Claude 4.7 Leads Coding Benchmarks but Burns More Tokens
Claude Opus 4.7 achieves state-of-the-art on SWE-Bench Verified and Pro via precise instruction following and output verification, excelling in agentic coding and UI generation, but uses significantly more tokens per task (shifting reasoning tiers up), increasing effective costs despite unchanged $5/$25 per million pricing.
Agentic Coding and Reasoning Upgrades Enable Production Workloads
Claude Opus 4.7 handles long-running tasks with higher rigor by following instructions literally, verifying outputs before reporting, and breaking complex requests into modular systems like physics engines, rendering, and cameras. This results in state-of-the-art SWE-Bench Pro and Verified scores, outperforming Claude 4.6, GPT-4o, and Gemini 1.5 Pro on hardest tasks, especially webdev UI generation now matching Gemini. Reasoning efficiency improves across tiers—low performs like prior medium, medium like high, high like max—allowing handoff of engineering work with minimal supervision. Memory gains support multi-session workflows, while vision processes images at 3x higher resolution for polished UI designs, slides, and documents. Retune prompts from 4.6 as literal interpretation breaks older setups.
Token Efficiency Trade-offs Raise Real Costs
Higher capability demands more reasoning tokens, using 2-3x more per task than 4.6 (visible in bar graphs shifting usage up), reducing usable context despite same pricing: $5 per million input tokens, $25 per million output. Early tests show weaker long-context retention, and max reasoning hits rate limits quickly (Anthropic raised limits post-launch). This trades quality for efficiency, making it pricier for high-volume use.
Demo Outcomes Highlight UI Strengths Over Complex Sims
In Kilo CLI tests (open-source agent harness), Opus 4.7 generated the best 3D SUV physics sim in mountains, modularizing logic for long-horizon planning; top Minecraft clone with ores, mobs, water physics, and procedural terrain (buggy execution); accurate MacOS UI clone with functional Finder, Launchpad, Spotlight, apps like Safari/Notes/Calculator (janky toolbar). Frontend landing pages match Gemini 1.5 Pro quality with dynamic typography and consistent styling. SVG outputs mixed: strong animated butterfly/painting with ambient effects, but weaker PS5 controller vs. Qwen 3.5B or Gemini. FPS shooter had recoil/movement but glitched controls. Overall, excels in ambitious creative UIs over flawless sim execution.