Claude Mythos 5

What it is

Claude Mythos 5 is the restricted sibling of Claude Fable 5: the same underlying capability, shipped without the consumer-facing safeguards, into Anthropic's Project Glasswing programme for defensive cybersecurity work. It is invitation-only, with no self-serve sign-up and no published price. Fable is the publicly described variant with the stricter safeguards; Mythos is the one you cannot simply buy.

Most readers will never touch it. It is on the board because it shows up in comparisons, and a score with no route to access is worth labelling as such.

How the composite is built

Overall 90 weights Coding 24%, Terminal 20%, Reasoning 20%, Tool use 15%, Context 13%, Speed 8%. Its profile is Fable 5's to within a point on every axis, which is exactly what you would expect from a variant that differs in policy rather than architecture.

Coding 94

Level with Fable 5 and two behind Opus 5. Frontier-grade, but not the top of this board.

Terminal 90

Identical to Fable 5, five behind Opus 5. Slow deliberation costs points on shell work regardless of what the safeguards allow.

Reasoning 96

Second only to Fable 5's 97 and the axis that carries this model. For security analysis — reading unfamiliar code for intent rather than correctness — this is the number that matters.

Tool use 92

Level with Fable 5, four behind Opus 5's board-leading 96. Solid across long chains of scanners, parsers and shell calls without being the best available at it.

Context 95

Matches Fable 5. Useful when the unit of work is a whole codebase or a long capture rather than a single file.

Speed 48

Joint slowest scored model, tied with Fable 5. Defensive analysis is usually a batch job, so this hurts less here than almost anywhere else on the board.

Where it fits

Defensive security work where the standard refusal surface gets in the way: malware analysis, exploit triage, reading attacker tooling, red-team support inside an authorised engagement. The access model is the product as much as the weights are — Anthropic is gating who gets an unguarded frontier model rather than selling one.

Limits

• You probably cannot get it. Invitation-only, no self-serve. Plan your system on Opus 5 or Fable 5 and treat Mythos as a bonus if an invitation arrives.
• No published price means no value rating. The board shows a dash, not a poor score. If it follows Fable 5's $10/$50 the economics would be the worst on the board, but that is an inference, not a published figure.
• Fewer safeguards is a liability transfer, not a free upgrade. Whatever the model will now discuss, your logging, authorisation and scope discipline have to carry.
• Scores here are an early read. A restricted model gets less independent testing, so treat these numbers as lower-confidence than the widely available tiers.

Price and access

No published price and no public endpoint as of this scoring. Access runs through Project Glasswing by invitation. Anything you read quoting Mythos pricing is inferring it from Fable 5, which shares its specifications.

Alternatives on this board

• Claude Fable 5 — the same specifications with safeguards, generally available at $10/$50. The realistic substitute.
• Claude Opus 5 — 92 overall, better coding and terminal scores, $5/$25, and buyable today.
• DeepSeek R2 — Reasoning 93 from a much cheaper reasoning line; worth testing before assuming only a frontier model can do the analysis.
• DeepSeek V4 — open weights you can run inside your own boundary, which is sometimes the real requirement behind wanting an unguarded model.

Sources