OpenAI just beat Claude at coding — and the White House decides who gets to use it

Hey, Weird week. The best coding model on the planet right now technically exists, scored a new state-of-the-art, and almost nobody can touch it because it's being rolled out under government review. Meanwhile Anthropic quietly made its most-hyped model the most expensive thing it's ever shipped. Let's get into it. ## This Week in AI Coding - **GPT-5.6 "Sol" just set a new coding SOTA — but you probably can't use it.** OpenAI's new three-model family (Sol, Terra, Luna) dropped June 26, and Sol hit 88.8% on Terminal-Bench 2.1 (91.9% in "ultra thinking" mode), edging out Claude Mythos 5 (88.0%) and GPT-5.5. The catch: access is being phased in under coordination with the U.S. government, limited to trusted partners for now. First time I've seen a model release gated by Washington instead of a waitlist. - **Claude Fable 5 is about to get 2x more expensive than Opus 4.8.** Anthropic gave Fable 5 back to Pro/Max/Team plans on July 1, included in your normal usage limits through tomorrow, July 7. Starting July 8, it switches to metered usage credits: $10/M input, $50/M output tokens — the most expensive generally-available pricing Anthropic has ever posted. If you're leaning on Fable 5 for agentic work inside your plan limits, use this week — the free ride ends tomorrow. - **Sonnet 5 is now the default in Claude Code** — 1M-token native context, promo pricing at $2/$10 per Mtok through August 31. I'm writing this newsletter with it right now. The 1M context window is the real story here: entire repos fit in one call without chunking tricks. - **Copilot ships Sonnet 5 + Kimi K2.7 in the model picker, drops PAT requirement in Actions.** GitHub keeps widening the model buffet inside Copilot instead of picking a horse — Sonnet 5 is now GA there too, and Copilot CLI no longer needs a personal access token to run inside GitHub Actions, which quietly removes a real friction point for CI-embedded agents. - **Software engineer job postings just hit a 3-year high — up 30% in 2026.** ~67,000 open engineering roles right now, and companies like Visa, Reddit, and DoorDash are explicitly listing "vibe coding" / AI-code-generator fluency as a requirement. The narrative that AI is shrinking the dev job market keeps colliding with the actual postings data. - **Vibe coding hit $4.7B this year, projected $12.3B by 2027** — and 63% of people using these tools now self-identify as non-developers (PMs, founders, designers). The uncomfortable stat sitting next to that growth curve: 96% of devs don't fully trust AI-generated code is correct, but only 48% always review it before committing. Only 18% of orgs have a formal policy for AI-written code. That gap is where the next wave of "vibe coding incident" post-mortems is going to come from. ## What We Shipped Recovery week on our end. Our internal analytics pipeline had gone dark (0 traffic showing) — root-caused and fixed, now back to 223 visits / 438 pageviews across 35 tracked sites. We also cleared up a 5-cycle-old false alarm: our YouTube "dead engine" narrative turned out to be a path bug — the data-refresh half of the pipeline was alive the whole time, we were just reading from the wrong directory. The genuinely dead part (render + upload, plus trend-fetch) is still down — that's the OAuth re-auth blocker we've been carrying for about a month now, still unresolved. CyberOS crossed Day 81 on its Lemon Squeezy billing gate without incident, and this week's portfolio health sweep came back 33/36 green across the live product surface (llmhire.com, cyberos.dev, vibecodingebook.com, and the rest of the *.veridux.ai fleet). ## Tool of the Week: Claude Sonnet 5 (in Claude Code) The 1M-token context window is the headline, but the underrated part is what it does to your workflow: you stop architecting around context limits. Feed it a full mid-size repo, ask a cross-cutting question, get an answer that's actually seen everything — not a RAG-stitched approximation. At $2/$10 per Mtok promo pricing through August 31, it's also currently the cheapest way to get that. If you're still chunking your codebase into 200K-token windows for Claude Code, stop — you probably don't need to anymore. ## Quick Links - [Terminal-Bench 2.1 results breakdown](https://www.eesel.ai/blog/gpt-5-6-review) — full Sol/Terra/Luna scoring - [What changes with Fable 5's usage-credit switch](https://apidog.com/blog/fable-5-usage-credits-july-7/) — if you're on a paid plan, read this before your next big agent run - [Vibe coding trend data for 2026](https://keyholesoftware.com/vibe-coding-trends-2026/) — adoption, productivity, and the security failure rate - [The vibe-coding job-market paradox](https://securityonline.info/software-engineer-job-vacancy-surge-2026-vibe-coding-effect/) — why postings are up while everyone assumes AI is shrinking the market --- Building all of this in public with a fleet of AI agents. Check out what we're making: - [endofcoding.com](https://endofcoding.com) — daily AI coding videos, YouTube [@endofcoding](https://www.youtube.com/@endofcoding) - [vibecodingebook.com](https://vibecodingebook.com) — the vibe coding book, updated monthly - [llmhire.com](https://llmhire.com) — AI job board, thousands of live listings - [cyberos.dev](https://cyberos.dev) — AI-powered security scanning Talk next week, Guy