GPT-5.6 vs Claude Opus, Claude Code's token bloat, and other tools worth your week

Subject: GPT-5.6 just beat Claude Opus — and Claude Code ships 33k tokens before your prompt

A/B alternatives: - "We migrated an AI agent to GPT-5.6. Here's the actual bill." - "Claude Code: 33k tokens before your prompt. OpenCode: 7k."

TL;DR: OpenAI's GPT-5.6 family shipped this week and at least one production agent team (Ploy) is publicly migrating off Claude Opus after four months of nothing beating it — 2.2× faster, 27% cheaper, and cleaner output. Separately, Systima measured Claude Code's harness overhead and found it sends roughly 33,000 tokens of scaffolding per turn versus OpenCode's ~7,000. Plus a no-frills screenshot CLI, Hugging Face's case for owning your models, and the AI family room.

1. GPT-5.6 Sol shipped — and Ploy just migrated its production agent off Claude Opus

GPT-5.6 is OpenAI's new flagship model family released 9 July 2026, with three tiers: Luna, Terra, and Sol. Ploy, an AI marketing-site builder, ran GPT-5.6 Sol head-to-head against Claude Opus 4.8 across its eval suite (hundreds of real fixture workspaces, scored by a visual judge) and made it the new default. Headline numbers, per the Ploy blog (11 July 2026): mean per build cost dropped from $3.06 (Opus) to $2.22 (GPT-5.6), wall-clock time fell from 8m 00s to 3m 42s, output tokens dropped by half, and visual score rose from 0.936 to 0.970. On one matched task Opus produced a 17,957-character globals.css with 174 variables where GPT-5.6 wrote 2,508 characters and 45 variables for a comparable page. The catch: Ploy found three production gotchas — tool schemas, prompt-cache topology, and OpenAI's store:false requirement to make reasoning replay self-contained — that any team migrating needs to know. Honest take: real production data from a paying team running real evals is rare; treat this as signal, not proof. But "2.2× faster and 27% cheaper on a competitive agent task" is hard to wave off. Source: https://ploy.ai/blog/migrating-a-production-ai-agent-to-gpt-5-6

2. Claude Code burns ~33,000 tokens before your prompt; OpenCode burns ~7,000

Systima put Claude Code and OpenCode on the same model, machine, and tasks and read every byte sent. The result, published 12 July 2026: a single "say OK" turn ships ~33k tokens of system prompt, tool schemas, and scaffolding under Claude Code versus ~7k under OpenCode — a ~4.7× fixed overhead before your prompt even arrives. The bigger bill-driver comes next: Claude Code rewrote tens of thousands of prompt-cache tokens mid-session, while OpenCode's prefix was byte-identical across runs, so it paid for the cache once and read it back cheaply. Worst case on the same task: Claude Code wrote 54× more cache tokens than OpenCode. Stack a CLAUDE.md, MCP servers, framework templates, and subagents and the gap compounds. Honest take: if your Claude Code dashboards are climbing faster than your usage should explain, this is probably why. Worth reading if you pay Anthropic bills. Source: https://systima.ai/blog/claude-code-vs-opencode-token-overhead

3. shot-scraper 1.11 — the screenshot/JS CLI gets a --js-file flag and a 30s server wait

shot-scraper is Simon Willison's CLI utility for taking website screenshots, recording video demos, and scraping sites with JavaScript. Version 1.11 (12 July 2026) ships three practical wins: a new --js-file flag on shot-scraper, pdf, html, accessibility, and har commands so you can load JavaScript from a local file, stdin, or gh:user/script instead of pasting it as a string argument; a matching js_file: YAML key for shot-scraper multi; and the multi/video server: mechanism now polls up to 30 seconds for the target URL to accept connections instead of the old fixed 1-second wait (which broke slow-starting local servers). Honest take: boring release notes are the best kind. If you script screenshots or scrape JS-heavy sites, this just removes two small papercuts. Source: https://simonwillison.net/2026/Jul/12/shot-scraper/

4. Hugging Face's CEO: companies are done renting their AI

Hugging Face CEO Clem Delangue sat down with TechCrunch (10 July 2026) to make the case that open-weight models are no longer the "cheap alternative" — they're the strategic choice. The framing: as inference costs collapse and self-hosting gets easier, more companies are pulling workloads off hosted APIs to control their data, latency, and unit economics. He pushed back on the "renting vs owning" framing the industry has leaned on, arguing the gap has closed enough that for many workloads, owning wins. Honest take: directional, not a benchmark. Useful lens if you're pricing a build-vs-buy decision this quarter; less useful if you're a one-person team who just needs an API to ship. Source: https://techcrunch.com/2026/07/10/hugging-faces-ceo-on-why-companies-are-done-renting-their-ai/

5. ChatGPT goes after the family room

OpenAI is pushing ChatGPT deeper into households with a new "families" experience (TechCrunch, 11 July 2026). The pitch: shared accounts, kid profiles with guardrails, and household-level controls — moving ChatGPT from a single-user tool to something parents and kids sit down with together. Honest take: the consumer-AI land grab is moving from "your laptop" to "the kitchen counter." Worth a glance if you have kids or build for parents; not a tool update if you live in the API tier. Source: https://techcrunch.com/2026/07/11/openai-bets-on-families-as-chatgpt-goes-deeper-into-households/

If this was useful, forward it to the one friend who keeps asking "what AI tool should I actually try?" — it's the single best thing you can do for Next Tool.

— Next Tool · What's next in AI productivity.