Your API costs may look fine until they don't, here's what's eating your budget in the dark.

You're paying for thinking you never see. Models like o3 don't just answer your question. They "reason" through it first, step by step, behind the scenes. That reasoning gets billed as output tokens. But it doesn't show up anywhere obvious in your dashboard. A 2025 study out of Polytechnique Montréal looked at complex tasks and found hidden reasoning tokens can account for over 90% of total token spend. Ninety percent. You could be looking at a bill and have almost no idea what you're actually paying for. DigitalOcean's newsletter surfaced this problem and it deserves more attention than it's getting. If you're using AI through an API for anything beyond simple text generation, go read this now: DigitalOcean newsletter. The workarounds are real: flat-rate providers like Groq and Cerebras sidestep the per-token problem entirely, and DigitalOcean's new Inference Router tries to route tasks to the cheapest model that can handle them. Neither is a magic fix, but knowing the problem exists is step one.

The "AI writes bad code" argument just got a lot harder to make. A study of 220,612 pull requests across 539 Python repositories compared the major AI coding tools head to head. Claude Code merged 84.3% of the time. Codex hit 73.5%. Cursor was 63.9%, Copilot 59.6%, and Devin came in at 43%. Bug rates for Claude Code were comparable to, or lower than, human-written code. That's not a marketing claim from Anthropic. That's production data from real repositories. For any owner who has been sitting on the fence about whether to trust one of these tools with actual work, this is the most concrete comparison out there. Start with the DigitalOcean newsletter writeup for the summary.

Trending on GitHub

Homebrew/BrewUI, an official visual interface for Homebrew, the popular Mac software installer, so you don't have to use the command line (271 stars today)

alphaXiv/OpenResearch, turns AI coding agents into research agents that can read and synthesize academic papers automatically (531 stars today)

NationalSecurityAgency/ghidra, the NSA's open-source tool for reverse engineering software, used by security professionals to examine how programs work under the hood (725 stars today)

danny-avila/LibreChat, a self-hosted ChatGPT alternative you control, supporting Claude, GPT, Gemini, Mistral, and a dozen other models in one interface (254 stars today)

pacifio/atlas, source control for AI coding agents, so you can run multiple agents, track what each one changed, and query all of it from one place (91 stars today)

Also Worth Knowing

If your site was built with AI, an AI crawler may not be able to read it: Vendasta flags the discoverability blind spot that owners who used AI to build or update their sites should audit before AI-powered search grows further.

DigitalOcean's Inference Router is now live and automatically routes your AI tasks to the most cost-effective model for the job: details here.

GLM-5.3-Flash is currently DigitalOcean's cheapest model but defaults to maximum verbosity out of the box, meaning it burns tokens quietly if you don't configure it: worth knowing before you use it.

Lumpy mail (physical mailers and gifts) outperforms cold email 60 to 1 for booking sales meetings, according to Morningside AI's newsletter: the old-school tactic worth revisiting.

DigitalOcean's Open Intelligence Summit runs October 12 to 13 in San Francisco, focused on open-source AI and model portability: registration info here.

DOKS now supports Spot GPU instances with automatic fallback to On-Demand if a spot gets reclaimed, relevant if you're running any GPU workloads: see the update.

-Scott

SmartOwner is published (almost) daily by the team at DigitalTreehouse. Want AI consulting or automations for your business? Reply to this email.