 Dax has found it has the same 'just works' reliability as Tailscale itself. Dax switched to a permanently running bare-metal server AMD 990X, 192GB RAM, $200/month years ago and never looked back. The setup enables seamless multi-device work, avoids the upgrade treadmill, and has become essential for running long-lived AI coding agent sessions. 1:49 5:05 Chapter 4 · 06:24 The Setup: Tmux and Long-Running Sessions Prompted by a tweet Dax had posted the night before, Wes asks about his remote bare-metal dev setup. Dax explains that he rented a serious server — an AMD 990X, 192GB RAM, $200/month — not from a major cloud provider but from a bare-metal host near his city. 1 — Dax Raad"$200/month bare metal server: Dax runs a personal AMD 990X server with 192GB of RAM for $200 a month, cheaper than buying a MacBook."06:50 The killer feature isn't raw performance; it's permanence. Always-on tmux sessions, no hardware upgrade treadmill, seamless multi-device continuity. What started as a niche setup for Vim users has become dramatically more accessible now that many developers interact with machines primarily through AI coding agents rather than local editors. As OpenCode's team has grown, they've provisioned geo-distributed servers via latitude.sh — $300–$400/month each — across Europe, the US, and Singapore. But the real recommendation is exe.dev, a new turnkey product from the former Tailscale founder that productizes exactly this setup. 2 — Dax Raad"Dax switched to a permanently running bare-metal server AMD 990X, 192GB RAM, $200/month years ago and never looked back. The setup enable…"01:49 Dax has found it has the same 'just works' reliability as Tailscale itself. Dax runs a personal AMD 990X server with 192GB of RAM for $200 a month, cheaper than buying a MacBook. Chapter 5 · 07:15 Brought to you by Sentry Scott Tolinski delivers the Sentry sponsorship read, emphasizing the tool's dual role in performance monitoring and error tracking. He notes that production bugs often go undetected without proper visibility tooling, and that the Syntax team has used Sentry for years. The offer: two months free at sentry.io/syntax. OpenCode's team servers for Europe, US, and Singapore run $300–$400 per month each — still cheaper than purchasing laptops for employees. Chapter 6 · 08:12 Integrating AI with Personal Projects Scott asks about tmux configuration and Dax offers a clean, muscle-memory-friendly setup: one named session per project, each with a standard window/pane layout so his hands always know where to go. The more interesting reveal is what's running inside those sessions — not just dev work, but persistent personal AI agents. One session tracks his gym performance in SQLite and reminds him about form cues between sets. Another is synced with his iMessage, giving him an always-available AI contact reachable from anywhere. 1 — Dax Raad"Dax configured OpenCode to message his wife via iMessage and suggest a gift. She replied furiously that she'd divorce him if he used AI for…"09:00 That second one led to the cold open story: the AI he pointed at his wife to buy a gift, which detected her mounting rage and terminated itself. The section lands a broader point — having an always-on cloud machine isn't just a developer productivity hack, it's an infrastructure layer for life automation. Dax configured OpenCode to message his wife via iMessage and suggest a gift. She replied furiously that she'd divorce him if he used AI for gifts. Claude detected her anger, declared it couldn't continue, and terminated its own system service. When Dax used an OpenCode session with Claude to buy his wife a gift via iMessage and she got angry, Claude detected her upset tone and shut down the service entirely. Wes asks about OpenCode 2.0 and Dax answers with characteristic self-awareness: 'My entire career it's always taken me three swings to get something right.' OpenCode 0 and 1 were the first two swings; 2.0 is the informed rewrite. The architectural centerpiece is running OpenCode as a persistent local service rather than a session-scoped process — everything syncs automatically across desktop and web, and users can write their own scripts and apps against it. More striking is the multi-host device graph: Dax's Linux server, Mac Studio, and Framework desktop are all Tailscale-connected and OpenCode-aware. 1 — Dax Raad"OpenCode 2.0 is a ground-up rewrite with a cleanly designed API, always-on service architecture, multi-host device awareness, and a new plu…"11:31 When the Linux server wants to send an iMessage, it SSHs into the Mac and uses that browser session. No special plumbing required — just the fact of connectivity is enough. Electron replaced Tauri for the desktop app though the desktop was never formally released out of beta . The beta was targeted for end of the recording week, full release a month after. OpenCode 2.0 is a ground-up rewrite with a cleanly designed API, always-on service architecture, multi-host device awareness, and a new plugin API. The team burned massive tokens on exhaustive design exploration — researching every prior art and possible API shape before committing to anything. OpenCode 2.0, a major rewrite with a new API and always-running service model, was planned for beta release by end