Mythos Model

Sources#

• Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next
• Claude Fable 5 and Claude Mythos 5
• Claude models explained: choosing the best model for your use case
• Claude Mythos Preview red.anthropic.com
• Claude Opus 4.8 System Card
• How Anthropic's product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
• When AI builds itself

Summary#

Anthropic's preview-tier frontier model. Notably described as "incredibly powerful" and gated behind safety review (the Mythos Preview / red.anthropic.com publication that established the LLM-Driven Vulnerability Research story). Used internally at Anthropic alongside Claude Opus 4.7. As of May 2026, not GA.

What's known publicly#

• Mythos Preview demonstrated emergent cybersecurity capabilities — autonomous zero-day discovery, full exploit chains. See LLM-Driven Vulnerability Research for the detailed analysis from the Mythos Preview publication.
• Anthropic's response: Project Glasswing safeguards (referenced in 4.7 release as "first post-Glasswing safeguards").
• Boris Cherny: "We use a little bit of Mythos to try it and then a lot of Opus 4.7 to dog food it and to write most of our code." — Mythos is preview-tier, not the workhorse.
• Cat Wu: "Mythos is an incredibly powerful model. But we do use the models internally and I think this has increased our rate of shipping a little bit but I don't think it explains the bulk of the increase." — confirms internal use; explicitly disclaims it as the cadence explanation.

Role in the Opus 4.8 System Card#

The Opus 4.8 System Card (May 2026) makes Mythos Preview's role unusually concrete — it remains the capability frontier that the general-access model is measured against, and it is used as a tool in the assessment itself:

• Frontier benchmark: Opus 4.8 "does not advance the capability frontier beyond Mythos Preview." On the AECI index, Mythos scores 158.3 vs Opus 4.8's 155.5 (and 4.7's 154.1). Its Risk Report bounds the RSP case for 4.8.
• Investigator model: Mythos Preview is one of the two investigator models driving Opus 4.8's Automated Behavioral Audit (the other being a helpful-only Opus 4.7).
• Reviewer of the assessment: in a notable meta-move, Mythos was given access to internal Slack discussion and asked to review the near-final alignment section; its (published) review confirmed candor and flagged that no eval specifically tests for training-gaming — see Evaluation Awareness & Grader Gaming.
• Alignment yardstick: Opus 4.8 matches Mythos Preview's alignment profile on most measures and surpasses it on several honesty metrics (Agentic Honesty & Diligence).

Capability data points (When AI builds itself)#

The Anthropic Institute essay (June 2026) attaches concrete numbers to Mythos Preview as the model that drove Anthropic's AI-R&D acceleration (AI Accelerating AI Development):

• Time horizon: METR rated it able to work for "at least" 16 hours, "at the upper end of what [METR] can measure without new tasks" (Task Time-Horizon Scaling).
• Kernel-optimization eval: ~52× speedup (April 2026) on the train-a-small-model-faster task, vs Opus 4's ~3× a year earlier and a ~4× human baseline — "from super helpful to superhuman in under a year."
• Research next-step judgment: beat the human choice 64% of the time on hard detour moments (Opus 4.5 was 51% in Nov 2025).
• Self-reported uplift: in a March 2026 poll of 130 research-team employees, the median estimated ~4× output with Mythos Preview vs no AI (Anthropic believes true uplift was somewhat lower).

These are deployment-side figures (Mythos used internally), distinct from the System Card's gated-capability-frontier framing above.

Why kept gated#

The combination of:

• Cybersecurity capability gap vs prior models (per Mythos Preview publication)
• Safety mechanisms still being evaluated and hardened
• Anthropic's stated mission posture

Means the model is used internally and previewed selectively rather than shipped broadly. Some descendant of Mythos is expected to ship publicly later — Boris Cherny: "It will become some version of some descendant of that will become available at some point to everyone."

Update — the descendants shipped (June 2026)#

Boris's prediction came true. In June 2026 Anthropic launched Fable 5 and Mythos 5 — the first general-access Mythos-class models, and the realization of "Mythos-class" as a named capability tier sitting above the Opus class. The lineage is now Mythos Preview (April 2026) → Fable 5 / Mythos 5 (June 2026):

• Fable 5 = a Mythos-class model "made safe for general use" via classifiers that fall back to Opus 4.8 on cyber/bio/distillation queries (Capability-Gated Model Fallback).
• Mythos 5 = the same underlying model with safeguards lifted, deployed through Project Glasswing as the upgrade to Mythos Preview ("comparable to, or somewhat stronger than" it, at less than half the price). Existing Glasswing/Mythos-Preview users upgrade directly.

