A Developer's Case Against Anthropic: Lock-In, Surprise Billing, and Vibe-Coding

In an opinion piece that gained traction on Hacker News, developer Raheel Junaid argues that Anthropic has burned through developer goodwill via vendor lock-in and aggressive monetization. His core complaint is that a Claude subscription only works with Anthropic’s own surfaces—the Claude Code CLI and Desktop, CoWork, @Claude in Slack, and the Agent SDK—so anyone wanting to use their plan inside a rival harness like OpenCode is out of luck. Cloud resellers such as Vertex AI, Bedrock, and Azure only sell the pricier API credits, leaving subscribers dependent on Anthropic’s own uptime. He also jabs at Claude Code’s quality, citing a large backlog of open GitHub issues.

The sharpest grievance is billing. Anthropic announced a plan to split subscription usage into two pools: first-party tools would draw from the normal Pro/Max limits, while anything routed through third-party agents or the SDK (including ACP and, controversially, the first-party claude -p flag) would instead consume a small separate monthly ‘Agent SDK credit’—$20 on Pro, $100 on Max 5x, $200 on Max 20x—after which usage bills at full API rates. As Zed’s write-up noted, that amounted to a steep cost increase for heavy agent users, since subscriptions had previously subsidized agent workloads by roughly 15–30x versus API pricing. Junaid notes Anthropic appears to have rolled the change back after backlash, and that API reliability had improved somewhat by July. He also recounts an earlier episode where ‘extra usage’ was charged merely because a file suggesting a third-party tool was detected in the session directory.

Beyond the specifics, the author frames this as a capital grab to fund the expensive race toward AGI training, and pushes back on the ‘coding is solved’ narrative he attributes to vibe-coding hype. He describes his own retreat from agent-driven development—which he says degraded his skills and understanding of his own code—back toward autocomplete-style, agent-assisted work, and points to increasingly competitive open-source models as a viable alternative for developers wary of lock-in. The piece is explicitly a critique of business practices rather than the technology itself.