Anthropic’s Claude Opus 4.7 faces backlash over hidden cost concerns

08:20

By: Dakir Madiha

Anthropic’s Claude Opus 4.7 faces backlash over hidden cost concerns

Anthropic released Claude Opus 4.7 as its most advanced public AI model, highlighting gains in coding, instruction following, and long running agent tasks. The San Francisco company also reported higher honesty scores, with the model reaching 91.7 percent on its MASK benchmark and challenging incorrect assumptions in more than three quarters of cases. Pricing remains unchanged on paper at five dollars per million input tokens and twenty five dollars per million output tokens.

Despite these claims, experienced users quickly raised concerns about real world costs. The update introduces a new tokenizer that can map identical text to up to 35 percent more tokens. Developers argue that this change increases effective usage costs without altering official pricing.

A second issue comes from default reasoning settings. The model now runs at an “xhigh” effort level in coding environments, which generates more internal reasoning tokens during complex tasks. Users on developer forums reported hitting usage limits far faster than before. Some also flagged weaker instruction adherence and increased hallucinations compared to the previous version.

The pricing shift is more visible in integrated tools. Within GitHub Copilot, Opus 4.7 launched with a premium request multiplier of 7.5 times, up from three times for Opus 4.6. This means a single interaction can consume a larger share of monthly quotas, at least during the current promotional period.

Anthropic acknowledged the tokenizer change, stating that the same content may now require between one and 1.35 times more tokens depending on format. The company advised users to manage costs by adjusting effort levels, setting task budgets, or refining prompts. Engineers also noted that deeper reasoning tokens often correlate with more accurate outputs, suggesting the added computation serves a functional purpose.

Some early complaints softened after initial deployment issues were addressed. Users reported improvements within a day, aligning with Anthropic’s claim that rollout bugs had been fixed. Still, the episode has intensified concerns among developers about reliance on AI services that can change behavior without clear cost signals.