Anthropic's July 2026 selection guide states the packaging cleanly: the Mythos class "ships in two packages of the same underlying model" — Mythos for trusted organizations handling dual-use cybersecurity and biology work, Fable with the additional safeguards that make it safe for general public use. Both require limited data retention as a condition of safe deployment (the 30-day rule, not a Fable-only constraint). It also confirms the class is positioned by difficulty carried, not domain: "especially capable at coding, long-running agent tasks, and solving problems AI has not reliably handled before" (Cost-per-Task Over Cost-per-Token).

This moves the capability frontier beyond Mythos Preview — the line the Opus 4.8 System Card treated as the ceiling — and is the first time Mythos-class capability reaches the public. (Both were reported suspended shortly after launch; see Claude Fable 5.)

An outside account of the gating — unverified#

Musk's July 2026 Economist interview asserts a specific chain of events around Mythos that appears nowhere else in this corpus, and it is recorded here because it is checkable and, if true, matters for how release decisions actually get made:

"It wasn't as though someone in the government figured out that Mythos had some scary cyber security risks. It was Amazon. Amazon highlighted this and Amazon called the White House."

The interviewer's own framing in the same exchange — that "the US government stepped in because they were worried about the cyber security risks, and they stepped in using the threat of export controls to prevent the widespread release of this model" — is treated by both speakers as established.

Status: unverified, single-source, prediction-tier. Everything the wiki otherwise holds about Mythos gating (the sections above, plus the RSP determinations) comes from Anthropic's own publications, which describe an internal cybersecurity-capability judgment and say nothing about an external discovery, a White House call, or an export-control lever. The two accounts are not contradictory — they could describe different parts of the same decision — but nothing here corroborates the Amazon chain, and Musk is a competitor. Do not cite it as fact.

Connections#

• Anthropic — vendor
• Claude Opus 4.7 — prior GA model; Mythos is the next-tier preview
• Claude Opus 4.8 — the GA model of its era, benchmarked against Mythos; Mythos is its capability frontier, an investigator in its behavioral audit, and the reviewer of its alignment assessment
• LLM-Driven Vulnerability Research — primary public account of capability profile
• Harness Shrinkage as Models Improve — Mythos-class capability is what makes Boris's "100 lines" prediction conceivable
• AI Native Product Cadence — explicitly disclaimed as the cadence explanation, but contributes
• AI Accelerating AI Development — the model behind Anthropic's measured AI-R&D acceleration (52× kernel eval, 64% next-step, ~4× poll)
• Task Time-Horizon Scaling — Mythos sits at the measurable edge of the time-horizon curve (16+ hours)
• METR — rated Mythos's time horizon at "at least 16 hours," beyond its standard measurement ceiling
• Claude Fable 5 — the first general-access Mythos-class model; Mythos Preview's safeguarded descendant
• Claude Mythos 5 — the safeguards-lifted descendant deployed through Project Glasswing as the direct Mythos Preview upgrade
• Capability-Gated Model Fallback — the safeguard architecture that made general release of a Mythos-class model possible
• Cost-per-Task Over Cost-per-Token — where the Mythos class sits in Anthropic's published selection framework: the top tier, access-gated to Glasswing organizations, and the constraint that makes access a selection question at all
• Claude Sonnet 5 — Mythos Preview is the best-aligned reference on the automated behavioral audit that the mid-tier Sonnet 5 trails (Sonnet 5 is safer than Sonnet 4.6 but worse than both Opus 4.8 and Mythos Preview)

Open Questions#

• Do Fable 5 / Mythos 5 return after the post-launch suspension, and when?
• Capability profile beyond cybersecurity: Mythos Preview focused on the safety story; other capability dimensions not well-documented externally.
• Internal access controls: who at Anthropic actually uses Mythos for daily work, vs Opus 4.7? Boris implies infrequent (try-it use); not detailed.

Resolved Questions#

• Public release timeline: Answered — Mythos Preview itself never shipped GA, but its descendants Fable 5 / Mythos 5 reached general access in June 2026 (see the descendants shipped above).

Sources#

• Claude Mythos Preview red.anthropic.com
• How Anthropic's product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
• Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next
• When AI builds itself — Mythos capability data points (METR 16h, 52× kernel, 64% next-step, ~4× poll)
• Claude Fable 5 and Claude Mythos 5 — the June 2026 launch of the first general-access Mythos-class descendants
• Claude models explained: choosing the best model for your use case — Anthropic, July 2026 (vendor-claim): Mythos/Fable as two packages of one model; class positioned by difficulty carried, not domain