Claude 4 Released: What Long-Task Capability Means for Translation Workflows

On May 22, 2025, Anthropic released the Claude 4 family (Opus 4 and Sonnet 4). The new generation's most watched trait is its ability to execute complex tasks coherently over long durations — the model can work toward a goal for hours while maintaining context and instruction fidelity. For translation workflows, this capability dimension matters more profoundly than a simple "quality bump".

Real translation projects are "long tasks": a technical document of several hundred pages demands consistent terminology, unified style and correct cross-references throughout. Previous LLMs tended to "drift" on long tasks — forgetting by page 300 what was agreed on page 3. Stronger long-task capability means AI is closer to independently running the full chain of "absorb the style guide → translate → self-verify → produce a quality report", with human review focused on sampling and high-risk segments.

Translator-community testing kept its customary sobriety: long-task capability reduces drift but does not eliminate hallucination; terminology consistency improves markedly, but precision in specialist domains still needs human control. Every step of AI progress pushes human value up to a higher tier of judgment.

The model capability race keeps accelerating. For the translation industry, what matters is not memorizing every model's name but building a methodology for "rapidly evaluating new models" — because the next model always arrives next month.