Most attention at Code with Claude went to the SpaceX compute deal and Claude Code rate limit doubling. But one slide may carry more information than all the hardware news combined.
Anthropic revealed three priority directions for their next-generation models. Not new features, not price cuts, not API updates — where the model itself is evolving.
Three directions
1. Higher judgment and code taste
This phrasing is interesting. Anthropic didn't say "stronger reasoning" or "higher accuracy" — they used the word "taste."
In coding, "taste" means: not just writing code that runs, but writing code that's maintainable, readable, and follows engineering best practices. This is actually the same thing the SWE Atlas refactoring benchmark measures — code quality isn't just functional correctness.
2. "Infinite" context windows
Anthropic used quotes, which means they know literal "infinity" is physically impossible. But the direction is clear: make context windows stop being a usage bottleneck.
Claude's context window is already large (200K tokens), but developers still hit truncation and information loss in long document processing, codebase understanding, and multi-session memory scenarios. The essence of "infinite" context is making the model stop making mistakes because it can't remember.
Technically, this might be achieved through sparse attention, layered memory, or linear attention approaches similar to FlashQLA. Anthropic didn't give a timeline, but putting it on the roadmap means technical feasibility has been internally validated.
3. Multi-agent coordination
This direction pairs directly with the new Claude Managed Agents features announced the same day (Dreaming, Outcomes, Webhooks, Orchestration).
Multi-agent isn't a new concept, but native model-level support for multi-agent coordination means: the model itself can understand and handle task distribution, information sharing, and conflict resolution between multiple agents — rather than relying entirely on upper-layer frameworks for orchestration.
Why this roadmap matters
Anthropic's model release cadence has clearly accelerated (the gap between Opus 4.7 and Sonnet 4.8 is shrinking). But rather than "when will what model ship," "what will the next model look like" tells more about the company's technical judgment.
Of the three directions, "code taste" is the most pragmatic — it directly targets the core competitiveness of coding agents. "Infinite context" is the most ambitious — if actually achieved, it would redefine long-text task workflows. "Multi-agent coordination" is the most forward-looking — it's a bet on the future shape of the agent ecosystem.
Whether all three can be pursued simultaneously, or whether they'll compete for resources, is the most interesting thing to watch over the next few quarters.
Sources:
- Anthropic Code with Claude conference official livestream
- Community meeting notes and discussion threads