QWE AI Academy The #1 mistake right now? Thinking you can access Claude Mythos. You can’t. This isn’t a launch – it’s an accidental leak of draft materials that weren’t ready for public eyes. Fortune broke it March 26, 2026. Hours later? Anthropic locked it down.
Real: Anthropic is testing an unreleased model internally. Uncertain: pricing, timeline, final name, exact capabilities. Here’s what actually matters.
The Leak
CMS misconfiguration. Draft blog posts sitting in a publicly searchable data store. Alexandre Pauwels (University of Cambridge) and Roy Paz (LayerX Security) found it independently, flagged it March 26.
Nearly 3,000 unpublished assets – blog drafts, images, PDFs, details of an invite-only CEO summit in Europe where Dario Amodei plans to pitch enterprise clients. Anthropic called it “human error” in their statement to Fortune. Same day? Access gone.
What’s left: screenshots, cached copies, second-hand reporting. Original materials? Deleted.
Mythos vs Capybara
Draft docs (per Fortune’s review): “Claude Mythos” = model version 1. “Capybara” = tier name. Fourth tier above Opus, Sonnet, Haiku. Both names, same model.
But Anthropic’s official statement only confirmed “developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity.” No names used. The draft was structured like a launch post – headings, publication date, the works. Still a draft though. Final branding could change completely.
Tiers Now vs What’s Coming
Today’s setup (March 2026, per Anthropic’s docs):
| Tier | Use Case | Current Model | Pricing (per 1M tokens) |
|---|---|---|---|
| Opus | Most capable | Claude Opus 4.6 | $5 input / $25 output |
| Sonnet | Balanced | Claude Sonnet 4.6 | $3 input / $15 output |
| Haiku | Fastest, cheapest | Claude Haiku 4.5 | $1 input / $5 output |
Capybara/Mythos slots above Opus. Leaked blog: “larger and more intelligent than our Opus models – which were, until now, our most powerful” and “by far the most powerful AI model we’ve ever developed.” Anthropic’s spokesperson: “a step change” and “the most capable we’ve built to date.”
Draft performance claims: “dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity” vs Claude Opus 4.6. Cybersecurity performance “far ahead of any other AI model in cyber capabilities” as of the draft date.
Think of it this way: if Opus is a Honda Accord (reliable, gets the job done), Mythos is a Bugatti. Faster, more powerful – but you’re not driving it to the grocery store, and you’re definitely not getting Honda pricing.
“Expensive to Run” Decoded
Draft says “expensive to run and not yet ready for general release.” Pricing figures? Zero.
API users budgeting for Capybara access: you’re guessing. Only anchor: “more expensive than Opus” ($5/$25 per million tokens as of March 2026). Reddit/Twitter speculation: 2-3x Opus pricing. That’s unverified.
Reality check: If Opus 4.6 already strains your budget, Capybara will hurt. “Expensive to run” means Anthropic hasn’t figured out how to make inference economically viable at scale yet. Early access pricing ≠ what public eventually pays – assuming it goes public at all.
Context: Anthropic dropped Opus 4.6 pricing 67% vs Opus 4.1 ($15/$75 → $5/$25). Precedent for aggressive cuts as infrastructure improves. But Mythos is positioned as fundamentally larger. Higher baseline costs.
One budgeting gotcha nobody’s talking about: the leaked draft says Mythos is “expensive to run” but gives NO pricing. Community speculation puts it at 2-3x Opus ($5/$25 per million tokens as of March 2026), but that’s pure guesswork from Reddit/Twitter – zero verification. If you’re planning Q3 API budgets around a 2x multiplier, you could be off by 50% in either direction. The only confirmed anchor: “more expensive than Opus.” That’s it.
Cybersecurity Risks
Draft doesn’t just hype capabilities – it flags risks. Per Fortune’s reporting, the leaked materials warned Mythos “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.”
Anthropic worries hackers could use this to find and exploit bugs faster than security teams patch them.
Planned release strategy (draft version): start with cyber defenders. Give them “a head start in improving the robustness of their codebases against the impending wave of AI-driven exploits” before wider access. Currently testing with “a small group of early access customers” focused on cybersecurity defense.
