Claude Mythos pricing: what Anthropic's leaked new model will cost developers
Anthropic accidentally exposed internal documents describing a new model tier above Opus. No pricing yet. But Anthropic already charges $30 per million input tokens for fast Opus 4.6 -- and that tells us something.
Image source: Anthropic
- Anthropic's CMS leak exposed Claude Mythos (codename Capybara) -- a new tier above Opus, confirmed real, not yet priced.
- Opus 4.6 Fast Mode already runs at $30/$150/M. That's the ceiling. Our best guess for standard Mythos pricing lands around $15/$75 -- back where Claude 3 Opus was.
What the leak actually showed
Security researchers Roy Paz (LayerX Security) and Alexandre Pauwels (University of Cambridge) discovered the exposed cache. Fortune reviewed the documents and notified Anthropic on Thursday; the company restricted access shortly after. Anthropic called it human error in CMS configuration.
The leaked draft blog post was direct about what Mythos is. It described Capybara as "a new tier of model: larger and more intelligent than our Opus models -- which were, until now, our most powerful." The Anthropic spokesperson confirmed it to Fortune: "We're developing a general purpose model with meaningful advances in reasoning, coding, and cybersecurity. Given the strength of its capabilities, we're being deliberate about how we release it."
The model hierarchy going forward looks like this:
| Tier | Current model | Role |
|---|---|---|
| Haiku | Claude Haiku 4.5 | Fast, cheapest |
| Sonnet | Claude Sonnet 4.6 | Balanced |
| Opus | Claude Opus 4.6 | Most capable (public) |
| Capybara | Claude Mythos | New frontier tier (unreleased) |
What the leaked documents said about capabilities
The leaked draft used strong language. Compared to Opus 4.6, Mythos gets "dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity." It said Mythos is "currently far ahead of any other AI model in cyber capabilities" -- with the ability to discover and exploit vulnerabilities in ways that Anthropic itself called a near-term risk requiring special handling before wider release.
Markets reacted immediately. Palo Alto Networks, CrowdStrike, and Fortinet all dropped 4-6% on the news. Worth noting: Anthropic's own CMS security failure exposed a model they were warning could enable large-scale cyberattacks.
What we don't have: specific benchmark numbers. The leak was qualitative -- no MMLU scores, no SWE-bench percentages, just directional language. For context, Claude Opus 4.6 already leads GPT-5.4 on SWE-bench Verified (81.4% vs 71.7%). Mythos being "dramatically higher" on coding suggests it could push well past 85%.
Anthropic's pricing history
Anthropic cut Opus pricing by 67% between the 4.1 and 4.5 generations. Claude 3 Opus and Claude Opus 4.1 both ran at $15/$75. Then Opus 4.5 dropped to $5/$25 and stayed there for 4.6. That pattern -- two generations at the same price, then a sharp cut -- makes the trajectory somewhat readable.
| Model | Input / 1M | Output / 1M | Notes |
|---|---|---|---|
| Claude 3 Opus | $15.00 | $75.00 | Launched Mar 2024, deprecated |
| Claude Opus 4 / 4.1 | $15.00 | $75.00 | Same tier, same price |
| Claude Opus 4.5 | $5.00 | $25.00 | 67% price cut |
| Claude Opus 4.6 | $5.00 | $25.00 | Current standard |
| Opus 4.6 Fast Mode | $30.00 | $150.00 | 6x premium, beta |
| Claude Mythos | ??? | ??? | Not yet published |
Pricing from Anthropic official docs, retrieved March 27, 2026.
What Mythos will probably cost
Two data points constrain the range. On the low end: Opus 4.5/4.6 is $5/$25. Anthropic wouldn't launch a new tier at the same price as the tier below it -- that erodes Opus revenue and confuses buyers. Some premium is structurally necessary. On the high end: GPT-5.4 Pro at $30/$180 is the current ceiling for what enterprises actually pay.
