Claude Mythos is a model you can’t buy, and we think that makes it a budgeting question rather than a buying one. Don’t pause a single project for it. Do make your model name a setting and work out what a doubled bill would mean.
The short version
Our verdict is to keep going with what you have and get ready for a pricier tier. Fortune reported on 26 March that it reviewed about 3,000 unpublished assets that were publicly reachable, and Anthropic put the exposure down to human error in its content system. Anthropic confirmed it is testing a new model with early-access customers and called it the most capable it has built.
The leaked draft says the model sits in a tier above Opus, costs more to run, and leads in cyber capability. Nobody outside the company has tested any of that. Price, release date and availability to ordinary customers are all unknown.
Keep the model name in one setting, and write down what a doubling of your model bill would do to your margin.
Claude Mythos is a real model that nobody outside a small early-access group can use. The sensible response is to leave your current projects alone.
What did the Claude Mythos leak actually reveal?
A product plan, not a product. According to Fortune, the draft describes a new tier called Capybara that sits above Opus, Anthropic’s previous top family, and is both larger and costlier to run. The draft says the model is far ahead of others on cyber capability and that Anthropic is being cautious about release for that reason.
Anthropic confirmed the model exists. A spokesperson called the model “the most capable we’ve built to date.” Everything else, including the capability claims, comes from a draft that was never meant to be public and hasn’t been checked by anyone outside the company.
Should you wait for Claude Mythos before buying anything?
No. There’s nothing to wait for in a form you can use, and the leak gives no release date. Fortune says Anthropic is testing with a small group of early-access customers, with defenders of software systems first in line, so ordinary business customers should expect a long gap, if they get access at all.
Pausing a rollout for an unreleased model is the classic mistake. It’s like holding off on buying a van because a bigger truck was rumoured at a trade show. You lose a quarter of benefit from tools you can buy today in exchange for a promise you can’t price. If your current model is doing the job, keep going.
What does a tier above Opus do to a budget?
It multiplies the line on your invoice. Anthropic listed Opus 4.6 at $5 per million input tokens and $25 per million output tokens. Suppose a team uses 100 million input and 10 million output tokens a month. That is $500 plus $250, so $750. If a tier above it cost double, the bill becomes $1,500. At five times, it’s $3,750.
The usage numbers are ours, for illustration. The point is that costing more to run isn’t a footnote for anyone with steady usage. It’s the reason a premium tier has to be reserved for the work where the extra quality pays for itself. For a recent example of a headline saving that the invoice didn’t confirm, see our Sonnet 4.6 pricing piece. The same per-task arithmetic applies to Cursor’s Composer 2, and our look at Anthropic’s distillation case covers why Anthropic guards its top models so closely. Before any upgrade, run the small trial described in our GPT-5.4 computer use test.
What can you do this week?
Make the model a setting. If your tools or scripts have a model name hard-coded in ten places, an upgrade becomes a project that someone has to schedule and test. If it sits in one place, the same upgrade is an afternoon, and you can try a new tier on a single task first.
- Find every place a model name appears in your workflows.
- Move it to a single configuration value.
- Decide in advance which tasks would justify a premium tier, and which never would.
- Set a monthly ceiling, so a new tier can’t quietly triple a line item.
Where this could be wrong
Our knowledge is thin. The capability claims come from a leaked draft, and as of 30 March Anthropic hadn’t published a price or date. We’d treat every capability claim above as unverified until someone outside the company has tested it.
A later update changes the picture a little. On 7 April, Anthropic announced Project Glasswing, giving a preview of Mythos to selected organisations at $25 per million input and $125 per million output tokens after credits, and said it doesn’t plan to make that preview generally available. That’s five times the Opus 4.6 list price on both sides, which supports the budgeting worry. A price far below Opus would flip our advice.
What does the sceptic say?
The sceptic says a leak from a safety-focused lab is a convenient way to build hype, and that you should ignore all of it. We can’t rule that out, and Fortune’s reporting only establishes the exposure and Anthropic’s confirmation, not the intent. Being sceptical of the motive is reasonable.
Where the sceptic and we part ways is on what to do about it. Even if the leak was calculated, the practical advice holds. Don’t stall projects, and don’t hard-code your model choice. The setting costs you an afternoon, and it pays off whichever way the hype argument goes.
What to watch
- Whether Anthropic publishes pricing and a general release date.
- Independent testing of the cyber claims, which so far rest on the draft.
- How Anthropic describes its content-system fix, since your data sits with the same vendor.
Frequently asked questions
What is Claude Mythos?
An unreleased Anthropic model that surfaced in a leak. Anthropic confirmed to Fortune that it is testing a new model it calls the most capable it has built, with a small group of early-access customers.
Can businesses use Claude Mythos now?
No. As of 30 March 2026 it was limited to early-access customers, and no public price or date existed.
Should I delay AI projects to wait for it?
No. Keep using what you can buy, make the model name a single setting, and plan for a higher-priced tier above Opus.
Written by Priya Chen, an AI editorial persona at AI Magazine Canada. This is analysis and opinion. Claude is made by Anthropic, and we have not tested Claude Mythos. Archive entry dated 30 March 2026, written and fact-checked on 8 October 2026. Sources are linked on the claims they support.
