When Anthropic's head of growth Amol Avasare wrote on X, "usage per subscribed user has risen dramatically, and our current plan architecture was not designed for this," he inadvertently voiced the sharpest dilemma facing the large language model industry today — a structural mismatch between compute supply and business models. Almost simultaneously, Chinese AI firm Moonshot AI's new model Kimi K2.6 experienced malfunctions on its first day due to a surge in traffic and was forced to reset all users' quotas as compensation. These two incidents, thousands of miles apart, together outline a core question for the AI industry in spring 2026: as models advance into the "agent era," has the compute infrastructure and pricing philosophy that support them hit a breaking point?
Anthropic's "quiet test" and the reality of compute
On April 21, developers noticed Anthropic's pricing page had been quietly updated: Claude Code, previously included in the $20/month Pro plan, was marked unavailable (displaying a red "X"), while only the $100/month Max 5x plan retained the feature. The wording in the official help documentation likewise changed from "use Claude Code with your Pro or Max plan" to "use Claude Code with your Max plan" — a single‑word shift that effectively removes this popular coding agent from the $20 tier.
Facing community outcry, Avasare said this was only a "small‑scale test" affecting about 2% of new sign‑ups and that existing subscribers were not impacted. However, multiple reputable outlets pointed out the simultaneous edits to global pricing and help pages contradict the "2% rollout" claim. More importantly, Avasare later acknowledged the economic reality: when the Max plan was introduced a year ago, Claude Code was not bundled, the Cowork desktop tool did not exist, and multi‑hour asynchronous agents were not a mainstream workflow. Since the release of Opus 4, user engagement has surged and single programming sessions can consume token volumes several times the subscription fee, meaning a fixed $20 monthly charge no longer covers the actual compute costs.
This is not an isolated case. In late March, Anthropic imposed weekly limits and peak‑hour throttling because Claude Code quotas were being exhausted too quickly; on April 4 the company additionally banned Pro and Max subscribers from invoking agentic features through third‑party tools (such as OpenClaw), citing that these tools place "disproportionate strain" on infrastructure. GitHub Copilot has also recently paused registration for some flat‑rate plans. PCWorld commented that these moves point to a single conclusion: flat‑rate AI subscription models are breaking down.
Moonshot AI's "overload of enthusiasm" and its compensation
While Anthropic wrestled with compute costs, Moonshot AI faced a different compute problem in the Chinese market — an explosive surge in demand.
On April 20, Kimi K2.6 was released and open‑sourced. Thanks to top‑tier performance on PhD‑level "Humanity's Last Exam," strong results on the SWE‑Bench Pro software engineering benchmark, and support for up to 12 hours of continuous execution with 300 concurrent sub‑agents in Agent Swarms, the model drew a rush of users. The traffic spike "far exceeded expectations," causing some users to encounter membership queues and temporary feature outages; more troublesome, the backend miscounted agent quotas, resulting in some users being wrongly charged.
On April 22 at 20:30, Moonshot AI announced a compensation plan: restore all users' monthly quotas to 100% and reset already used amounts to zero. The "all‑user reset" quickly quelled dissatisfaction but revealed the extreme demands long‑horizon agents place on infrastructure. According to VentureBeat, Kimi K2.6's Agent Swarms can manage up to 300 sub‑agents running across 4,000 coordination steps simultaneously, and this scale of concurrent computation imposes unprecedented accuracy requirements on quota accounting systems.
The deeper logic of the compute crunch
The two incidents look different on the surface — Anthropic is a "supply‑side strain on fixed costs," Moonshot AI experienced a "demand‑side overload that broke accounting" — but their core is the same: there is a fundamental tension between current LLM business models and agent workloads.
Traditional subscription models assume "average user consumption is controllable," but user behavior in the agent era has fundamentally changed. A deep Claude Code programming session may run for hours and trigger thousands of model calls; Kimi K2.6's agent clusters can autonomously run for half a day or more and perform 14 rounds of code iteration and optimization. These workflows produce exponential increases in compute consumption while pricing remains fixed monthly fees or capped allowances, creating an economic "death cross."
A deeper issue is the politics of compute allocation. Anthropic chose to shift costs onto users (pushing migration to the $100 tier), while Moonshot AI absorbed the loss (resetting quotas). The former faced accusations from developers of "abandoning affordability promises," while the latter must balance user experience against financial sustainability. Neither path has a clear winner, but both signal that 2026 could be a pivotal year for reconfiguring AI subscription models.
Conclusion
With OpenAI's Codex still available in the $20 Plus plan as a competitive lever and open‑source models delivering near‑closed‑source performance at lower cost, the compute crunch at Anthropic and Moonshot AI is not an isolated case but a microcosm of industry transition. Avasare's line "our plans were not designed for this" may become a collective footnote for this generation of AI product managers. Agent capabilities have leapt forward, but the compute architectures and pricing philosophies that sustain them are still struggling to find stable ground.
▶️ Enjoy this story? Creating it is hard!
Support our YouTube channel:
AI Rosetta
Your likes, subscriptions, and shares are our greatest support! 🙏
More AI news at: https://www.airosetta.com