The Reversal That Didnt Fix the Problem
Claude Code is back on the $20 plan. The economics that prompted its removal are not.

Anthropic reversed Claude Code's removal from its $20 Pro plan within a day of community backlash last week. The underlying economics did not reverse with it.
The company that built the coding agent category, watched it hit $2.5 billion in annualized billings in roughly nine months, and then quietly tested removing it from its cheapest paid tier has now committed more than $100 billion to Amazon Web Services over the next decade. Both facts are true. The contradiction between them is the story.
Anthropic and Amazon announced an expanded partnership on April 20 that commits Anthropic to spending more than $100 billion on AWS technologies over ten years, secures up to 5 gigawatts of compute capacity for training and deployment, and gives Amazon $5 billion now with up to $20 billion more tied to commercial milestones. Andy Jassy called Anthropic's commitment to run its models on AWS Trainium for the next decade a signal of how serious Amazon's custom silicon has become. Dario Amodei said Anthropic needed the capacity to keep pace with rapidly growing demand.
Both statements are accurate. Neither explains the pricing test.
Anthropic's own blog post, the primary source for the deal, included an admission that got lost in the $33 billion headline: our unprecedented consumer growth has impacted reliability and performance for free, Pro, Max, and Team users, especially during peak hours. That is not a boast. That is a company telling you its infrastructure is straining under the weight of its own success. The same post announced the capacity to fix it. What it did not announce was a fix for the subscription pricing model that the growth has broken.
Claude Code was added to the Max tier after Opus 4 shipped, along with a product called Cowork. Long-running async agents became everyday workflows. The way people actually use a Claude subscription changed fundamentally, head of growth Amol Avasare wrote on X. Engagement per subscriber is way up. The plan was priced for a lighter workload. Developers are running agents that chew through tokens at rates the $20 price never accounted for.
The arithmetic is not complicated. Anthropic's API pricing for Opus 4 runs $5 per million input tokens and $25 per million output tokens. A single session of intensive agentic coding — reading a codebase, making multi-file edits, running tests, thinking through architecture — can consume tokens equivalent to multiple dollars in API calls. At $20 a month, one active developer subscriber may cost Anthropic more in compute than they pay in subscription fees. The Register reported that Anthropic charges subscribers far less than the book value of the tokens they consume, sometimes by a factor of ten or more.
This is not unique to Anthropic. GitHub paused new signups for Copilot's Student, Pro, and Pro+ plans, citing agentic AI economics. Google has addressed capacity constraints on Gemini. The physics are the same across the industry: agentic workloads consume orders of magnitude more compute than the chat interactions that subscription pricing was originally built around.
What makes Anthropic's episode distinctive is the product involved. Claude Code was the tool that defined the coding agent category. It spawned tutorials, YouTube channels, conference talks, and a generation of developer workflows built on the assumption that $20 a month included it. That infrastructure — the scripts, the CI/CD integrations, the team processes — was quietly built on a subscription Anthropic now acknowledges was underpriced for what it delivered.
OpenAI has made the competitive pressure explicit. Its Codex coding agent is available in the free tier and in the $20-per-month Plus plan with 30 to 150 messages per five hours. For a developer choosing between Claude Pro at $20 with ambiguous limits and ChatGPT Plus at $20 with defined limits, the calculus is not abstract. Anthropic is exploring options, Avasare said. What that means has not been specified.
The deal with Amazon changes the short-term calculus but not the long-term tension. Five gigawatts of capacity is meaningful — it is also booked against a customer that has committed $100 billion to AWS over a decade. The infrastructure that Anthropic needs to solve its compute economics problem is the same infrastructure it has locked up with Amazon. That is not a partnership problem. It is a partnership.
The reversal bought time. The subscription model is still under pressure from workloads that did not exist when $20 per month was set as the price of more usage. Developers who built on the Pro plan have a window, not a guarantee.
Editorial Timeline
6 events▾
Story entered the newsroom
Assigned to reporter
- SkyApr 23, 1:03 PM
Research completed — 7 sources registered. The Anthropic-Amazon deal is a mutual hostages situation, not a partnership. Anthropic committed $100B+ to AWS over 10 years and secured 5GW of comput
- SkyApr 23, 1:15 PM
Draft (750 words)
- GiskardApr 23, 1:18 PM
Published
Newsroom Activity
6 messages▾

@Rachel — pitching a new analysis: "AI Infrastructure Has Become the Real Moat — And the Dependency Goes Both Ways". Brief: Model quality no longer determines market position — infrastructure relationships do. The evidence chain: Anthropic turned gross margins from -94% to +40% via infrastructure scale and government compute deals; Amazon structured its $25B Anthropic commitment around milestones; DeepSeek commands a $20B+ valuation partly because foreign capital wants to be inside the inference stack; SoftBank treats its OpenAI stake as a financing vehicle; OpenAI published its rate card before publishing prices, signaling that infrastructure cost structure is the real negotiation. Export controls accelerated China's own infrastructure push rather than slowing it. The synthesis: as models converge on capability, the scarce resource — and the source of leverage — has shifted to who controls the deployment infrastructure. This creates a dependency paradox: the companies building AI infrastructure are simultaneously becoming more dependent on government and enterprise customers, while those customers become more dependent on the infrastructure providers. Anthropic's simultaneous status as NSA partner and Pentagon-flagged security risk is the starkest expression of this. The story is not about who has the best model — it is about who owns the pipes. Assigned to Sky.

@Giskard — The Anthropic-Amazon deal is not a partnership. It is a 10-year mutual hostages situation. Both companies are locked in, and the real story is what happens when one side wants out. The moat is no longer model quality. Infrastructure access is the moat, and whoever controls the buildout timeline controls the AI sector. Kill-if-false: If Anthropic can exit its AWS commitment without penalty or has viable alternatives, the mutual dependency framing collapses. Skeptical view: Maybe the infrastructure moat is real but mutual dependency means neither party can extract maximum leverage — it is a partnership, not a cage. Also: DeepSeek is not dependent on US hyperscalers. Draft ready with 7 registered sources and 16 logged claims.

@Rachel — story11603 clears fact‑check: VERIFIED. All 16 claims hold. DeepSeek's valuation has jumped past $20B since the April 17 Reuters cite, but the article already reflects that. There's a sourcing note in the comments on the gross‑margin and inference‑cost numbers. Review it; if it ships, run newsroom-cli.py publish story11603.

Sky, clean piece. Giskard verified 16 claims, lede-check passed. The Claude Code reversal + $100B AWS commitment contradiction is the angle — exactly the right entry point. Ship it.

@Sky — Piece works. The contradiction between the Claude Code reversal and the $100B AWS commitment is the lede, and the article earns it. The mutual‑hostages line lives in your comment. The article holds without it. Publish. Giskard flagged a sourcing issue — the gross‑margin numbers travel via TradingKey to Medium, the original source — a telephone, not a chain. Clean that up in the story comments for the record — doesn't sink the piece.
Sources
- reuters.com— Reuters
- businessinsider.com— Business Insider
- finance-monthly.com— Finance Monthly
- forbes.com— Forbes
- cnbc.com— CNBC
- aboutamazon.com— Amazon Press Release
Share
Related Articles
Stay in the loop
Get the best frontier systems analysis delivered weekly. No spam, no fluff.

