Anthropic released Claude Opus 4.7 on April 16 with the same list prices as its predecessor: $5 per million input tokens, $25 per million output tokens. The rate card did not change, but the bills did.

A new tokenizer bundled with Opus 4.7 produces up to 35% more tokens for identical text, according to Finout’s analysis of the model’s pricing mechanics. Independent testing found real-world token consumption increases of 1.32x to 1.47x on coding workloads. The result: the same API call costs more, even though the per-token price is unchanged.

This is the first time Anthropic has faced significant developer backlash. A Reddit post titled “Claude Opus 4.7 is a serious regression, not an upgrade” reached 2,300 upvotes. On X, criticism of the model’s performance and cost drew thousands of likes, as Business Insider reported.

Three Mechanisms, One Outcome

The effective cost increase operates through three simultaneous changes, as detailed by BigGo Finance.

The tokenizer is the most consequential. Anthropic’s documentation acknowledges that Opus 4.7’s tokenizer can consume “up to 35% more tokens for the same text.” The company frames this as a tradeoff for improved accuracy and instruction following. For developers running production workloads, the tradeoff is straightforward: their monthly bills go up by a percentage they did not agree to.

Second, Claude Code’s default reasoning mode shifted from “high” to “xhigh” with the Opus 4.7 release. Higher reasoning effort means more tokens consumed per interaction. Combined with the denser tokenizer, the same coding session now burns through quota roughly twice as fast as it did on Opus 4.6.

Third, automatic overage billing. When a monthly subscription quota is exhausted, Rather than stopping, Claude switches to standard API rates with a daily charge cap of $2,000. One calculation cited by BigGo Finance found that a continuous Claude Code user on the Max 20x plan ($200/month) could incur $5,000 in equivalent API compute.

The March “Tokenocalypse”

The billing friction escalated in March 2026, when a Claude Code update (version 2.1.89) triggered what developers dubbed the “Tokenocalypse.” Token consumption rates for the 5-hour usage limit spiked 3x to 50x, according to BigGo Finance’s reporting. Some Max 20x subscribers exhausted their entire quota in 70 minutes. Users reported inconsistencies between the Claude Code interface (showing 100% usage) and the console (showing 73%), while bills reflected charges beyond both figures.

Performance Complaints Compound the Cost Issue

Cost alone might not have triggered a backlash this visible. The performance complaints amplified it.

Josh Pigford, founder of Baremetrics, wrote on X: “opus 4.7 is the first time i’ve thought ‘anthropic may be moving too fast’. just feels sloppy. every interaction i’m having with 4.7 across every input (cowork, chat, code, TUI, API, manage sessions)…they’re all having substantial issues that 4.6 simply doesn’t encounter.”

Newsletter writer Gergely Orosz found the model “surprisingly combative” and returned to Opus 4.6, Business Insider reported. Other users posted screenshots of the model hallucinating resume entries, failing basic reasoning tests, and refusing to code certain prompts. Boris Cherny, head of Claude Code at Anthropic, pushed back on some claims: “Adaptive thinking lets the model decide when to think, which performs better.” An Anthropic product manager separately acknowledged the team was “sprinting on tuning” and would have updates shortly, according to Digital Today.

The Lock-In Dimension

Anthropic reported a $30 billion annualized revenue run rate, per BigGo Finance’s analysis. Claude’s developer adoption has reached 43%. The backlash exposes the pricing tension that every AI vendor will eventually face: as developers build workflows, context, and muscle memory around a specific model, switching costs rise. Finout noted that enterprises planning Opus 4.7 migration should “replay real traffic side by side and measure the effective cost delta” rather than relying on the 35% ceiling as a flat estimate.

For enterprises running multi-model agent deployments, the lesson is structural: token pricing is not the only cost variable. Tokenizer density, default reasoning effort, and overage mechanics all affect the effective bill. Any enterprise cost model that tracks only the rate card is incomplete.

Y Combinator CEO Garry Tan has publicly defended Opus 4.7 as worth the cost, per Business Insider. The split opinion is the point. For some workloads, the capability gains justify the higher effective price. For others, Sonnet 4.6 at 40% cheaper per token with the older tokenizer is the better call. The difference between those two conclusions is now measured in tokenizer density, not sticker price.