Claude Opus 4.7’s Visual and Long‑Context Leap: Near‑Full Vision and 1M‑Token Tasks Redefine Knowledge Work

Anthropic officially released Claude Opus 4.7, positioning it as the strongest Claude model for complex software‑engineering and long‑chain tasks, with a focus on reduced human supervision.

Visual ability jump

The model’s visual score rose from roughly 50 % with Opus 4.6 to near‑full marks, a leap demonstrated by benchmarks such as XBOW (54.5 % → 98.5 %) and ScreenSpot‑Pro, where low‑resolution UI localization improved from 57.7 % to 69.0 % (+11.3 pts) and high‑resolution handling reached 79.5 % → 87.6 %.

Complex‑task performance

On SWE‑bench Multilingual the overall score increased from 77.8 % to 80.5 % (+2.7 pts). In the 1 M‑token GraphWalks benchmark, the Parents task rose from 71.1 % to 75.1 % (+4 pts) while the BFS task surged from 41.2 % to 58.6 % (+17.4 pts). Vending‑Bench 2 showed a profit rise from $8,018 to $10,937 (+36 %). OfficeQA Pro jumped from 57.1 % to 80.6 %, and structural‑biology reasoning climbed from 30.9 % to 74.0 % (2.4×).

Three user‑visible changes

Stronger instruction following: Opus 4.7 executes prompts more literally, reducing “prompt‑engineering” guesswork, though legacy prompts may need adjustment.

Finer visual perception: Supports images up to 2,576 px on the long side (≈3× previous size), enabling precise UI element detection, dense screenshots, complex charts, and high‑resolution posters.

Outputs closer to deliverable: Improved aesthetics and creativity for slides, documents, and code; better multi‑turn memory reduces repeated context provision, making the model act more like a finished‑product assistant.

Security and cost considerations

Anthropic added automatic detection and blocking of high‑risk network‑security requests; overall security posture is comparable to Opus 4.6, with modest gains in honesty and resistance to malicious prompts but slight regressions in a few sub‑metrics.

The new tokenizer can increase token counts by 1.0‑1.35× for the same input, raising usage costs for high‑effort workloads. Pricing remains unchanged from previous versions, though the baseline is already high.

Who benefits and cautions

Developers, analysts, legal professionals, researchers and any heavy document‑oriented users stand to gain the most, while users must watch image size (to avoid token bloat) and remain aware of the added security guardrails.