Anthropic's 'Claude-Mythos-5' Briefly Appears in Dev Mode, Likely Claude 5's Top Tier
A model slug labeled "claude-mythos-5" briefly surfaced in the model registry of Anthropic's internal developer tool "Dev Mode" around June 6, 2026, before being removed within minutes to tens of minutes. It is not an official announcement and appears to be an internal test or staging entry, but as a new generation of "Mythos"—the company's independent top-tier model class—it has drawn major attention among developers (details).
The slug's appearance was confirmed and spread by multiple developers, described in the context that "Mythos 5 appears first, followed by Opus 5 and Sonnet 5." Mythos is planned as a separate, independent model class from the conventional Haiku/Sonnet/Opus lineup, and the interpretation that it will become the top tier of the Claude 5 family is spreading. An output example generating a music video purely from code was also attached, drawing attention for its expressive power.
The backdrop is "Claude Mythos Preview," which Anthropic released on April 7, 2026. Positioned as the company's most capable frontier model ever, it significantly outperformed Opus 4.6. Mythos Preview is particularly strong in cybersecurity, coding, and reasoning, and is provided to only about 40 organizations through "Project Glasswing," an invitation-only program limited to defensive cybersecurity use. The appearance of Mythos 5 is taken as suggesting the next iteration of this Preview, or the full rollout of a new tier replacing Opus.
The performance of Mythos Preview is detailed in the official System Card (244 pages). The main metrics are as follows.
| Item | Mythos Preview (predecessor) | Notes |
|---|---|---|
| SWE-bench Verified | 93.9% | +13 points vs Opus 4.6, record-breaking |
| Benchmark lead | 17 of 18 | Only MMMLU close to Gemini 3.1 Pro |
| Cyber capability | Saturates existing benchmarks | Zero-day discovery and PoC creation possible with agentic harness |
| Parameter scale | ~10 trillion (unverified) | Not officially confirmed, leak reports only |
Regarding cyber capability, internal evaluations saturated existing benchmarks, and cases such as discovering a 27-year-old dormant bug in Firefox have been reported. At the same time, the decision to withhold general release due to this high capability—keeping it limited to Project Glasswing—has been flagged as a risk by outlets such as BBC.
As for Mythos 5 itself, official benchmarks, pricing, availability, and release timing are all unannounced, and only the Dev Mode slug has been confirmed. Limited access via Project Glasswing is expected to continue, and general release appears undecided. As for delivery environments, in addition to the Anthropic API and Claude platform, there is precedent of Mythos Preview being added via AWS Bedrock.
Reception is largely anticipatory, and because public access is extremely limited, widely shared real-world usage reviews remain scarce. On the positive side, some say music video generation and world-simulation-like outputs are possible purely from code, and there are voices recommending joining Dev Mode. In comparative contexts, there is debate that Mythos leads OpenAI's GPT-5.6 in some cyber capabilities, while GPT holds an edge in speed and cost. Concerns cited include inference costs from the enormous parameter scale and a cautious stance on alignment evaluations. Amid intensifying competition with GPT-5.6 and GPT-5.5, a structure is evident in which Anthropic is prioritizing cyber and agentic capabilities above all, and attention is focused on the official trajectory of Mythos 5.