Same pattern: OpenAI’s GPT-5.3-Codex (released February 2026) was the first model classified “high capability” for cybersecurity under its Preparedness Framework. Both companies acknowledge latest frontier models cross a threshold – dual-use risk (helping attackers AND defenders) is real now.
Security teams: you might get access sooner. Everyone else: expect heavy gating, at least initially.
Still Unknown
- Public release timeline: Draft: “not yet ready for general release.” No date, quarter, or year.
- Actual pricing: “More expensive than Opus” = only hint. Figures? Unknown.
- Context window: Opus 4.6 offers 200K tokens (1M preview for Sonnet 4.5). Mythos specs weren’t in leaked materials.
- API availability: Early access happening. Whether Capybara hits public API or stays enterprise-only (like some OpenAI models)? Unclear.
- Draft vs final product: These were “early drafts of content considered for publication.” Specs, branding, strategy may all shift before launch.
Docs also don’t clarify benchmark conditions. If test setups for Mythos differed from public evals for Opus 4.6, comparisons get murky.
Another cross-reference gotcha: Mythos and Capybara refer to the SAME model according to leaked docs – v1 is “Claude Mythos”, tier name is “Capybara”. But Anthropic’s official Fortune statement only confirms “a new model” without using either name. If you’re tracking model releases for compliance docs or vendor evaluation, the final branding may change entirely before launch. Draft materials ≠ shipping product.
How to Prepare
Currently using Claude API in production? Here’s what to do:
1. Budget for tier inflation. Fourth tier above Opus = Anthropic is pushing capability ceiling higher and charging accordingly. If your workload requires the “most capable” model, per-token costs may jump when Capybara launches.
2. Monitor Opus 4.6 performance now. If it already handles your tasks well, you might not need Capybara. “Step change” language suggests diminishing returns for simpler use cases.
3. Track early access eligibility. Anthropic starting with cybersecurity-focused orgs. If that’s you: reach out to account rep (if you have one) or watch Anthropic’s announcements. Early access typically requires existing enterprise relationship.
4. Revisit prompt caching strategy. Capybara will likely support same cost optimizations as other Claude models (prompt caching, batch API). Not already using these to cut Opus costs 88-95%? Learn now – savings matter even more at higher price points.
Honestly? Don’t overreact. This is a leak of draft materials for a model that isn’t publicly available. Treat everything as provisional until Anthropic makes an official announcement with real specs, real pricing, real release date.
The leak also exposed details of an invite-only CEO summit in Europe with Dario Amodei – part of Anthropic’s enterprise sales strategy. Nearly 3,000 assets included summit invites, pricing deck drafts, and pitch materials. After Fortune notified Anthropic March 26, access shut down within hours. The researchers who found it (Pauwels at Cambridge, Paz at LayerX) likely have copies, but those aren’t public. If you’re looking for original materials, they’re gone – everything now is cached screenshots and second-hand reporting.
Frequently Asked Questions
Can I access Claude Mythos or Capybara tier right now?
No. Early access testing only. No public API, no waitlist, no timeline.
How much will Capybara tier cost compared to Opus?
Unknown. Draft: “more expensive” than Opus ($5/$25 per million tokens as of March 2026). No figures given. Community guesses range 1.5-3x Opus pricing – unverified. Anthropic has cut pricing aggressively before (Opus 4.6 was 67% cheaper than 4.1), so launch pricing may not stick long-term. One real scenario: if you’re running a chatbot that processes 50M tokens/month on Opus 4.6 (currently ~$1,500/month), switching to Capybara at 2.5x pricing would jump you to $3,750/month. At 3x? $4,500. Budget the high end until official numbers drop.
Is “Claude Mythos” the final name, or will Anthropic rebrand before launch?
Unclear. Leaked materials: “Claude Mythos” (model), “Capybara” (tier). Anthropic’s official Fortune statement didn’t confirm either. These were drafts – branding often changes between internal docs and public launch. Wait for it on anthropic.com or in API docs before treating names as final.