of the recording week. Chapter 8 · 17:15 Software Engineering Methodology and Design Philosophy Prompted by a genuine question about polish — why does OpenCode feel so much better than Anthropic's own Claude Code? — Dax gives a surprisingly philosophical answer. Step one is simply deciding that quality matters, which he frames as non-obvious given that billion-dollar companies ship rough UX and succeed anyway. For OpenCode, caring is a prerequisite to everything else. Step two is investing in primitives: OpenTUI, a Zig-based TUI framework their team built from scratch, is painstakingly correct across all platforms and enables even non-experts to build high-quality terminal interfaces on top. 1 — Dax Raad"OpenCode's terminal UI quality comes from OpenTUI, an open-source TUI framework written in Zig that's extremely performant across all platf…"20:35 And then there's the token strategy — deliberately burning tokens to 'indulgently over-design everything.' Before committing to any API shape, the team researches all prior art, generates every possible structure, and compares exhaustively. This was never feasible before LLMs. The result is software that reflects more considered decisions. Wes draws a parallel to Pierre Computer, whose simple, well-designed diff and sidebar tree primitives have become the foundation for much of the agentic coding UI ecosystem. OpenCode's terminal UI quality comes from OpenTUI, an open-source TUI framework written in Zig that's extremely performant across all platforms. It supports React, SolidJS, and potentially Vue bindings, letting average developers build high-quality terminal apps on solid expert-built foundations. 20:35 21:40 Chapter 9 · 21:51 Model Routing and AI Integration Scott raises the model routing question that's been circulating in developer circles: can you meaningfully route between models for cost or capability gains? Dax is skeptical of most commercial routing products, which he sees as inference middlemen inventing a problem to solve. The fundamental challenge is that switching models mid-session is a full cache bust — expensive and disruptive. The routing decision has to happen at the beginning, often before there's enough signal to make it well. But there's a version that does work: the orchestrator pattern. 1 — Dax Raad"The model routing category is inflated by middlemen desperate for a business model. Real value comes from the orchestrator pattern: an expe…"21:51 Some OpenCode team members run expensive, capable primary models configured to never take direct action — they only spawn cheaper subagent models to do the actual work. This nets out cheaper overall because the expensive model's intelligence is applied only to orchestration, while the grunt work goes to commodity-priced models. And newer-generation models are significantly better at this parallel orchestration than their predecessors, making it practical now in a way it wasn't a year ago. The model routing category is inflated by middlemen desperate for a business model. Real value comes from the orchestrator pattern: an expensive, smart primary model that never acts directly, instead spawning cheaper subagents for all tasks. It ends up cheaper and more capable. Some OpenCode team members run an expensive orchestrator model that never acts directly, instead spawning cheaper subagents for all tasks, which ends up being cheaper overall. Chapter 10 · 23:58 The Evolution of AI Models Wes digs into an earlier remark about burning tokens on 'some models not available yet.' Dax is visibly constrained — he posts things he thought were okay, got told they weren't. He can confirm that both OpenAI and Anthropic run significant preview programs, and that the latest model from one of them is responsible for the 5x token usage surge. 1 — Dax Raad"OpenCode's team describes itself as conservative and skeptical of AI hype. But after getting preview access to an unreleased next-generatio…"24:50 What's striking is his framing: OpenCode is a conservative team, historically skeptical of AI hype, not the 'AI psychosis people.' And yet. When preview access ended, nobody could talk about anything else. AI-generated funeral imagery circulated internally. 'What's the point of working anymore?' The model didn't feel smarter in a raw capability sense — it felt like a better partner. It listened. It picked up on things the user missed. It was trustworthy in a way previous models weren't. For Dax, that trust delta is what drove the usage explosion — and what makes the next public release genuinely exciting. OpenCode's team describes itself as conservative and skeptical of AI hype. But after getting preview access to an unreleased next-generation model, their monthly token usage jumped 5x. When the preview period ended, the whole team mourned. After losing the unnamed preview model, the OpenCode team split — half went back to GPT-5.5, half migrated to GLM 5.2, an open-source model that Dax describes as very comparable in day-to-day use. Post-preview-model experience, the frontier models all feel 'roughly the same' anyway, so the choice is mostly pragmatic. The broader argument is that the gap between open-source and proprietary frontier models is closing — and that the economics differ dramatically. GLM 5.2 replacing GPT-5.5 is a meaningful data point. 