Anthropic released Claude Opus 4.7 on April 16, the latest step in its flagship line, alongside a new website-design tool. The model is a meaningful but measured upgrade from Opus 4.6: stronger coding autonomy on tasks that previously required close human supervision, and image-processing at resolutions up to 2,576 pixels on the long edge, roughly three times what earlier Claude models handled. What makes the announcement worth watching is not only 4.7 itself but the model Anthropic is not shipping. Mythos, a more powerful internal model, remains gated under the Project Glasswing safety framework and has been made available only to 11 partner organizations for cybersecurity vulnerability research.
On the 4.7 specifics, Anthropic's positioning is clear. The coding upgrade is the headline capability; the model can now drive multi-step coding work with less hand-holding, which matters for anyone shipping coding agents against the Anthropic API. The vision improvement is quieter but structurally important: 2,576 pixels on the long edge means Opus 4.7 can read dense interfaces, full-page screenshots, and larger diagrams without the downsampling artifacts earlier models introduced. On Mythos, the public posture is consistent with Anthropic's stated responsible-scaling policy. The model exists, it is more capable than Opus 4.7 on some dimensions, and it is in "Mythos Preview" with 11 named partners granted access specifically to help find and fix cybersecurity vulnerabilities. This is explicitly not a "the model is too dangerous to release" framing. It is a "the model is released, but only for a use case where its capabilities are net-positive, and broad access is gated."
The gated-access pattern is the part worth watching. Anthropic is threading a real needle here. If you refuse to release capable models, you concede the frontier to less cautious competitors; if you release them broadly, you accept capability-abuse risk with no counterbalancing deployment pattern. The Mythos posture is neither. It is a controlled-access deployment where the approved use case, cybersecurity research, produces a societal return on the capability release. Whether that framework scales to the next generation of models is the open question. Eleven partners is a size you can manage by hand, and capability-abuse risk is not the only axis of concern as models get broader. OpenAI's posture, by contrast, has been broader rollouts with RLHF guardrails rather than access gating. Both approaches have real costs, and the next 12 months will produce evidence on which one actually contains harm.
If you are building on Anthropic's API today, Opus 4.7 is a drop-in upgrade worth testing against your workloads, particularly for coding-agent and image-heavy use cases. The coding-autonomy improvement will quietly change the failure mode of your multi-step agents, so budget evaluation time before bumping model IDs in production. Mythos is not a public-API concern and will not be one for most builders in the near term. What it is worth tracking is the access-gating pattern, because some version of "capable model available only to vetted partners" is a plausible shape for future frontier deployments from multiple labs. If you have a credible cybersecurity research program, this is the kind of access posture you want to be eligible for. For everyone else, keep your abstraction layers honest: if model capability is gated by partnership status, the moat is relationship-dependent, not API-key-dependent, and that shifts how procurement should look over the next year.