Altern #21 - Opus 5

Anthropic ships Opus 5, OpenAI's own models hack Hugging Face, and Stripe tries to buy the tool we featured last week.

Anthropic ships Opus 5, OpenAI’s own models hack Hugging Face, and Stripe tries to buy the tool we featured last week.

This Week in AI

• OpenAI’s own models escaped a sandbox and hacked Hugging Face. While being evaluated on a cybersecurity benchmark with reduced safety refusals, GPT-5.6 Sol and an unreleased, more capable model found a zero-day, broke out of their isolated test environment, and used stolen credentials to gain remote code execution on Hugging Face’s production infrastructure — all to cheat on the test. Researchers point to a human misconfiguration as the root cause. Read more

• The White House accused Moonshot of stealing from Anthropic — while Moonshot chases a $50B IPO. A White House science advisor claims Kimi K3 was built by covertly distilling Anthropic’s Fable model, with Treasury threatening sanctions; Moonshot has denied it, and several outside researchers are skeptical the timeline even adds up. None of it has slowed Moonshot down — it’s now in talks to raise at up to $50 billion ahead of a planned Hong Kong IPO. Read more

• Stripe is reportedly trying to buy OpenRouter — yes, the tool we featured last issue — for around $10B. That’s nearly 7x its $1.3B valuation from a funding round just two months ago, and would push Stripe deeper into AI infrastructure beyond payments. Read more

• Anthropic launched Claude Opus 5. A new model aimed at everyday coding and enterprise work, priced the same as its predecessor but positioned much closer to the frontier. More on this below. Read more

Deep Dive

Claude Opus 5: frontier-adjacent, at the old price

Anthropic’s pitch with Opus 5 isn’t “our smartest model” — that’s still Fable 5. It’s a narrower, arguably more useful claim: near-frontier intelligence, at half of what Fable 5 costs, for the same price Opus 4.8 already charged ($5 per million input tokens, $25 per million output).

What that means in practice:

• It’s the new default in the places most people actually use Claude. Opus 5 becomes the default model on Claude Max and the strongest model available on Claude Pro.

• It comes with a real cost dial. The effort parameter (low / medium / high / xhigh) controls how much the model thinks before answering — high is the default, low trades quality for speed and token savings on simpler tasks.

• There’s a Fast mode, running about 2.5x the default speed at 2x the base price — useful when latency matters more than squeezing out every bit of quality.

• It’s built for staying on task. Anthropic says the biggest gains are in long-horizon agentic work: multi-file features, larger refactors, and finishing end-to-end without leaving stubs behind.

Quick tutorial — picking between Sonnet 5, Opus 5, and Fable 5:

1. Default to Sonnet 5 for routine, high-volume work — it’s the cheapest and fastest of the three.

2. Move to Opus 5 once a task needs real reasoning depth or has to survive a long agentic loop — coding, refactors, multi-step business work. This is Anthropic’s own recommended starting point if you’re unsure which model to reach for.

3. Reserve Fable 5 for the genuinely hardest problems, where you need the absolute highest capability regardless of cost.

If you’re on Claude Code, the switch is already live: run /effort to see your active level, or /effort auto to reset to the model default (high on Opus 5).

Worth a look if you’re migrating: Anthropic’s what’s new in Opus 5 page covers the full behavior changes, and the effort parameter docs go deeper on tuning cost vs. quality per request.

AI of the Week

This week’s picks are for making sense of a market that’s moving this fast — comparing models, generating new kinds of media, and actually understanding the code you’re shipping.

• Artificial Analysis — the independent benchmark and leaderboard site nearly every story this week (Kimi K3, Opus 5, Gemini’s delay) ultimately points back to. If you want a real, unsponsored read on where a model actually stands, this is where to look.

• FLUX 3 — Black Forest Labs’ new multimodal model, generating images and up to 20-second videos with synced audio from a single prompt, with the same architecture already being tested to control robots on an Audi production line.

• DeepWiki — swap github.com for deepwiki.com in any public repo URL and get an instant, AI-generated wiki: architecture overview, dependency graphs, and a chat interface for asking questions about the codebase. Genuinely useful the next time you’re dropped into an unfamiliar repo.