1 — Dax Raad"GLM 5.2 comparable to GPT-5.5: Dax says open-source model GLM 5.2 is very comparable to OpenAI's GPT-5.5, with half the team switching to i…"26:55 Meanwhile, OpenCode's cheap Go plan — designed for international developers who can't afford frontier model pricing — has the US as its single largest subscriber base. The $200/month frontier model plan is out of reach for more people, even domestically, than the tech community typically assumes. Contrary to expectations, the US is the top subscriber country for OpenCode's cheap Go plan, designed for open-source models, showing price sensitivity even among US developers. Chapter 12 · 28:27 Cost Implications of AI Model Usage Wes asks the inevitable: are AI inference costs going to spiral into thousands per employee per month? Dax has actual data to offer. At their 5x surge level, OpenCode's inference spend amounts to roughly 15% of their payroll — notable, but manageable for a tech company. 1 — Dax Raad"AI model cost ~15% of payroll: At 5x usage, OpenCode's AI inference costs represent roughly 15% of their total payroll — a manageable overh…"28:44 And the structural trajectory is downward. His estimate: Anthropic and OpenAI are currently running approximately 90% margin on inference, excluding R&D. Breakeven is 10x cheaper than current prices. 2 — Dax Raad"Dax estimates Anthropic and OpenAI are running ~90% margins on inference, not counting R&D. That means breakeven is 10x cheaper than curren…"29:10 Training losses are separate — they don't factor into inference economics. For open-source models, OpenCode can already host at a 70% discount to cost even using GPU middlemen. Direct GPU ownership would approach those same 90% margins. The narrative that OpenAI and Anthropic are perpetually unprofitable confuses training investment with inference margins — these are very different line items on the P&L. Dax estimates Anthropic and OpenAI are running ~90% margins on inference, not counting R&D. That means breakeven is 10x cheaper than current prices. For open-source models with middlemen, OpenCode already hosts at 70% discount to cost. Wes asks the inevitable: are AI inference costs going to spiral into thousands per employee per month? Dax has actual data to offer. At their 5x surge level, OpenCode's inference spend amounts to roughly 15% of their payroll — notable, but manageable for a tech company. 1 — Dax Raad"AI model cost ~15% of payroll: At 5x usage, OpenCode's AI inference costs represent roughly 15% of their total payroll — a manageable overh…"28:44 And the structural trajectory is downward. His estimate: Anthropic and OpenAI are currently running approximately 90% margin on inference, excluding R&D. Breakeven is 10x cheaper than current prices. 2 — Dax Raad"Dax estimates Anthropic and OpenAI are running ~90% margins on inference, not counting R&D. That means breakeven is 10x cheaper than curren…"29:10 Training losses are separate — they don't factor into inference economics. For open-source models, OpenCode can already host at a 70% discount to cost even using GPU middlemen. Direct GPU ownership would approach those same 90% margins. The narrative that OpenAI and Anthropic are perpetually unprofitable confuses training investment with inference margins — these are very different line items on the P&L. OpenCode can host some open-source models at a 70% discount even when using middlemen for GPU hosting, enabling massive margins. Chapter 14 · 32:48 Navigating Claude Code and Third-Party Integration Scott asks directly about the Claude Code situation — confused policy, Max plan integration being blocked, unclear stance on third-party wrappers. Dax offers a structural explanation rather than a grievance. Anthropic's business model is consumer-to-enterprise: Claude Code hooks people cheaply, those users convert their companies, companies pay per-token enterprise prices. OpenCode disrupts this funnel by letting users switch freely to any model. The Max plan integration was definitely blocked after a fight; the SDK/headless usage is in a gray area for now. 1 — Dax Raad"Claude Code exists as top-of-funnel to get companies hooked and convert them to per-token enterprise pricing. Tools like OpenCode disrupt t…"34:10 The deeper issue is cultural: Anthropic isn't consumer-oriented in the way OpenAI is. OpenAI will burn and raise unlimited money to maximize reach; Anthropic has an enterprise sales team that needs compute allocation to justify itself. Any compute going to consumer-friendly paths is compute competing with enterprise deals. Dax doesn't cast this as villainous — just structurally different from OpenAI's posture. Claude Code exists as top-of-funnel to get companies hooked and convert them to per-token enterprise pricing. Tools like OpenCode disrupt that funnel by letting users switch models. Anthropic's compute limitations and sales incentives make it structurally hostile to third-party integrations. Scott asks directly about the Claude Code situation — confused policy, Max plan integration being blocked, unclear stance on third-party wrappers. Dax offers a structural explanation rather than a grievance. Anthropic's business model is consumer-to-enterprise: Claude Code hooks people cheaply, those