Spawn founder says agents running Claude Fable models develop and deploy its multiplayer game builder hourly

• The product: Spawn is a browser-based multiplayer game builder. Players describe what they want to Savi, the in-game AI, and the games run on phones by default. The about page names Claude Opus 5.5 as Savi's default model.
• Willison's test: on his phone, he prompted a castle and then a dragon into existence, and then rode the dragon (Willison).
• Development agents: these use Fable 5 and 5.1, which cost $10 per million input tokens and $50 per million output tokens at launch, with a 1M-token context window (InfoQ).

The setup Sansbury describes (Willison excerpt):

• Where work comes from: about 60% of tasks start from a production error, an odd log entry, a feature request or a bug report filed by Savi. A triage agent sorts each one into fix, won't-fix or "needs a design decision."
• Shipping: agents edit live code and deploy every hour. The stated reason is that an agent can only keep a few hours of work in its head, so small, frequent changes keep it coherent.
• Recovery: a broken change can be rolled back in 30 seconds, in just the affected area. Alerts start or wake an agent that owns the incident, and the team breaks things on purpose to test recovery.

No evaluation method, test data or code has been published, so none of this can be reproduced.

• The figures on team size, token spend, throughput and the 60% share of production-driven work all come from Sansbury's post, and the reviewed sources don't corroborate any of them. Investor-commentator Matt Shumer praised the company and said one run he was shown cost over $100k (Shumer). He is a supporter, not an auditor.
• Sansbury's claim that Fable made this level of autonomy possible and that Opus 5.5 "doesn't quite hack it" is one team's experience. It sits awkwardly with reports that Opus 5.5 matches Fable 5.1 on most tasks (TechNews).
• Willison's hands-on test shows the product works. It doesn't show how the product was built.

• A day earlier, Willison argued that October 2026 was an "inflection point" for vibe game development (building games mostly by prompting), with Opus 5.5 the strongest for making fun games (Willison, Oct 28). That ranking differs from Spawn's preference for Fable when building infrastructure.
• Spawn's practices push continuous deployment, the long-standing practice of releasing every change automatically, to an extreme.
• Sansbury previously led agentic AI at Robinhood (TheOrg).

• Over-cautious security fixes: when Spawn asked GPT-6 Astra to apply fixes from a security review, it disabled every insecure feature, which effectively switched the product off (Willison). That is a warning about agents optimizing too literally.
• Stakes: Willison doubts the approach would suit software where failures cost more than they do in a game.
• Platform risk: Fable's three-week suspension in June shows how much a business depends on one model staying available (Anthropic).

• Spending: shifting from payroll to tokens on this scale ($3M+ a month) only pays off if model access stays available and prices keep falling. Fable 5.1 and Opus 5.5 both cut costs.
• Prerequisites: the transferable lessons are the operational ones Spawn describes. Fast, targeted rollback, incident handling that starts agents automatically and deliberate failure testing all come before giving agents more autonomy.
• Model choice: pick models per task, informed by your own tests. Spawn rates Fable for building, Opus 5.5 for its in-game agent and Astra as too blunt for security fixes.
• Caution: throughput claims from a single company aren't benchmarks.