Claude Opus 4.7 Revamps System Prompt Compared to 4.6
anthropic claude
| Source: Mastodon | Original article
Anthropic rolled out Claude Opus 4.7 on April 16, 2026, and, for the first time since July 2024, the company published the full system prompt that drives the model’s behaviour. The newly released prompt differs markedly from the one used in Opus 4.6, tightening instruction compliance, swapping in a revised tokenizer, and reshaping how the model handles tool use, long‑running workflows and “agentic” coding tasks.
The changes matter because the system prompt is the hidden rulebook that determines how Claude interprets user requests, prioritises safety, and allocates compute. By making the prompt public, Anthropic offers developers a rare glimpse into the levers that steer model performance, a transparency move unmatched by other major labs. The stricter instruction set reduces “hallucination” on complex software‑engineering queries, a claim backed by Anthropic’s own benchmarks that show Opus 4.7 outperforming 4.6 on the toughest coding challenges. The new tokenizer also alters token accounting, meaning existing API calls may see different cost calculations and token limits.
Beyond the prompt, Opus 4.7 adds high‑resolution image handling up to 3.75 MP and introduces an “xhigh” effort tier that allocates extra compute for demanding tasks. These upgrades broaden Claude’s appeal for visual‑heavy workflows and for enterprises that need deeper reasoning without sacrificing speed.
What to watch next is how the community reacts to the disclosed prompt. Early adopters are likely to experiment with prompt‑engineering hacks, while competitors may feel pressure to follow Anthropic’s transparency playbook. Analysts will also monitor whether the new tokenizer reshapes pricing models and whether the stricter instruction regime impacts the model’s flexibility in creative domains. The next model update, slated for later this year, will reveal whether Anthropic can sustain the performance gains while keeping the prompt open for scrutiny.
Sources
Back to AIPULSEN