Claude Opus 4.7: Benchmarks, Cyber Safeguards and Open Questions

In April 2026, Anthropic released Claude Opus 4.7, a new version of a model family I use frequently. I am struck by the reported improvement over 4.6 and by its proximity on some tests to Mythos Preview, which Anthropic had restricted to a limited set of defensive cybersecurity partners. I want to understand how the company drew that deployment distinction.

I speculate about whether model size, training runs or future versions influenced the decision. The video repeats a rumor that Mythos has ten trillion parameters and guesses a smaller size for Opus; neither figure is established by the cited materials. My proposed explanations of Anthropic’s motives are hypotheses, not confirmed internal facts.

Anthropic’s table reports SWE-bench Pro scores of 53.4% for Opus 4.6, 64.3% for 4.7 and 77.8% for Mythos Preview. On SWE-bench Verified, the displayed scores are 80.8%, 87.6% and 93.9%. These are substantial differences on those evaluations, but benchmark setups and real coding workflows can differ.

I interpret the emphasis on coding as part of a business and research feedback loop: useful coding tools can attract enterprise demand and help teams build future systems. That is my strategic interpretation, not proof of a fully autonomous self-improvement loop. I also wonder whether the limited Mythos announcement has marketing effects, but I have no evidence that this was the reason for the decision.

The displayed results show a roughly seven-point gain on Humanity’s Last Exam without tools and a smaller gain with tools, while agentic search drops from 83.7% to 79.3%. On the displayed computer-use test Opus 4.7 reaches 78%, compared with Mythos Preview’s 79.6%. Improvements are uneven across tasks rather than a uniform increase everywhere.

The CyberGym vulnerability-reproduction scores shown are 73.8% for Opus 4.6, 73.1% for 4.7 and 83.1% for Mythos Preview. I ask whether Anthropic intentionally reduced cyber capability. The company says it experimented with differential reductions during training, but a small difference on one test cannot isolate the cause or measure all cyber abilities; Anthropic characterizes 4.7 as roughly similar to 4.6 in its cyber evaluations.

I conjecture that Mythos is a new, much larger training run and that future iterations could improve further. Those model sizes and the details of its training are not confirmed in the video. Anthropic’s stated distinction concerns its evaluations and safeguards, not a publicly specified parameter threshold.

Anthropic describes Opus 4.7 as stronger on difficult software-engineering tasks and says it follows instructions carefully and checks work. It also reports gains in vision and professional outputs such as interfaces, slides and documents. I like Opus for front-end work, but that is my experience and preference rather than a controlled comparison.

The company says it limited Mythos Preview under Project Glasswing while testing new cybersecurity safeguards on a less capable model first. It says Opus 4.7 includes safeguards intended to detect and block prohibited or high-risk cyber requests, and that deployment experience will inform future decisions. Legitimate security work can also be affected, so the company describes a verification program.

I speculate that a stronger model might supervise a weaker one to detect harmful activity. The release statement does not establish that Mythos is monitoring Opus in this way, so that remains my conjecture. The broader question is how to test safeguards against misuse while preserving legitimate defensive work.

Anthropic says Opus 4.7 follows instructions more literally, which can make prompts written for older models behave unexpectedly. I recommend reviewing model-specific prompting guidance when changing models and adjusting prompts or agent harnesses based on testing. Greater instruction adherence can be useful, but it also makes precise instructions and evaluation important.

The release also reports better multimodal support, professional work and use of filesystem-based memory over longer tasks. The ability to use stored notes depends on the application and its tools, not simply on a model possessing a universal persistent memory. I then review Anthropic’s prerelease charts for several distinct tasks.

On the displayed GDPVal-AA knowledge-work evaluation, Anthropic reports an Elo score of 1,753 for Opus 4.7, compared with 1,619 for 4.6 and 1,674 for GPT-5.4. Other charts show gains in screenshot navigation, document reasoning, long-context reasoning, biomolecular reasoning and a simulated vending-machine task. These are specific tests and scoring methods; they are not proof of superiority on all professional work.

I notice that Anthropic’s automated behavioral audit gives Mythos Preview a lower misaligned-behavior score than the Opus models, and ask why the stronger model remained limited. The company’s release decision also considers capability and misuse risk; a score from one alignment evaluation does not establish overall safety or make the two deployment decisions interchangeable.

Opus 4.7 adds an extra-high effort setting between high and max and a Claude Code review command. These options let users choose different tradeoffs between effort, latency and output. The usefulness and cost of a setting should be measured on the actual task rather than assumed from its name.

I suggest that compute constraints may also have influenced the limited release of Mythos, citing user quotas and access changes. That is speculation; the public explanation emphasizes cybersecurity risk and safeguards. Capacity and pricing can change, and the video does not establish Anthropic’s internal serving constraints or financial position.

Anthropic says its updated tokenizer can map the same input to roughly 1.0–1.35 times as many tokens depending on the content, and that higher effort can produce more output tokens in agentic settings. It also reports improved overall token usage on an internal coding evaluation and recommends measuring real traffic. A higher input token count alone does not determine total task cost or efficiency.

The system card says Opus 4.7 does not advance Anthropic’s capability frontier relative to Mythos Preview on the relevant evaluations. I reduce this to a possible release ‘bar,’ but the actual assessment includes multiple risks and mitigations, not simply one numerical benchmark threshold.

The card says its automated AI R&D threat model does not apply to Opus 4.7 under Anthropic’s current assessment. I question the conclusion because AI tools can already assist research and generate data. That assistance is not the same as the card’s defined threshold: fully automating or dramatically accelerating large top-tier research teams in sensitive domains. The 4.7 finding alone does not establish that Mythos crossed that threshold.

I discuss Anthropic’s model-welfare research and its examination of models’ reported circumstances and internal representations. I find precaution worth considering, but those measurements and a model’s statements do not establish subjective experience or consciousness. Describing the work as treating a model as certainly conscious would overstate what the evidence shows.

The card describes one autonomy threat model as applicable and a second, concerning highly accelerated or automated research, as not applicable to Opus 4.7 under its assessment. It also discusses model responses to questions about ending conversations. These are research and deployment evaluations with uncertainty, not a conclusive answer to model welfare.

Opus 4.7 looks promising to me, but I have not completed the practical testing discussed in the video. I plan to try it over the following days and compare its behavior in real workflows rather than relying only on launch charts.