The biggest news this week is the release of Opus 4.7 and announcement of Mythos. Let’s see what is special about these models.
Opus 4.7 is best described as “marginally better at everything than 4.6.” I’m using it now and that feels true. It is finding the right files in my codebases a bit faster, giving me answers that are noticeably smarter and overall just feels smoother to work with. It’s kind of like what upgrading your iPhone felt like 6-7 years ago when upgrades actually meant something. Here is a comparison of 4.6 and 4.7’s benchmark scores:
It’s not a massive step change but it’s enough that you’ll see a difference.
Now onto Mythos. It’s aptly named as Anthropic as claimed it is finding zero day vulnerabilities in all kinds of publicly used software is too dangerous to release resulting in widespread rumors about its capabilities. Here is what we do know:
SWE-bench score of 93.9% and USAMO of 97.6% (but benchmarks ran by Anthropic)
It was purposely trained for cybersecurity usage
Only fixed consortium of institutions through Project Glasswing have access. I guess the rest of us can wait to be attacked once other models catch up. Anthropic has actually said they do not plan on releasing it to the public.
It will be priced at 5x Opus 4.7
The increase in price suggests a significant increase in parameters and compute used to train it.
This is the first model by Anthropic that is only available to certain institutions and not the public. On April, 14, 6 days after Mythos was announced, OpenAI announced GPT 5.4-Cyber with a similar institution-only rollout. Both labs have their own policies around what they deem safe for a public release and both those policies were triggered by these new models. It does seem that OpenAI’s announcement was reactive to Anthropic.
This does put us in new territory of unequal access to intelligence (beyond ability to pay). Up until now AI was sort of an equalizer- small startups, individuals and bootstrappers had access to the same frontier tech as large companies. Constraints around dev time and cost weakened at a similar amount for everyone- and arguably smaller companies were better positioned to take advantage by acting faster.
That is now over. The larger organizations have access to tools no one else has, though it is unclear what that means practically. If access to these tools allows companies to outcompete players that don’t have them then the labs behind the models can essentially choose which companies will thrive, begging the question where public policy belongs in this.