Claude Mythos: Anthropic's Most Powerful Model Is Here

The Preview Release

Anthropic released the Mythos Preview in April 2026. It is the new top of the Claude lineup, sitting above Opus 4.7 in reasoning capability and positioned for the workloads where Opus is not quite enough. The underlying API model ID is claude-opus-4-7-6-20261120, exposed under the public alias mythos.

The preview is intentionally limited. Anthropic is gathering feedback on real workloads before opening it to the full customer base, so access is restricted to invited accounts via the Anthropic dashboard or the API waitlist. If you already have access, you will see mythos in your model list. If you do not, the waitlist is open and turnaround has been on the order of weeks.

Pricing and Context

Input: $25 per million tokens
Output: $125 per million tokens
Context window: 1,000,000 tokens
API model ID: claude-opus-4-7-6-20261120
Alias: mythos

The pricing is roughly five times Opus 4.7. That sounds steep until you do the math on the problems Mythos is designed for: tasks where a wrong answer costs hours of engineering time, where the alternative is shipping a security bug, or where the work is high-stakes enough that the model cost is rounding error against the value of getting it right.

Mythos vs Opus 4.7

Dimension	Opus 4.7	Mythos Preview
Context window	1M tokens	1M tokens
Input price	$5 / Mtok	$25 / Mtok
Output price	$25 / Mtok	$125 / Mtok
Reasoning depth	Strong	Significantly stronger on multi-constraint problems
Agentic workflows	Excellent	Holds coherence across longer chains
Code generation	Excellent	Better on novel problems, similar on common patterns
Availability	General availability	Preview — invite or waitlist

Where Mythos Dominates

The clearest wins for Mythos show up on problems with many interacting constraints. Security reviews where the threat model is non-obvious. Refactors that touch a system's invariants. Scientific or analytical work where the model has to hold a complex hypothesis space in view. Long agentic chains where each step builds on the last and drift kills the work.

On routine coding tasks the difference is smaller. A function-writing task that Sonnet 4.6 handles in two minutes will still get done in two minutes on Mythos, just at five times the cost. The economic case for Mythos is task-shaped, not session-shaped — pull it in for the specific problem that needs it.

Calling Mythos from Claude Code

# One-shot via the -p flag
claude -p "audit this auth flow for token leakage" --model mythos

# Interactive session pinned to Mythos
claude --model mythos

# Switch mid-session
/model mythos

Direct API calls work the same way, just with the full model ID:

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-opus-4-7-6-20261120",
    "max_tokens": 4096,
    "messages": [{"role": "user", "content": "..."}]
  }'

Why the Preview Is Limited

Anthropic is being deliberate. The preview window lets them collect real-workload feedback before opening Mythos to the full customer base. It also gates capacity — the model is more compute-intensive per token than Opus 4.7, and rolling it out to everyone immediately would push response latencies in a direction Anthropic does not want.

The practical upshot: if Mythos solves a problem that matters to your work and you do not have access yet, get on the waitlist. The preview phase is the right time to be evaluating it, not after general availability when capacity pressure has set in.