Detailed Analysis
Anthropic launched Claude Opus 4.7 on April 16, 2026, positioning it as a meaningfully upgraded but deliberately scoped successor to Claude Opus 4.6, with targeted improvements across coding performance, vision capabilities, and operational efficiency. On CursorBench, Opus 4.7 scores 70% compared to 4.6's 58%, while its vision accuracy on high-resolution images (3.75MP) leaps from 54.5% to 98.5% — a dramatic jump that significantly expands its usefulness for analyzing documents, screenshots, and complex diagrams. The model also achieves three times more production task completions on Rakuten-SWE-Bench and resolves four additional tasks on a 93-task coding benchmark that its predecessor could not handle. Efficiency gains are equally notable: Opus 4.7's low-effort mode now matches the medium-effort output of 4.6, and the model maintains stronger consistency across its one-million-token context window. Pricing remains unchanged at $5 per million input tokens and $25 per million output tokens, making the upgrade cost-neutral for existing users.
The central tension in Opus 4.7's release is its relationship to Claude Mythos Preview, the more powerful model released on April 7, 2026, under Project Glasswing, which remains restricted to a limited set of trusted teams. Mythos outperforms Opus 4.7 on raw capability benchmarks by substantial margins — achieving 93.9% on SWE-bench Verified and saturating Cybench at 100% — while also demonstrating autonomous agentic abilities in high-stakes domains such as exploit development, zero-day vulnerability detection, and reverse engineering. These capabilities, precisely because of their power, make broad public deployment of Mythos a significant safety risk, which is why Anthropic has withheld it from general availability with no published pricing or confirmed context window specifications. Opus 4.7, by contrast, incorporates lessons and architectural refinements derived from Mythos development, effectively functioning as a "halfway" model — more capable than 4.6, but constrained to remain safely deployable at scale.
This release reflects a deliberate two-track strategy that Anthropic has been refining as frontier AI capabilities accelerate into increasingly sensitive domains. By enhancing Opus 4.7's cybersecurity safety features — specifically its improved detection and blocking of high-risk requests — Anthropic is acknowledging that the gap between what a model *can* do and what it *should* be allowed to do in general deployment is widening. The existence of Mythos as a restricted research model signals that Anthropic is actively developing capabilities it is not yet willing to release, a posture that reflects both competitive pressure from other frontier labs and internal commitments to safety-gated deployment. This approach mirrors practices seen across the broader AI industry, where the most powerful models are increasingly evaluated and staged through trusted-user pipelines before any public release.
The broader trend illustrated by the Opus 4.7 launch is the maturing of a tiered AI ecosystem in which public-facing models serve as carefully calibrated proxies for internal research capabilities. For the vast majority of developers and enterprises, Opus 4.7 represents a genuine and immediately actionable upgrade — fewer iterations, higher-quality outputs, and better multimodal performance — without the access restrictions that make Mythos impractical for most workflows. However, the widening benchmark gap between Mythos and publicly available models raises a longer-term question about how Anthropic plans to eventually bridge that divide, and under what safety conditions more of Mythos's capabilities might be made broadly accessible. The answer to that question will likely define the company's competitive positioning through the remainder of 2026 and into the next generation of model releases.
Read original article →