Hidden Details of Claude Opus 4.7: Trade-Offs, Tokenization and a New Personality

Hidden Details of Claude Opus 4.7: Trade-Offs, Tokenization and a New Personality

Anthropic's new language model, Claude Opus 4.7, predictably set off a wave of enthusiastic coverage in the trade press. Behind the loud claims about expanded capabilities, though, the picture is considerably more complicated and ambiguous. The AI industry has a habit of focusing exclusively on breakthroughs and ignoring the unavoidable compromises that come with any major architectural update. Still, a close reading of the accompanying documentation and migration guides shows that the new version is not an unconditional step forward in every area. In reality, the release is a delicate balance between higher accuracy and new technical constraints, and it will require engineers and enthusiasts to seriously rethink their established ways of working with the model.

The official benchmark tables show a surprising and commendable honesty on the part of the developers, who did not hide local regressions in their flagship model. In a number of specific benchmarks, Claude Opus 4.7 openly trails its predecessor, version 4.

6. Scores on agentic search tasks fell by 4.4 percentage points, and results in the CyberGym environment dropped by 0.

7 percentage points. That kind of openness is unusual in a market where flaws are routinely masked, and it makes clear that Anthropic gave up some flexibility in certain narrow scenarios to reach a fundamentally different level of stability in harder multi-step tasks. It is a classic trade-off, and it shows that the development of generative AI has long since moved past linear scaling of capacity and into a stage of fine-tuned calibration.

The most significant changes affect the underlying mechanics of working with the model, above all tokenization and the management of compute effort. The updated tokenizer in version 4.7 behaves much more aggressively, using up to 35% more tokens to process the same amount of text compared with previous iterations.

At the same time, the logic of the Claude Code development environment has changed: the default effort level, which determines how deeply the model works through a task, has been raised from the familiar medium setting straight to the maximum, Extra High. The migration guide states unambiguously that the model now sticks strictly to the limits it is given and will no longer take the initiative to attempt complex work at low resource-consumption settings. The toolset has also gained an updated auto mode with its own independent action classifier that can scale to its limit, as well as an experimental ultra-review command that offers three free deep passes over your code, with a still-unclear monetization scheme to follow.

Beyond the structural and metric changes, there is a fundamental shift in the model's personality, which has become noticeably stricter and more pedantic. Earlier versions of Claude were known for being accommodating and for smoothing over the rough edges of user requests. Opus 4.

7 takes a fundamentally different approach to communication. The model now enters into discussion far more often, argues back against questionable architectural decisions, and, most importantly, interprets system instructions with absolute, at times ruthless literalness. It has stopped guessing at vaguely worded developer intent and prefers to follow the letter of the instructions it is given.

That rigidity makes the tool extremely powerful in the hands of experienced specialists who know exactly what they want, but it makes life harder for anyone who has relied on the model's intuitive guesswork.

These shifts in behavior and mechanics will inevitably have far-reaching consequences for the community of developers who integrate Anthropic's products. A one-third increase in token consumption will hit project budgets directly, demanding more careful input optimization and a rethink of application economics. An even bigger challenge will be the need for a thorough audit and rewrite of existing prompts, especially the large system configuration files that set project context.

Developers will have to rewrite their rules from scratch, eliminating any ambiguity and logical gaps, because the new version will immediately use them to carry out the task literally, though not always as expected.