Project Glasswing announcement
Official Anthropic page describing the initiative, partner access, Mythos Preview, benchmark context, credits, and participant pricing.
← Back to Compute College
Compute College
What is Claude Mythos Preview?
Claude Mythos Preview is an unreleased Anthropic frontier model used in Project Glasswing for defensive cybersecurity work.
Compute & Pricing LessonsLearning path
One concept connected to AI compute market decisions.
5-8 minutesRead time
A practical introduction designed to be completed in one sitting.
Benchmarks / Security / Serving CostTags
Useful for security leaders, ai infrastructure buyers, analysts, and developers tracking frontier-model compute demand.
Plain-English definition
Claude Mythos Preview is an unreleased Anthropic frontier model made available to selected Project Glasswing partners for defensive cybersecurity work. It matters to AI compute markets because a model can create concentrated, high-intensity inference demand before it becomes a general public API product.
Why it matters
Why it matters
Mythos Preview shows that frontier-model demand is not limited to public chat or coding APIs. Anthropic describes partner use across vulnerability finding, exploitability analysis, and critical-software defense, backed by a large usage-credit commitment and later participant pricing. That makes it a restricted-access compute signal: limited users, unusually demanding workloads, and premium token economics.
- Restricted access can concentrate usage among a small number of high-value partners instead of broad consumer traffic.
- Cybersecurity agent workflows can be long-running and token-intensive because they combine code reading, reasoning, testing, and verification loops.
- Public benchmark claims become market evidence only when they connect to partner usage, cloud availability, or priced API access.
Simple example
Simple example
Anthropic says Project Glasswing includes a $100 million model-usage-credit commitment during the research preview and later participant access to Mythos Preview at $25 per million input tokens and $125 per million output tokens. An illustrative 100,000-input-token and 20,000-output-token task at those later participant rates would cost $2.50 for input and $2.50 for output before any platform, caching, or workflow differences.
- This arithmetic is illustrative and uses Anthropic-published participant pricing, not a current public retail quote.
- The same task can cost more or less depending on context size, output length, retries, tool rounds, and verification steps.
- Usage credits can hide near-term spend while still representing real model-serving load on the underlying infrastructure.
Example figures are illustrative calculations, not current quoted market prices.
Current example
What Anthropic published
Anthropic announced Project Glasswing in April 2026 as a defensive cybersecurity initiative using Claude Mythos Preview with launch partners including major cloud, software, security, and infrastructure companies. Anthropic describes Mythos Preview as restricted, not generally available, and says Project Glasswing partners use it to find and fix vulnerabilities in critical software. A May 2026 update says Anthropic and partners had used Mythos Preview to find more than ten thousand high- or critical-severity vulnerabilities across important software.
Project Glasswing initial update
Official update with early public evidence from Project Glasswing and vulnerability-triage context.
Mythos Preview cybersecurity assessment
Anthropic Frontier Red Team post about high-level capability testing and defensive-security motivation.
Source discipline: this lesson summarizes Anthropic-published claims and keeps the market read at the level of access, pricing, workload intensity, and defensive use. It does not provide vulnerability-finding instructions. Last checked: June 1, 2026.
Market signal
How to read the market signal
The market signal is restricted frontier capacity. Mythos Preview is not a normal public model rollout, so demand is visible through partner programs, usage credits, cloud-platform access, pricing, benchmark disclosures, and follow-up evidence about the volume of defensive work being processed.
- Watch whether Anthropic expands Mythos-class access, launches safeguards, or moves similar capabilities into an upcoming Opus-class model.
- Watch cloud distribution: access through the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry changes where inference demand can land.
- Watch workload intensity: long-running security agents can consume far more tokens than short chat or summarization tasks.
Market read: Mythos Preview is evidence that some frontier-model compute demand may appear first as restricted partner access, not as open public API traffic.
Common mistake
Common mistake
Do not treat Mythos Preview like a standard model launch. It is restricted and defensive-use focused, so public benchmarks, availability, and pricing should not be read the same way as a generally available chat or coding model.
Practical takeaway
What you can do with this
Use Mythos Preview as a case study for restricted frontier-model deployment. Track who can access the model, which clouds can serve it, what usage credits or prices apply, and whether the work creates repeatable high-token security workflows.
- Security buyers: separate defensive evaluation interest from actual access, policy, and procurement terms.
- Infrastructure analysts: treat usage credits and partner programs as demand clues even when revenue timing is unclear.
- Developers: compare benchmark claims with real workflow cost, latency, tool use, and verification burden.
Decision check: identify whether the model is generally available, who can use it, what the published price or credit commitment is, and what workload pattern would turn capability into sustained compute demand.
Helpful memory trick
Helpful memory trick
Public API models are storefront demand. Mythos Preview is partner-program demand: fewer doors, heavier workloads, and stricter rules.
Compute College
Follow model releases as market signals
Follow model releases as AI compute market signals in the ComputeTape Morning Brief.
Compute College track
Model Benchmarks & AI Compute Economics
Step 15 of 23: What is claude mythos preview