At Anthropic's Code with Claude 2026 conference last week, the company doubled usage limits on Pro and Max plans and announced a SpaceX compute deal. In an interview with Ars Technica, Claude Code product lead Cat Wu put a number on what was happening behind the scenes: the team planned for 10x year-over-year growth and got 80x. That gap โ Dario Amodei's words from the keynote โ is the root of the compute crunch builders have been hitting all year, and it is what the new "programmatic credit pool" announced for June 15 was designed to manage. Wu also confirms the product naming: alongside Claude Code there is now Claude Cowork and Managed Agents as distinct surfaces โ which lines up with the PwC alliance piece this week that mentioned Cowork without explanation.
The product philosophy Wu describes is "no grand plan," week-long iteration cycles, and what Anthropic calls the lean harness. Concretely: the team defaults to NOT shipping opinionated tools unless they clearly improve token efficiency or accuracy on internal evals. LSP plugins that semantically navigate the codebase are available as extensions, but Anthropic does not ship them in the default Claude Code surface because their own evals do not show a measurable performance gain from doing so. Wu invokes Richard Sutton's "Bitter Lesson" essay as a guiding principle โ bet on general-purpose methods that scale with compute rather than baking domain-specific structures into the agent. That is the architectural reason Claude Code looks lighter than some competitors that have built more semantic-aware scaffolding around the same models.
The growth-shape commentary is the other useful detail. Wu confirms that the center of gravity is still the CLI for power users, but the team is gradually shifting to desktop because the multi-agent multi-tab workflow has overrun what a terminal can comfortably display. The Wild West framing โ short feedback loops, productize-from-conviction in roughly a week, dogfood internally before betting on a direction โ explains why the surface count keeps expanding (CLI, IDE, desktop, mobile via ChatGPT integration on the Codex side, web). The "Claude anticipates what you want" next-step Wu describes โ monitoring Slack and GitHub for feedback on the features you ship โ is the kind of capability that creates a new compute draw, which loops back to the rate-limit question.
For builders, the practical takeaway is twofold. First, the lean-harness bet means that if you want semantic codebase navigation, fine-grained LSP integration, or domain-specific scaffolding, you wire it in yourself rather than waiting for Anthropic to ship it as a default โ and you take the token cost yourself. Second, the 80x-versus-10x context says that pricing and quota changes are demand management responses, not a product strategy shift; the new programmatic credit pool is part of the same engineering response as the SpaceX compute deal. The honest signal Wu gives is that nobody at Anthropic claims to know what the right shape of Claude Code is in 18 months. They are iterating on weekly cycles and assuming the models will keep moving fast enough that any locked-in product plan would be obsolete by the time it shipped.