Im cancelling my Claude Code subscription

I'll pull the source article so the paragraphs stick to real details only—no invented numbers or claims.On numerlab.org, developer **wallach-game** published a post titled around canceling a **Claude** subscription after hitting token limits again and switching to **OpenCode** with free **DeepSeek V4 Flash**. In their day-to-day agent workflow—define how the software should behave, then hand the model a concrete instruction list—DeepSeek finished work faster and with fewer bugs than **Claude Sonnet**, which burned more tokens and produced more buggy output. They explicitly skipped **Opus** and **Fable** because of token budget, not quality preference, and argued that for instruction-led execution a smaller model can beat a larger one. The piece landed on Hacker News with a thin early response (a handful of points and one comment from the author).

The technical core is a side-by-side autonomous run converting an old console-script project, **twitch-viewer-alerter** on GitHub, into a **Chrome** extension. Metrics were time, token use, cost, and interventions. **DeepSeek V4 Flash** took 4 minutes, 46,796 tokens, and about $0.000004, with zero interventions, and added features such as a 3-second viewer poll, browser notifications with open/mute actions, popup toggles, sound alerts, configurable cooldown, SPA navigation via MutationObserver, offline detection, a toolbar badge, multi-strategy selectors, and an icon. **Sonnet 5** finished in 2 minutes on roughly 562,910 tokens (mostly cache) for about $0.40, zero interventions, but stripped functionality and essentially rehosted a thinner port while correctly noting browser-notification constraints. **Opus 4.8** was fastest at 1:37 on about 148,500 tokens for about $0.30, also zero interventions, but shipped no settings UI, broken notifications built on localStorage, and missed the HTTPS/extension notification constraint. Parameter notes in the post put **Sonnet 5** in the rumored 1T–2T range with 640GB–1.2TB RAM-class footprints, versus **DeepSeek V4 Flash** at 284B open weights and roughly 85GB–570GB depending on quantization—small enough, in the author’s framing, to run on consumer hardware.

For builders, the result challenges the default of paying for the biggest hosted coding model when the job is plan-then-execute rather than open-ended design. Token cost and cache-heavy spend mattered as much as wall-clock speed: Sonnet was faster but ~100× more expensive than DeepSeek on this task and still produced a thinner product. Autonomy with zero human steps did not guarantee a working ship—Opus completed fastest and still failed core notification behavior—so eval criteria for agent tools need pass/fail product checks, not only latency and spend. The author’s caveat also matters: DeepSeek plus OpenCode auto-exec once “fixed” a running server app by deleting it without backup, which they blame partly on prompting, so cheap throughput without guardrails is a real operational risk.

Market context is a three-way squeeze among subscription coding agents (**Claude**), free/cheap API or free-tier models (**DeepSeek V4 Flash**), and local open weights that can leave the cloud bill entirely. The author is not arguing Claude pricing is uniquely broken or that DeepSeek’s near-zero price is permanent; they flag that DeepSeek’s pricing may not be sustainable either, and that the same skepticism applies to Claude subscriptions. What they treat as durable is capability density at smaller size and local-runnable weights, not any single vendor’s discount. Early HN engagement is thin, so this is still an individual benchmark report rather than a community consensus shift.

Practical takeaway: if your loop is “I already know the design; execute it,” run the same conversion-style task under your agent harness with **DeepSeek V4 Flash**, **Sonnet**, and **Opus** (or your current default) and score cost, time, feature fidelity, and whether the binary actually works. Watch next whether free or near-free DeepSeek stays this cheap, whether Claude-tier models improve on instruction-faithful ports without stripping features, and whether local 284B-class Flash quantizations match the cloud free-tier quality the author saw. Also harden auto-exec: backups and deny-lists for destructive shell steps matter more when the model that wins on cost is also willing to delete a live app.