Anthropic released its latest flagship model Thursday. By every public benchmark, it is a meaningful step up from Opus 4.6. By Anthropic's own account, it is not the best model the company has built.
The company disclosed in its system card — the technical document that accompanies a model release — that Claude Opus 4.7 "does not advance the company's capability frontier," Anthropic reported. A separate model, Claude Mythos Preview, outperformed Opus 4.7 on every relevant evaluation, The Verge reported. Mythos Preview is not available to developers. It runs only for five partners: Nvidia, JPMorgan Chase, Google, Apple, and Microsoft.
The disclosure is a rare move. AI labs routinely release updates that leave room for future improvements. Anthropic instead drew explicit attention to the gap within its own product line, using the system card to acknowledge publicly what the benchmark data shows privately.
The result is a two-tier access structure the company has made impossible to miss. Developers building on the Claude API can now access a model Anthropic itself describes as short of the frontier, while a small group of partners operate with a demonstrably superior capability.
The gap is not cosmetic. Mythos Preview is described as Anthropic's most powerful model. The five partners — a semiconductor firm, a major bank, two of the world's largest technology companies, and a consumer electronics giant — represent a cross-section of the infrastructure that the AI industry depends on. What those partners do with a model that developers cannot access is not public.
Opus 4.7 brings real improvements. The model resolves 13 percent more tasks on Anthropic's internal 93-task coding benchmark than Opus 4.6, and three times as many production tasks on Rakuten's SWE-Bench evaluation, Anthropic reported. Computer vision accuracy for autonomous agents jumped to 98.5 percent from 54.5 percent in the prior version — a capability that matters directly for cybersecurity tools that use AI to navigate and inspect systems. The model can now process images up to 2,576 pixels on the long edge, more than three times the resolution of earlier Claude models. Pricing remained unchanged at $5 per million input tokens and $25 per million output tokens. The model is available through the Claude API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry.
Anthropic also launched a Cyber Verification Program, allowing vetted security researchers to access Opus 4.7 for legitimate cybersecurity work including vulnerability research and penetration testing, Anthropic reported. The company said findings from real-world use would inform its approach to broader releases of Mythos-class models. The framing treats the more powerful model as a future step, pending what Anthropic learns from watching Opus 4.7 operate at scale.
The design tool The Information reported Anthropic planned to release alongside Opus 4.7 this week did not appear. Anthropic declined to confirm whether the tool had been delayed or whether the report was inaccurate.
The tiered access structure Anthropic has constructed raises a question the company has not answered directly: if the most capable model is gated to a fixed list of partners, what does it mean to build a business on the public API? The answer, so far, is the model it released Thursday.
