Claude Fable 5, GPT-5.6, and the month AI started getting government sign-off

Anthropic released Claude Fable 5 on June 9. Three days later the US government put export controls on it and access was suspended worldwide. The controls lifted June 30, everything came back July 1, and today OpenAI answered with three new models called GPT-5.6 Sol, Terra and Luna. That's the fastest, strangest month this industry has had. We build websites and AI systems on these tools every day, so we followed it closely. Here's the short version, minus the jargon.

What Claude Fable 5 actually is

Fable 5 is Anthropic's most capable model to date, and the first one they've openly said is too capable to release without restraints. The same model exists in a second form, Claude Mythos 5, with fewer restrictions, reserved for approved security teams working with the US government. Fable 5 is the version the rest of us get.

The guardrails are narrow and specific. Ask it something genuinely dangerous, a bioweapon question or an offensive hacking request, and it blocks the answer and quietly hands your question to the older Claude Opus 4.8 instead. Anthropic says that happens in under 5% of sessions. We've never hit it doing normal business work. Day to day it just feels like a much sharper model that costs more: $10 per million input tokens and $50 per million output, exactly double the Opus tier under it.

Where it's best

The coding stories are the headline. Anthropic's launch post describes Stripe finishing a 50 million line code migration in a day, a job the team had scoped at two months. We can't audit Stripe's codebase, but the direction matches what we see on much smaller jobs: builds that used to take a week of back and forth now land the same day.

It also sees properly. Show it a screenshot and it can rebuild the page as working code. Hand it a chart and it reads the exact numbers off it. And it holds a thread: Anthropic measured it at roughly three times better than their previous model at keeping track of details across very long tasks.

None of that matters to you as trivia. It matters because the practical stuff improves underneath: the AI answering your phone keeps the whole conversation straight, the system building your website gets more right on the first pass, and long jobs run unattended without wandering off.

Why this changes the AI space

Two firsts happened in the same month. The obvious one: governments are now part of AI launches. Fable 5 was frozen for export three days after release because there was no reliable way to verify who was using it. While it was dark, researchers at Amazon reported a way around the guardrails. Anthropic patched it, says the reported technique is now blocked in over 99% of attempts, doubled its safety team, and got the restrictions lifted on June 30. OpenAI took the other route and cleared GPT-5.6 with the US Department of Commerce before launching at all. Frontier AI is being handled like critical infrastructure now, not like an app update.

The quieter first: these are not chatbots anymore. Both companies are selling models that run for hours and hand back finished work instead of a paragraph of suggestions. The shift is from search box to staff member, and that's the part that actually reaches small businesses.

So what did OpenAI just release?

GPT-5.6 comes in three sizes, public as of today after a preview limited to approved partners. Sol is the flagship at $5 per million input tokens and $30 out. Terra is the middle option at $2.50 and $15. Luna is the small fast one at $1 and $6.

Early numbers say Sol is a real rival. On Terminal-Bench, a benchmark for command line coding work, DataCamp's roundup puts Sol at 88.8% against Fable 5's 83.4%. One benchmark is one benchmark, and the order flips depending on the task. The real takeaway is that the two labs are trading blows weeks apart, and prices keep sliding while capability climbs.

Quietly, Luna might matter most for businesses like our clients. The always-on, high-volume work, answering phones, texting back missed calls, booking jobs, runs on cheap fast models. Cheap just got cheaper.

What we're doing about it

Not marrying either company. We test per job: big models for builds and audits, fast cheap tiers for the around-the-clock work, and we switch when something better ships. You shouldn't have to care what a Sol or a Fable is, the same way you don't care which brand of nail gun your roofer swears by.

If you'd rather someone else keep up with all of this, that's literally the job. Here's what we set up for local businesses, and the free visibility report shows where you're leaking leads today, whichever model ends up doing the work.