users convert their companies, companies pay per-token enterprise prices. OpenCode disrupts this funnel by letting users switch freely to any model. The Max plan integration was definitely blocked after a fight; the SDK/headless usage is in a gray area for now. 1 — Dax Raad"Claude Code exists as top-of-funnel to get companies hooked and convert them to per-token enterprise pricing. Tools like OpenCode disrupt t…"34:10 The deeper issue is cultural: Anthropic isn't consumer-oriented in the way OpenAI is. OpenAI will burn and raise unlimited money to maximize reach; Anthropic has an enterprise sales team that needs compute allocation to justify itself. Any compute going to consumer-friendly paths is compute competing with enterprise deals. Dax doesn't cast this as villainous — just structurally different from OpenAI's posture. Wes asks whether a future exists where the best models are only accessible through a lab's own app. Dax thinks it's a real risk. Product teams at these labs will always push for model-product bundling; it's good lock-in strategy. But the sales organization's revenue targets create counterpressure — APIs generate money that product-only strategies can't always match. The real danger is the unfair button: safety claims used as cover for competitive restriction. 'This model is too dangerous to use outside our harness' is a much more palatable public justification than 'we want the market share.' 1 — Dax Raad"Dax thinks AI models do have genuine potential for harm, and some government review makes sense. But when labs publicly claim their models …"37:30 And regulation that accepts this framing uncritically could codify harmful asymmetry. His broader warning: labs making extravagant public claims about the danger of their models are playing with political fire that could result in outcomes nobody wants. Dax thinks AI models do have genuine potential for harm, and some government review makes sense. But when labs publicly claim their models are nuclear-level dangerous, they attract irrational political attention that could result in overly aggressive or corrupt regulation. Scott asks what MCP tools and techniques are actually worth using today. Dax's personal setup is relatively vanilla, but the team's Discord bot is anything but. Running OpenCode as a Discord bot with a full suite of MCP servers connected to their data lake, the team can ask it complex analytical questions about subscriber behavior and get answers immediately. 1 — Dax Raad"OpenCode runs as a Discord bot for their internal team, with MCPs connected to all company systems including a full data lake. Team members…"39:55 The 'Gang Growth' skill invites collective team participation on design decisions. The pattern: always tag OpenCode before tagging a person. Then the voice prompting admission — the whole team, including casual Discord messages, runs on voice now. Kit Langdon introduced it; once you see someone doing it, it unlocks something. 2 — Dax Raad"OpenCode's team now uses voice prompting for everything — not just AI coding sessions, but casual Discord messages too. The LLM's tolerance…"42:20 Dax uses Handy on Linux, Hex on Mac. Scott has a foot pedal; Wes double-taps a mouse button. Someone sent Wes a prototype ring with a button on it. The LLM's tolerance for messy, rambling, imprecise speech is exactly what makes voice input viable where it failed before. OpenCode runs as a Discord bot for their internal team, with MCPs connected to all company systems including a full data lake. Team members tag it before tagging a human — it often solves the question on its own. The 'Gang Growth' skill helps the team work through design decisions collectively. OpenCode's team now uses voice prompting for everything — not just AI coding sessions, but casual Discord messages too. The LLM's tolerance for rambling and imprecise speech makes it work where traditional voice input would fail. Scott uses a foot pedal; Wes double-taps his mouse button. Wes raises the emerging question of what happens when an AI coding agent needs to show you something a terminal can't. Dax is planning an artifacts feature — think Claude's canvas-style HTML/SVG documents — for OpenCode. MCP UI is on the radar, especially for non-technical users who'd benefit from richer dynamic interfaces. But the dynamic part of the spec is underdeveloped, and current implementations feel slow. Wes is bullish on the concept but agrees it's not yet worth building around. The consensus: it's a clear direction, just not the right moment to commit. For his sick pick, Dax recommends exe.dev — not just because it's good, but because it fills a gap that was genuinely unserved. You could rent cloud servers, but they had slow disks. You could find cheap bare-metal, but the providers were unreliable to the point of comedy. 