Two pieces of data landed on the same day. One is a product release. One is a research finding. Together, they say something neither says alone.
Anthropic released Claude Opus 4.7 to general availability on April 22, 2026. In the same announcement, the company disclosed that a more powerful internal system, Claude Mythos Preview – exists above Opus 4.7 in their internal capability hierarchy. On the same day, Epoch AI published updated analysis showing that frontier AI capability growth has nearly doubled in pace since early 2024. The two data points are not coincidental noise. They’re a pattern.
What Claude Opus 4.7 Actually Is
Start with what’s confirmed. Claude Opus 4.7 is generally available via the Claude API. It supports a 1M token context window and 128K maximum output tokens, confirmed via platform.claude.com’s release notes. Adaptive thinking and enhanced visual analysis for document workflows are among the listed capabilities. GitHub Copilot Pro+, Business, and Enterprise users can select it in the model picker, per GitHub’s changelog.
On benchmarks: Anthropic reports a score of 46.9% on Humanity’s Last Exam (HLE) without tool access. The company characterizes this as the highest score among current frontier models on this benchmark variant. That characterization requires a specific qualifier: HLE results differ meaningfully depending on whether models use tool access during evaluation. Comparing a without-tools score to a with-tools score produces a misleading picture. MiniMax M2.7 is reported at 44.9% on HLE with tools, a number that cannot be placed on the same row of a comparison table as Opus 4.7’s without-tools result without a header explaining the difference.
Anthropic’s technical report accompanies the release at arXiv:2604.09881. This is a vendor-authored document, not an independent evaluation. Independent verification of the 46.9% HLE figure through Epoch AI’s benchmark tracking is pending URL confirmation; readers should treat the benchmark claim as Anthropic-reported until that confirmation is available.
The Benchmark Question: With Tools or Without?
Practitioners evaluating frontier models are increasingly encountering a problem that today’s release makes visible: benchmark scores are often compared across incompatible conditions.
HLE is a rigorous exam-style benchmark designed to test models at the limits of human expert knowledge. It’s one of the more credible evaluation frameworks for frontier capability assessment precisely because it resists the kind of saturation that has invalidated earlier benchmarks. But the distinction between tool-assisted and tool-free performance is not a minor footnote, it’s a fundamental difference in what’s being measured.
A model that can call a calculator, a search API, or a code interpreter during an HLE evaluation is being tested on a different capability than a model working from parameters alone. Both results are useful. They answer different questions. A without-tools score tells you something about the model’s internalized reasoning. A with-tools score tells you something about the model as part of an agentic workflow.
Teams building applications that invoke Claude Opus 4.7 as one node in a multi-agent system should care most about with-tools performance. Teams evaluating the model’s raw reasoning quality for use cases without tool access should care about the without-tools figure. Both audiences deserve scores framed in their relevant context, not a single headline number.
The hub’s coverage of AI benchmarks will return to this question. It’s not a Claude-specific issue. It’s a structural problem with how frontier model benchmarks are communicated to practitioner audiences.
Anthropic’s Safety Tiering Architecture
The Claude Mythos Preview disclosure is editorially significant in a way that’s easy to underappreciate. Anthropic didn’t have to say this publicly.
According to Anthropic’s own disclosure, corroborated by multiple independent reports, Claude Opus 4.7 is intentionally “less broadly capable” than Claude Mythos Preview, and Opus 4.7’s cyber capabilities are explicitly described as less advanced than those of the internal model. Anthropic chose to characterize the gap and explain why it exists as a feature of their safety approach, not a limitation to apologize for.
This is a meaningful signal about how Anthropic thinks about the frontier labs problem: the most capable AI you can build is not necessarily the most capable AI you should release. The public release is a deliberate decision point, not the natural terminus of the development cycle. What gets deployed broadly is a safety governance choice, shaped by evaluation of risks that don’t always appear in benchmark scores.
