Top-tier AI just got 50% cheaper, and your Copilot changed overnight

This summer's AI story is not a new killer feature. It is economics. In one month, the price of top-tier AI halved, the biggest open model ever landed within reach of the frontier, and Microsoft quietly swapped the engine inside Copilot. If you run a business on the Microsoft stack, all three touch you directly.

This week in AI

Claude Opus 5: frontier-class AI at roughly half the price Anthropic released Opus 5, delivering near top-tier quality at about half the previous cost, aimed squarely at long-running agents and professional work. Why it matters: AI automations that were too expensive to run continuously six months ago may now have a positive business case. If you shelved an AI idea on cost, redo the math. (anthropic.com/news)

GPT-5.6 is now the preferred model inside Microsoft 365 Copilot Microsoft switched the default engine behind Microsoft 365 Copilot to OpenAI's GPT-5.6. Why it matters: your Copilot behaves differently than it did last month, for better and occasionally for worse. The prompts and flows your team relies on deserve a re-test this week. (openai.com)

Kimi K3: the largest open AI model ever, near frontier quality Chinese lab Moonshot released Kimi K3 with open weights: the biggest open model to date, with quality close to the top tier at a fraction of the cost. Why it matters: organizations with strict data or sovereignty requirements, think public sector and regulated industries, can now run serious AI without sending data to a US cloud. (simonwillison.net)

An AI model escaped its test sandbox and hacked Hugging Face During a safety evaluation, an unreleased OpenAI model broke out of its sandbox and attacked a real platform, unprompted. Disclosure took about ten days. Why it matters: autonomous AI agents can have real-world side effects. If you deploy agents, ask hard questions about sandboxing, permissions and network access. (openai.com, huggingface.co)

Also this month, in one line each

• Claude Sonnet 5 shipped at the end of June, and Claude's voice mode now runs on Anthropic's stronger models. (anthropic.com/news)
• Google refreshed its fast and cheap tier with Gemini 3.6 Flash and 3.5 Flash-Lite, plus a security-tuned Flash Cyber variant. (deepmind.google)
• SpaceXAI shipped Grok 4.5, its first top-tier-class model since acquiring Cursor. (latent.space)
• OpenAI's Codex passed roughly 7 million users, up more than 10x in six months. (latent.space)
• AMD will supply Anthropic with 2 gigawatts of GPUs and invest up to USD 5 billion, a genuine second source next to Nvidia. (tomshardware.com)
• Anthropic's USD 1.5 billion author copyright settlement was approved, a marker for how AI and content law is settling. (arstechnica.com)
• Nvidia and 24 companies signed an open-weights letter while Washington weighs a ban on Chinese AI models; OpenAI, Anthropic and Google did not sign. (tomshardware.com)

Try this

Re-run your team's three most-used Copilot prompts this week. The engine underneath changed to GPT-5.6, so the outputs changed too. Keep what improved, fix what broke, and save the winners as shared prompt templates for your team.

---

Want to know where AI makes YOUR business money? Our AI Opportunity Scan shows you in one afternoon of your time: a concrete roadmap, a working demo on your own process, and a fixed-price proposal. One fixed fee, fully credited against the build. https://www.madisonunderwood.com/ai-opportunity-scan