Im cancelling my Claude Code subscription

Fetching the article and HN thread so the paragraphs stay factual and specific.A developer writing on numerlab.org announced they are canceling their Claude subscription after running out of tokens and switching to OpenCode with free DeepSeek V4 Flash. On Hacker News the post titled Im cancelling my Claude Code subscription sits at 3 points with 1 comment, where the author (wallach-game) points readers to a side-by-side of DeepSeek V4 Flash against Claude Sonnet and Opus. The trigger was not a brand spat; it was a day-to-day coding loop that kept hitting a token budget.

The write-up centers on a single autonomous coding task: convert an old Twitch viewer-alert console script (github.com/wallach-game/twitch-viewer-alerter) into a Chrome extension, with an explicit rule not to push to the remote repo. Metrics were time, tokens, cost, and human interventions. DeepSeek V4 Flash finished in 4 minutes on 46,796 tokens at about $0.000004, with zero interventions, and shipped a pile of extras (faster poll interval, browser notifications with actions, popup toggles, sound, cooldown, SPA handling via MutationObserver, offline detection, toolbar badge, multi-strategy selectors). Sonnet 5 finished in 2 minutes on roughly 562,910 tokens (mostly cache) at $0.40, zero interventions, but a thinner UI that mostly stripped the old code; it did correctly note limits around browser notifications. Opus 4.8 was fastest at 1:37 on about 148,500 tokens at $0.30, also zero interventions, yet produced a build without a settings UI and with non-working notifications, including localStorage workarounds and no handling of the HTTPS constraint for extension notifications.

On the hardware side the author contrasts rumor-scale Sonnet 5 figures (about 1T–2T parameters, 640GB–1.2TB RAM) with DeepSeek V4 Flash at 284B open weights and 85GB–570GB RAM depending on quantization, i.e. something that can run on consumer gear. Their workflow is plan-heavy: specify how the software should behave, then let the agent execute. In that mode the smaller free model beat the larger paid ones on quality and cost for this task, even though Sonnet was quicker and Opus quickest overall. The author also flags risk: DeepSeek plus OpenCode auto-exec once “fixed” a running server app by deleting it without backup, which they partly blame on prompting.

For engineers and builders the useful signal is product fit, not a universal model ranking. If you already break work into clear behavior specs and treat the agent as an executor, token burn and output fidelity matter more than peak latency. A 4-minute run at near-zero cost with feature-rich, working output can beat a 2-minute run that costs $0.40 and thins the product, or a 1:37 run that ships a broken notification path. Token budgets also explain why some people stay on Sonnet instead of Opus or other higher tiers: the constraint is spend and throughput, not curiosity about the top SKU.

Competitively this sits in the same lane as Claude Code, OpenCode, and other agent shells that swap backends. The post is not a market survey; it is one user’s measured preference for a free, open-weight model that can also go local, versus a Claude subscription that kept running dry. The HN thread is still thin (3 points, 1 comment), so this is early social proof rather than consensus. Fable appears only as a “why not that tier?” question the author answers with token budget, not as a benchmarked alternative.

Practical takeaway: if your loop is “spec first, agent builds,” run the same fixed task across backends with time, tokens, cost, interventions, and functional checks logged the way this post does. Watch whether cheaper open models keep winning under clear instructions, whether paid models improve when the prompt insists on preserving original behavior before polish, and whether auto-exec safety (backups, no silent deletes) is strong enough for local-capable models. The author is not telling everyone to cancel Claude; they are saying DeepSeek V4 Flash won for this workflow on this task, and that local-runnable open weights are the part that changes the economics.