Opus 4.6 Fast Mode is the most useful reference point. Anthropic already charges $30/$150 per million for fast inference on their existing best model -- it's in beta, but the pricing infrastructure exists. A new tier doesn't have to exceed that; it just has to justify why you'd pay more than standard Opus.
| Scenario | Input / 1M | Output / 1M | Basis |
|---|---|---|---|
| Conservative (2x Opus) | $10.00 | $50.00 | Modest premium for new tier |
| Mid-range (3x Opus) | $15.00 | $75.00 | Back to Claude 3/Opus 4.1 levels |
| Fast Mode parity | $30.00 | $150.00 | Matches existing Opus 4.6 Fast Mode |
| GPT-5.4 Pro floor | $30.00 | $180.00 | Matches current ultra-frontier ceiling |
My read: the $15/$75 mid-range scenario makes the most narrative sense. It mirrors Claude 3 Opus pricing, which most enterprise buyers already have budget line items for. It's defensible against GPT-5.4 Pro ($30/$180) on price-performance. And if Anthropic wants a cybersecurity specialist tier that gets careful, deliberate adoption -- not mass deployment -- a $15 input price achieves that without the sticker shock of $30.
How it fits in the ultra-frontier tier
The ultra-frontier bracket is currently thin. GPT-5.4 Pro is the only widely available model above the standard $2-5/M input range. Gemini 3.1 Pro Preview is priced much lower ($2/$12) -- Google is competing on access rather than claiming an above-Opus-class tier.
| Model | Input / 1M | Output / 1M | Context | Availability |
|---|---|---|---|---|
| Claude Mythos | TBD | TBD | TBD | Not released |
| GPT-5.4 Pro | $30.00 | $180.00 | 1.05M | GA |
| Opus 4.6 Fast Mode | $30.00 | $150.00 | 200K | Beta |
| Claude Opus 4.6 | $5.00 | $25.00 | 200K | GA |
| Gemini 3.1 Pro | $2.00 | $12.00 | 1M | Preview |
| GPT-5.4 | $2.50 | $15.00 | 1.1M | GA |
Pricing from official provider docs, retrieved March 27, 2026.
What this means if you're building on Claude
Mythos is not a drop-in replacement for Opus 4.6. Anthropic is being "deliberate about how we release it" -- which means restricted access, probably tiered by use case. The cybersecurity warning in the leaked docs suggests Mythos will have tighter API access controls than previous models, possibly requiring separate agreements or tier verification.
If you're running cost-sensitive workloads, nothing changes today. Opus 4.6 at $5/$25 is still the best value in the upper tier. When Mythos ships, the calculus will depend on whether the capability jump on your specific tasks is worth a 2-6x price premium.
The interesting edge case is cybersecurity. If Mythos is genuinely "far ahead of any other AI model in cyber capabilities," it could create a specialized market segment with different pricing tolerance than general-purpose AI -- closer to professional security tooling than chatbot APIs. That context might justify higher prices than the pure general-purpose math suggests.
Timeline
No release date has been announced. From the leaked materials, Anthropic was scheduling demonstrations to enterprise clients at a European CEO summit -- suggesting months away from public availability, not weeks.
Community speculation centers on Q4 2026, possibly timed with Anthropic's reported IPO plans (Fortune separately reported Anthropic is exploring a Q4 2026 public listing). That's speculative -- but launching a frontier-beating model alongside an IPO roadshow would be a logical sequence. We'll update this page when Anthropic publishes official pricing.
Sources
- Fortune: Anthropic says it is testing Mythos, a powerful new AI model, after data leak reveals its existence (March 26, 2026)
- Fortune: Anthropic data leak reveals powerful secret Mythos AI model (March 27, 2026)
- Fortune: Leaked Anthropic AI Mythos model poses unprecedented cybersecurity risks (March 27, 2026)
- Winbuzzer: Anthropic confirms leaked Mythos model, step-change reasoning (March 27, 2026)
- Anthropic pricing docs (official)
- OpenRouter: GPT-5.4 Pro pricing
When Mythos pricing drops, you can model the exact cost for your workload on TokenCost.