Anthropic just shipped a weaker model on purpose while sitting on something stronger.
The Summary
- Anthropic released Claude Opus 4.7, its most powerful "generally available" model, with improvements in advanced software engineering, image analysis, and document creation
- The release comes barely a week after Anthropic limited access to Mythos Preview, a more powerful cybersecurity-focused model the company says is "too dangerous" for general release
- The delta between what Anthropic can build and what it will ship is now a deliberate product strategy, not a bug
The Signal
Anthropic positioned Opus 4.7 as a step up from Opus 4.6 for "complex coding areas that in the past required more hand-holding." The model handles advanced software engineering tasks with less intervention, analyzes images better, and shows more creativity when generating slides and documents. Standard capability bump stuff.
But the real story is the model Anthropic isn't releasing. Mythos Preview, announced earlier this month, is Anthropic's most powerful model overall, with a specific focus on cybersecurity capabilities. The company has restricted access to it deliberately. Bloomberg reports Anthropic considers Mythos "too dangerous to be released to the general public."
"Anthropic is shipping a weaker model while competitors race for raw power."
This is a fork in the road for the agent economy. Every other major lab is sprinting toward "most capable model wins." Anthropic is running a different play: tiered releases based on risk assessment, not just capability benchmarks. They're building frontier models, then deciding which ones get API access.
The calculus matters because cybersecurity capabilities cut both ways. A model that can identify and patch vulnerabilities at scale can also identify and exploit them. If Mythos can reason through complex security architectures, it can also reason through how to compromise them. Anthropic looked at that capability profile and said "limited preview only."
Key distinctions emerging:
- Mythos: Most powerful overall, cybersecurity-focused, restricted access
- Opus 4.7: Most powerful for general availability, software engineering focus, open API
- The gap between them: A conscious policy decision, not a technical limitation
This creates an interesting market dynamic. Developers building agent systems for software engineering get Opus 4.7 through standard API channels. Security researchers who want Mythos-level reasoning have to go through whatever vetting process Anthropic set up for preview access. That's a capability moat with a velvet rope.
The Implication
If you're building agents that need advanced reasoning for coding, document generation, or image analysis, Opus 4.7 just raised the ceiling on what's possible through a standard API. But if your use case touches security research, penetration testing, or vulnerability analysis, you're now in a different tier of access entirely.
Watch how other labs respond. Does OpenAI start talking about o4 variants they won't release? Does Google tier Gemini access by risk profile? Anthropic just made "we built it but won't ship it" a viable competitive position. That only works if investors and customers believe the restricted model is actually more powerful, not just more marketing. Mythos Preview is the test case.