Anthropic Revises Claude Opus System Prompt Between 4.6 and 4.7

Anthropic published the updated system prompt and behavior notes that differentiate claude-opus-4-7 from claude-opus-4-6. The release pairs core capability upgrades with explicit changes to the system prompt, a new tokenizer, and token-usage-affecting features that make existing prompts and agent flows behave differently. Key capability headlines include 13% reported coding benchmark lifts, high-resolution vision at 2576px / 3.75MP, and improved low-level perception and image localization.

The system-prompt changes are coupled with platform and runtime changes you need to know. claude-opus-4-7 introduces a new tokenizer that changes how inputs are tokenized and therefore alters token counts for the same text and images. The model adds a new xhigh effort level to tune capability vs cost, and a beta task budget mechanism that provides the model a running token countdown when executing multi-step agentic loops. High-resolution image support raises the pixel cap to 2576px / 3.75MP, and image coordinates are now 1:1 with pixels, eliminating scale-factor math.

Practitioners should audit three classes of integration changes:

• •Prompt behavior: stricter instruction following means prompts that relied on implicit leniency from Opus 4.6 may fail or be truncated.
• •Token economics: the new tokenizer and higher-resolution images increase token usage; downsample images when extra fidelity is not required and budget task budgets conservatively.
• •Agent orchestration: task budget introduces an in-band budget signal the model will use to prioritize work, and xhigh can be used for more capability-sensitive agent steps.

Anthropic is unique among major labs in publishing system prompts for user-facing chat systems, and making prompt changes explicit reduces uncertainty for teams that run production agents or integrate Claude across platforms like Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. The combination of stricter instruction adherence and task-budgeting signals a design choice: make the model more deterministic and cost-aware for long-horizon agentic tasks, even if that breaks some previously implicit behaviors. This aligns with industry trends toward predictable, instrumentable models when used in agent loops and safety-sensitive automation.

The upgrade is meaningful for coding-heavy and vision-heavy workloads. Benchmarks cited by third-party analyses show Opus 4.7 pulling ahead in coding and visual navigation metrics, which raises the bar for competitive comparisons with other frontier models. However, these gains come with migration costs: prompt rewrites, retuning of effort levels, and possible tool-calls redesign to respect task budgets and new token accounting.

Run parallel validation: deploy Opus 4.7 in shadow mode against Opus 4.6 for the most common agentic flows and prompt templates, monitoring both functional outputs and token usage. If you rely on low-friction prompt behavior, expect to harden instructions and add explicit verification steps. Finally, treat task budgets as part of your orchestration API: they change model behavior and can be used to enforce graceful degradation in long-running interactions.