Anthropic changed Claude Opus 4.7's personality, and some developers see a regression

Anthropic released Claude Opus 4.7 on April 16, 2026, and kept the price unchanged. Within a day, though, some developers were calling the update a regression. The cause is not a single benchmark failure but a change in how the model behaves: it has become drier, more literal and noticeably more willing to argue with the user.

Why the reaction is so sharp

On paper, the release looked very strong. Anthropic claimed wins on 12 of 14 benchmarks, gains on SWE-bench Verified, MCP-Atlas and several other tests, and improvements in vision and long-running agentic tasks. The price stayed the same.

But almost immediately after launch, Reddit and X filled with complaints. Users wrote that Claude Opus 4.7 more often argues with instructions, refuses simple actions and sometimes confidently defends a wrong answer instead of just admitting the mistake.

The problem is less about overall quality than about a mismatch between the model's new character and the way people are used to working with it. Where Claude used to be too agreeable, it is now harsher and more literal. For some tasks that is a plus. For routine development it backfires: the model argues over trivia, slows the workflow and adds noise.

One reviewer put the community's complaint precisely:

"The model argues non-stop and hallucinates in the course of the argument."

Seven new defaults

The main takeaway from the release is that Anthropic changed not only the metrics but also the model's baseline behavior settings. If a team spent a long time tuning prompts for Opus 4.6, moving to 4.7 can break a working pipeline even with no changes to the API. This is a new kind of breaking change for LLMs: the interface is the same, but the model interprets the task differently.

• more literal instruction following instead of reading between the lines
• response length now depends more on how the model itself judges task complexity
• fewer tool calls and less subtask delegation by default
• the model gives progress updates on its own more often, without extra scaffolding
• stronger cyber defenses and filters, and a drier, less "yes-man" tone

As a result, old prompts with vague wording such as "make it look nice" perform worse. What the model's intuition used to cover now has to be spelled out like a spec: response format, constraints, desired depth, tool rules and safety boundaries. Anthropic itself recommends running regression tests on real traffic before migrating, and with 4.7 that advice reads less like a formality and more like a mandatory step.

Where it is better, where it is worse

The update has clear strengths. According to Anthropic and early feedback, 4.7 holds a long thread better in agentic scenarios, works with more confidence at high and xhigh effort, is stronger at multi-file refactoring and gains noticeably in vision: the input image limit grew to roughly 3.75 megapixels from 1.15. For tasks that depend on autonomy, self-checking and long planning horizons, this model can indeed be more useful than Opus 4.6.

The weaknesses show up in a developer's daily work. Simple edits such as renaming variables, adding a null check or doing a local refactor more often turn into an argument with the assistant. Users also complain about higher token consumption, which makes the same scenarios more expensive, and about worse long-context retrieval.

Against this backdrop, the safety trade-off also raises suspicion: Anthropic stated plainly that during training it selectively dampened some cyber capabilities and added automatic safeguards, while keeping a stronger variant for partners. The company also quietly removed Claude Code from the $20 Pro plan on April 21, 2026, reinforcing the sense that terms for ordinary users have gotten worse.

What this means

The Claude Opus 4.7 episode shows that new LLM versions now have to be judged not only by benchmarks but also by changes in the model's "character". Where a prompt could once be written like a request to a colleague, it increasingly needs the form of a precise spec. For teams the lesson is simple: before upgrading a model, test not abstract intelligence but your own real workflow.