Anthropic launches Opus 4.7 after weeks of user criticism

The complaints that couldn't be ignored

The most substantiated complaint came from a Senior Director in AMD's AI group, who reviewed thousands of Claude Code sessions and concluded that the model could no longer reliably handle advanced engineering tasks. The analysis showed that the depth of reasoning had decreased by 67 percent since February. The model also read fewer files before editing code, and the number of aborted tasks had increased significantly.

Complaints quickly spread on Reddit and X. The term "AI shrinkflation" emerged as users felt they were paying the same price for a product that was suddenly inferior. Speculation arose that Anthropic had deliberately downgraded the model to save computational power, a claim the company has denied.

What actually happened?

Anthropic denied that the model weights had changed, and that was probably true. But the explanation was more complex. During February and March, the company had adjusted the default settings for how deeply the model reasoned per task. This meant shorter planning sequences and fewer files reviewed before code was changed. For simple queries, this was barely noticeable, but for complex, long-term processes, the effect was clear.

The changes were documented in update logs, but without clear communication about what they would mean in production environments. Customers were not given the chance to understand and adapt in time.

What's new in Opus 4.7?

Opus 4.7 addresses several of the issues users raised during the spring. The model is better at staying focused in long, complex tasks and gives up less often. It also checks its own results more thoroughly before reporting back, reducing the need for human review at every step.

Three other changes are worth knowing about:

1. 1

   Image processing has been significantly improved and can now handle images with much higher resolution, which opens up for more advanced computer usage and analysis of complex diagrams.

2. 2

   The model also has improved file-based memory, meaning it can carry key information between sessions without needing everything explained from scratch each time.

3. 3

   A new effort level has been introduced, giving more control over the trade-off between reasoning depth and response time on hard problems.

A practical warning for those migrating: Opus 4.7 uses an updated tokenizer that may produce slightly higher token usage for the same input. This affects cost and is worth measuring on actual traffic before switching.

Anthropic openly admits the limitations

Anthropic communicated differently than what is the norm for a model release. They openly admitted that Opus 4.7 does not match their most advanced system, Claude Mythos Preview.

Mythos is Anthropic's most powerful model, but it is not yet available to the public due to safety concerns about its advanced cyber capabilities. Instead, it is being selectively tested by a small number of chosen tech and security companies.

Anthropic also presented performance comparisons against competitors and clearly warned that existing setups may need adjustment after migration. Customers got the basis to actually make an informed decision, a clear shift from how the Opus 4.6 changes were handled.

What should you expect from your LLM provider?

This is not a problem unique to Anthropic. Models are launched, optimized and adjusted continuously by all major providers, and those who notice changes first are usually the ones using the tools intensively in complex workflows.

If you use AI in your processes today, it's worth setting three concrete requirements for your provider:

1. 1

   Clear communication when default settings change in a way that affects performance in production environments.

2. 2

   Sufficient advance notice so you have time to adapt your setups before the change takes effect.

3. 3

   Your own benchmarks: keep test cases representing your most important use cases and run them against new model versions before migrating, so you spot changes before they affect your results.

How Anthropic handled the trust loss

How Anthropic communicated around the Opus 4.7 launch differed from what preceded it. In an industry where launches normally focus on communicating strengths and staying silent about limitations, the openness around Mythos Preview and the migration risks was unusual.

While this action might not be enough to regain the trust lost this spring, it demonstrates that honesty can be a competitive advantage. For companies choosing an AI provider today, it is a factor worth weighing as heavily as performance and price.