1 — Dax Raad"The market for fast, persistent, affordable servers has historically been filled by unreliable or outright fraudulent VPS providers — inclu…"47:35 Dax proves the point with a story: years ago he found a cheap VPS provider in Miami. The provider sent an email saying he was going in for a medical procedure and would be unavailable for three days. Three days later, the server went down. A month passed with no word. A forum investigation eventually connected this provider to a previous VPS service that had disappeared under similar circumstances. Dax still doesn't understand the scam — the guy was being paid for a service, so why disappear? exe.dev, with the Tailscale DNA behind it, promises the opposite of this experience. The market for fast, persistent, affordable servers has historically been filled by unreliable or outright fraudulent VPS providers — including one Dax used who faked his own death. exe.dev, from the former Tailscale founder, productizes the bare-metal remote dev setup with the same 'just works' reliability as Tailscale. A cheap VPS provider Dax used years ago faked his own death to disappear from the service, taking the servers offline, a pattern he later discovered the provider had done before. Dax configured OpenCode to message his wife via iMessage and suggest a gift. She replied furiously that she'd divorce him if he used AI for gifts. Claude detected her anger, declared it couldn't continue, and terminated its own system service. OpenCode's team describes itself as conservative and skeptical of AI hype. But after getting preview access to an unreleased next-generation model, their monthly token usage jumped 5x. When the preview period ended, the whole team mourned. Dax estimates Anthropic and OpenAI are running ~90% margins on inference, not counting R&D. That means breakeven is 10x cheaper than current prices. For open-source models with middlemen, OpenCode already hosts at 70% discount to cost. 29:10 32:30 Snapshots Key Quotes Sign in to keep viewing Create a free account to keep exploring this episode's insights, snapshots and quotes. AI lab discussed for its inference pricing margins, Claude Code product strategy, and restrictive stance on third-party API integrations. AI lab discussed alongside Anthropic for inference margins, and contrasted with Anthropic for being more consumer-oriented and allowing OpenCode integration. Bare-metal server hosting provider currently used by OpenCode for their team dev servers across US, Europe, and Singapore. AI coding agent tool co-founded by Dax Raad; the central product discussed throughout the episode including its 2.0 rewrite. Anthropic's AI model; featured in the viral story about shutting itself down after detecting spousal anger, and discussed for its sensitivity and quality. Anthropic's coding agent product; discussed as Anthropic's consumer top-of-funnel product and compared unfavorably in UI quality to OpenCode. A new product from the former Tailscale founder offering turnkey bare-metal remote dev environments; Dax's sick pick for the episode. Open-source Zig-based TUI framework built by the OpenCode team; powers OpenCode's terminal interface and used by other TUI products. VPN mesh networking tool used by Dax to connect all his devices; exe.dev was founded by the former Tailscale founder and compared favorably to it. Open-source AI model described by Dax as comparable to GPT-5.5, used by half the OpenCode team after losing access to unreleased preview models. Episode sponsor; application monitoring and error tracking tool recommended for visibility into production application performance. AI coding tool mentioned in the context of its iOS app release and as a competitor in the agentic coding space. Mac voice prompting application created by Kit Langdon used by the OpenCode team for AI voice interaction. Stats Episode stats Insight Overview insights chapters Insight distribution Sub-Categories Speaker breakdown Talk Time This episode Claims & Sources 0 / 12 cited 0% Factual claims made this episode, and whether a source was named. ⚠ OpenCode's monthly token usage increased 5x over a couple of months after accessing a next-generation unreleased AI model. Dax Raadno source cited ⚠ A bare-metal AMD 990X server with 192GB RAM can be rented for approximately $200 per month. Dax Raadno source cited ⚠ OpenCode's professional team servers in Europe, US, and Singapore cost $300–$400 per month each. Dax Raadno source cited ⚠ At 5x token usage, OpenCode's AI inference costs represent roughly 15% of their total payroll. Dax Raadno source cited ⚠ Anthropic and OpenAI are likely making approximately 90% margin on inference, meaning breakeven could be 10x cheaper than current prices. Dax Raadno source cited ⚠ OpenCode can host some open-source models at a 70% discount even when using GPU hosting middlemen. Dax Raadno source cited ⚠ GLM 5.2 is very comparable in capability to GPT-5.5 for coding tasks. Dax Raadno source cited ⚠ The United States is the number one subscriber country for OpenCode's low-cost Go plan, which was designed for international developers. Dax Raadno source cited ⚠ A Claude model shut down its own system service after detecting that a user's wife was extremely angry during an iMessage conversation. Dax Raadno source cited ⚠ exe.dev was founded by the former Tailscale founder. Dax Raadno source cited ⚠ OpenCode's OpenTUI TUI framework is written in Zig and powers the OpenCode terminal interface. Dax Raadno source cited ⚠ The Groq CLI is built in Rust using a library called Ratatouille rather than OpenTUI. Dax Raadno source cited Sign up free to see the full analytics Cast, category & speaker breakdowns and fact-checked claims — free with an account.