Anthropic dropped prices on Claude Opus 5.5 this week, a move that looked straightforward on paper. Lower costs for a flagship model should be good news for teams running complex agent workflows. Instead, the update shipped with four regressions that are quietly breaking pipelines developers have spent months tuning. The situation is a useful reminder that price changes rarely arrive alone.
What Changed, and What Broke
The pricing reduction itself is real. Opus 5.5 now costs less per million tokens, narrowing the gap with mid-tier options and making it more viable for high-volume agentic tasks. That context matters, because Anthropic has been cutting AI agent costs across its model lineup, and Opus 5.5 was one of the last holdouts at the premium tier. But the rollout also introduced breaking changes across four areas engineers depend on: tool call formatting, multi-turn memory handling, structured output reliability, and context window behavior near token limits. Each issue is individually manageable. Together, they create real friction for teams running autonomous agents in production.
Key Facts
- Opus 5.5 received a meaningful price reduction this week
- Four separate regressions were introduced alongside the pricing change
- Affected areas include tool call formatting, memory handling, structured outputs, and context window edge cases
- Developers running multi-step agent loops are most exposed
- Anthropic has acknowledged some issues but has not yet issued a full timeline for fixes
Tool call formatting is drawing the most complaints. Developers report that Opus 5.5 now occasionally returns malformed JSON in tool responses, particularly when tool schemas include nested optional fields. This is the kind of silent failure that only surfaces downstream, often after an agent has already taken an action based on corrupted data. Multi-turn memory handling has also regressed, with some users seeing context from earlier turns bleed into later responses in ways the model previously handled cleanly.
When a model gets cheaper and breaks things at the same time, you end up paying more in engineering time than you saved on tokens.Developer comment, The New Stack
The Broader Pattern
This is not the first time a cost-reduction rollout has carried unexpected side effects. Anthropic's recent changes to agent SDK billing also forced developers to revisit assumptions baked into their infrastructure. The pattern points to a structural challenge: as Anthropic moves faster to compete on price and capability, the surface area for regressions grows. Teams that have built tightly coupled agent systems are the most vulnerable, because they often have no easy fallback when a model behaves differently than it did last week.
Structured output reliability is the third failure mode getting attention. Opus 5.5 is producing valid but semantically inconsistent outputs in some constrained-format tasks, meaning the JSON parses but the values are wrong. That is harder to catch than a parse error. The fourth issue involves context window behavior: the model appears to truncate or drop earlier context more aggressively when approaching its limit, which disrupts long-horizon planning tasks that agentic systems depend on. Claude's model family has generally been strong on long-context reliability, which makes this regression more surprising to engineers who chose Opus specifically for that reason.
Anthropic has not released a patch timeline as of publication. Developers are advised to add validation layers around tool call outputs and to test structured output pipelines against the updated model before deploying to production. Whether the pricing benefits outweigh the current instability depends heavily on your use case. For teams running simple, single-turn workflows, the update is probably fine. For anyone operating multi-step agents in production, the calculus is more complicated right now.
