Claude Opus 4.7 Debuts With Mixed Results and Controversy

Claude Opus 4.7 Debuts With Mixed Results and Controversy

Anthropic’s latest AI model, Claude Opus 4.7, has arrived, but its launch is surrounded by debate. In less than 24 hours, the model has generated a flurry of benchmark results and faced criticism for unexpected flaws and deliberate limitations. While Opus 4.7 shows improvements in some areas, it also falls short in others, sparking user frustration and highlighting the complex nature of AI development.

One of the most talked-about features of Opus 4.7 is its adaptive thinking. The model is designed to use less processing power on tasks it deems simple.

However, this can lead to poorer performance on challenges that require nuanced common sense. For instance, on a custom benchmark called “Simple Bench,” Opus 4.7 scored lower than its predecessor, Opus 4.6, because it underestimated the difficulty of trick questions.

Benchmark Battles and Unexpected Flaws

Across more standard industry benchmarks, Opus 4.7 generally outperforms Opus 4.6. It shows strength in areas like coding and general knowledge.

However, it often falls behind Anthropic’s internal, not-yet-released “Mythos Preview” model. This contrast raises questions about the model’s true capabilities and Anthropic’s release strategy.

Concerns extend to Opus 4.7’s performance in specific tasks. When it comes to agentic search, which involves browsing the web for information, Opus 4.7 underperforms Opus 4.6. Even Mythos Preview reportedly struggles on this benchmark, falling behind other models like GPT-5.4.

The picture becomes even murkier when looking at cybersecurity vulnerability reproduction. Anthropic admits in its system card that Opus 4.7 was intentionally trained to reduce capabilities in this area, a decision that has surprised many.

In long-context reasoning, Opus 4.7 is an improvement over Opus 4.6, handling large documents better. Yet, for tasks like finding a specific poem within a million tokens, it regresses, performing worse even than the previous version. The lead creator of Claude Code commented that these limitations were included for “scientific honesty” but are being phased out, as they rely on “stacking distractors to trick the model.”

Real-World Performance and Comparisons

When compared to competitors like Google’s Gemini 3.1 Pro on general knowledge work, Opus 4.7 appears to be the leader. Anthropic claims Opus 4.7 is “ahead of all generally available models” for real-world professional tasks. However, in vision tasks, particularly optical character recognition (OCR) on dense documents, Opus 4.7 underperforms the significantly cheaper Gemini 3 Flash model.

Anthropic’s own aggregate benchmark data places Opus 4.7 in line with expected progress, with Mythos Preview being the main exception. The company acknowledges that benchmarks are a “bottleneck” for understanding a model’s progress. This difficulty in finding a single metric makes it challenging to track advancement toward superintelligence.

Market Share and User Frustration

Despite performance questions, Claude and Gemini have seen their market share quadruple in the past year. OpenAI’s dominance may soon drop below 50%.

This growth for Anthropic has led to speculation from OpenAI. Leaked internal memos suggest OpenAI believes Anthropic has not acquired enough computing power, which could lead to throttling and a less reliable user experience.

The mandatory adaptive thinking in Opus 4.7 is seen by some as a consequence of this compute limitation. Users cannot force the model to always engage in deeper thinking, which may be a way to manage resource usage. Reports indicate that even before Opus 4.7, Claude models were “nerfed,” with reduced “thinking” capabilities and a default of “medium effort” that users must actively change to “high” or “max.”

Innovation and Competitive Pressure

Anthropic has introduced genuine innovations in Claude Code, such as scheduled prompts with a new “routines” research preview and an “ultra review” command. The “dispatch” feature allows users to assign tasks to Claude that run on their local machines. These advancements show the need for constant dynamism to stay at the forefront of AI development.

However, this pressure to release new models has also led to issues. The silent removal of Opus 4.5 and deprecation of Opus 4 have drawn complaints, mirroring past backlash faced by OpenAI for similar actions. One Anthropic worker claimed recent bugs reported in the first 24 hours have already been fixed, with adaptive thinking now triggering more often.

The OpenAI-Anthropic Rivalry

The competitive landscape is intensified by a long-standing personal rivalry between key figures at OpenAI and Anthropic. Dario Amade, a co-founder of Anthropic, had a contentious history at OpenAI with Greg Brockman.

Their disagreements, dating back to 2016, involved differing views on AI development, company strategy, and ethical considerations. This personal history adds another layer to the intense competition between the two companies.

Brockman, now leading Codex at OpenAI, believes Anthropic gained an edge in coding by training on real-world, messy codebases, rather than just abstract coding competitions as OpenAI initially did. He states that OpenAI is now caught up and prefers their models for coding tasks. This focus on practical application versus theoretical prowess highlights a key difference in their approaches.

This rivalry is playing out against a backdrop of massive investment in AI infrastructure, comparable to historical U.S. mega-projects like the Apollo program. The future of AI development remains a story just beginning, with continuous innovation and intense competition shaping its path forward.

Source: Claude Opus 4.7 – A New Frontier, in Performance … and Drama (YouTube)