For practitioners, this creates a useful mental model. When evaluating any frontier lab’s “most capable public model,” the relevant question is not just “how capable is this?” but “how does this compare to what the lab has internally, and what does the gap reflect?” Anthropic is one of the few labs making that gap visible explicitly. Most labs are not. The ones that aren’t are not necessarily making different choices, they may just be less transparent about the architecture behind those choices.
The prior TJS coverage of Anthropic’s AWS compute partnership and hyperscaler infrastructure dynamics provides context for the infrastructure capacity underpinning these releases. A lab running at AWS scale has different release-decision calculus than a smaller organization constrained by infrastructure.
The Epoch AI Context: What the Breakpoint Data Means for Release Decisions
This is where the two data points converge.
According to Epoch AI’s updated analysis, frontier model improvement accelerated sharply in early 2024. The rate of progress on Epoch AI’s composite Epoch Capability Index (ECI) moved from approximately 8.2 points per year to approximately 15.5 points per year, a near-doubling. Epoch AI characterizes this as a piecewise linear inflection, not gradual drift.
Two clarifications before reading this number against the Claude release. First, ECI is a composite index proprietary to Epoch AI, it reflects their methodology for aggregating capability signals across frontier models. It’s not a benchmark score on a specific task. Second, the acceleration attribution, reinforcement learning on reasoning tasks and compute investment, is Epoch AI’s analytical framing, and should be read as a credible hypothesis from a respected research organization rather than an independently validated conclusion.
With those qualifications in place: if frontier capability is growing at twice the pace it was two years ago, then a lab’s internal capability in 2026 may be substantially ahead of what it was when its safety and deployment protocols were last fully updated. The gap between Claude Mythos Preview and Claude Opus 4.7 might represent, among other things, a safety evaluation process that hasn’t yet cleared the most recent internal capability tier. That’s a reasonable inference, not a confirmed fact. But it’s consistent with what Anthropic disclosed.
For compliance teams: the compute-based thresholds in frameworks like the EU AI Act were calibrated against a capability growth environment that Epoch AI’s data suggests has meaningfully changed. If the same amount of compute produces substantially more capable models than it did when thresholds were set, then FLOP-based classification boundaries may misclassify current systems. See the Regulation pillar for EU AI Act threshold coverage. This is a live issue.
Practical Implications: Who Should Evaluate Opus 4.7 and for What
For developers building against the Claude API: the 1M token context window is the most operationally significant confirmed capability. If your application involves large document processing, multi-file code analysis, or long-context reasoning, the threshold is genuinely useful. The GitHub Copilot integration means enterprise developer teams on existing Copilot contracts can evaluate it without a separate procurement step.
For enterprise technical decision-makers: the safety tiering disclosure is relevant to your evaluation framework, not just the capability metrics. A lab that publicly maintains a gap between internal and released capability is making a different kind of institutional commitment than one that ships everything it can build as fast as it can build it. That’s a vendor risk signal worth incorporating.
For teams building AI governance frameworks: the Epoch AI data strengthens the case for shorter re-evaluation cycles. If frontier capability is compounding faster than your governance schedule assumed, model approvals and policy documents that are eighteen months old may be evaluating a different capability environment than today’s deployments operate in.
TJS Synthesis
Claude Opus 4.7 is a capable model. The confirmed details, 1M token context, GitHub Copilot integration, strong benchmark positioning for extended reasoning, give practitioners real tools to work with. But the more durable story from today’s data isn’t about Opus 4.7’s specific benchmark numbers.
It’s about what happens when accelerating capability growth meets deliberate release governance. The Epoch AI data shows the frontier is moving faster. The Claude Mythos disclosure shows that at least one lab is explicitly managing what reaches the public, and why. Those two facts together describe an AI landscape where the gap between what exists and what’s deployed is a purposeful design feature, one that will have increasing implications for how developers plan, how enterprises procure, and how regulators draw lines.
The question isn’t whether Opus 4.7 scores well on HLE. The question is what it means that today’s most capable public model is intentionally not the most capable model, in an environment where capability itself is accelerating. That’s the decision every practitioner, procurement team, and policy group is now navigating, whether or not they’ve framed it that way yet.