Max Woolf traces months of falling model prices and argues that competing models and hosting companies constrain OpenAI and Anthropic. Coding-agent interfaces make trying alternatives easy, weakening the assumption that today’s cheap usage must eventually become expensive.

  • Repeated context matters: cached-input pricing covers reusing previously processed conversation text. Agents send long histories repeatedly, so cuts here can change the bill substantially.
  • Headline rates miss that: Woolf reports Fable 5.1 cut cached-input prices by 75% while keeping its ordinary input/output rates unchanged. Actual task savings depend on how much context gets reused.
  • The alternatives have suppliers: he counts 29 hosts offering DeepSeek V4.1 Flash 1017. Competition can come from different companies serving the same downloadable model.
  • Subscriptions follow different rules: cheaper API usage can coexist with tighter subscription allowances. Woolf says cheaper models let him downgrade his own Codex plan.

Woolf makes a case for competition and efficiency, but the article cannot establish the labs’ private margins or isolate what caused each cut. Its practical implication is to compare the cost of finished tasks, including retries and quality checks, alongside the advertised token rates.

The 177-comment thread on Hacker News adds reasons switching might succeed and reasons cheaper service needs checking.

What the thread adds

  • jameson — questions how customers can verify unchanged quality: “it’s difficult to know if the vendor’s supposed “optimization” affects performance negatively because it’s difficult to define an SLA”. An SLA is a contractual service guarantee.
  • buckleyourshoe — reports a switch prompted by refusals as well as price: “We tried Kimi and GLM models after hitting ‘cybersecurity’ blocks on OpenAI and Anthropic.” Their account is anecdotal.
  • Iolaum — pushes back on another commenter’s hope for unlimited subscriptions: “But if you get unlimited tokens, with scheduled/automated jobs and subagents you will burn a LOT of money.” Automated workloads can change the subscription comparison.
  • minimaxir, Woolf’s handle — adds a practical limit to easy switching: “I noticed discrepancies when using Zed’s External Agent feature with Claude Code”. He uses Zed with Codex but the Claude app for Claude Code. The surrounding agent software can affect the comparison; this is his experience, not a controlled test.

HN handles are pseudonymous; HN publishes no per-comment scores. Ordering reflects HN’s ranking; this is a slice of the thread, not a consensus.