Anthropic's Claude Keynote: A New Era for

There's a particular kind of corporate keynote that tells you everything by what it chooses to celebrate. At Anthropic's first-ever Code with Claude event outside San Francisco—held in London earlier today—the opening speaker, Boris, didn't lead with benchmark numbers or funding rounds. He started with a story about programming a TI-83 calculator at age 13 to pass math tests.

That's a deliberate choice. The message it signals: we're not selling you software; we're selling you back a feeling you used to have.

Whether that feeling survives contact with production infrastructure is a more complicated question.

The Exponential, The Linear, and the Gap in Between

The most analytically honest moment in the keynote wasn't in the demos. It was this observation from Boris: "Even though model capabilities are improving on an exponential, most organizations are still adopting AI on a linear path. That means there's a growing gap between what AI can do and what it's actually doing for people."

That's a real tension, and Anthropic doesn't hide from it. API volume on the Claude platform is up nearly 17x year-over-year. The average Claude Code developer is now spending over 20 hours a week running the model. Those are big numbers—but they're also Anthropic's numbers, self-reported at a promotional event. The 17x figure tells you adoption is accelerating; it doesn't tell you how much of that translates into shipped products that non-developers actually use.

What Anthropic is arguing, essentially, is that the bottleneck is no longer the model. It's the scaffolding, the deployment infrastructure, the evaluation pipelines—all the unglamorous work that sits between "Claude can do this" and "our users can do this." The whole London keynote was structured around that argument.

Eight Models in Twelve Months

Lisa from the research PM team offered the clearest picture of where Claude has come. She's been involved in all 17 versions of Claude since joining Anthropic in 2023. The trajectory she described is worth sitting with:

Opus 3 (two years ago): first model good at writing long-form code. Sonnet 3.5/3.6: first safe computer use. Sonnet 3.7: first model that thought before answering. Opus 4 (last year's headline): turned out to be unexpectedly good at Excel and PowerPoint. Opus 4.7 and Mythos preview (current): models that, per Lisa, "can own outcomes end to end and apply judgment to complete tasks with high ambiguity."

Eight frontier models in twelve months. That's a pace that would make most OSS maintainers quietly spiral.

The headline capability claim for Mythos is striking: last month, the model reportedly read the entire OpenBSD source tree and found a 27-year-old vulnerability that had survived every human reviewer, fuzzer, and static analyzer thrown at it for nearly three decades. If that holds up to independent verification, it's genuinely significant—not just as a party trick, but as a signal about what autonomous security research might look like. The autonomous PR handling capabilities Anthropic has been quietly rolling out start to look less like features and more like scaffolding for something larger.

Lisa introduced a metric called "task horizon"—how long can a model work before losing the thread? A year ago: minutes. Today: hours. Anthropic's stated expectation for future Claude generations: continuous. "Agents that are proactive, always on, that know what to do without being told."

That's a significant claim, and it deserves more scrutiny than a keynote allows. "Always on" agents that "know what to do without being told" sound elegant in a slide deck. In practice, they raise questions about failure modes, auditability, and who's accountable when an autonomous agent does something its operators didn't intend.

The Scaffolding Problem (and What to Do About It)

One of the more practically useful observations from Lisa: as models get smarter, scaffolding that used to help can hold Claude back. The complex orchestration layers, the elaborate prompt chains, the careful hand-holding—all of it was a workaround for model limitations. Better models need less of it. "More intelligent models can often get further with generalized primitives like a file system and sandbox computing environment."

This matters for anyone who's built on Claude over the past two years. Architecture decisions made for Claude 3 may be actively constraining what Claude 4.7 can do. The developers who win, per Lisa, are "the ones whose architecture is ready to absorb the next big jump."

That's honest advice, and it has teeth. It means that if you've spent months building elaborate prompt scaffolding, a model upgrade might render half of it obsolete. The implicit ask is to build loosely enough that you can hand Claude more rope when the model can handle it. The simplification trend in Claude Code that's been observable across recent releases reflects exactly this—less framework, more trust in the model itself.

The Cost Problem and the Advisor Strategy

Angela and Caitlyn, presenting the platform layer, named the tension plainly: businesses need frontier-level intelligence, but at lower cost. Their proposed solution—the "advisor strategy"—is architecturally interesting. You run a smaller, cheaper model (Haiku or Sonnet class) for execution, and that model can reach out to a larger model (Opus) when it needs help.

The results they cited: Eve Legal got "frontier model quality at five times lower cost." Sonnet running with Opus as advisor outperformed Sonnet alone—and came in cheaper, because Opus advised it to work more efficiently.

The mechanism is clever. The open question is how generalizable it is. Legal document workflows—structured, high-stakes, relatively well-defined—seem like a natural fit. Whether the same cost curve holds for messier, more ambiguous agentic tasks is an empirical question Anthropic is inviting developers to answer for themselves.

Managed Agents: Production in Days, Not Months

The other major platform announcement was Claude Managed Agents, described as "an agentic harness paired with production-grade infrastructure" that Anthropic claims can cut time-to-production for a new agent from months to days. Two new features landed at the event: self-hosted sandboxes (Claude can now execute work on your own server rather than Anthropic's) and MCP tunnels (secure access to internal MCP servers without public internet exposure).

The Asana example was instructive: they built AI teammates that let humans delegate tasks directly within Asana projects. The goal was speed and scale simultaneously—the exact combination that Angela noted is notoriously hard to hit when you're building from scratch. Whether managed agents become the default deployment path for serious Claude integrations is a live question, but the direction is clear.

The self-hosted sandbox announcement is particularly worth watching. It signals that Anthropic understands enterprise data governance concerns aren't just compliance theater—they're real blockers. Giving teams control over where code executes is a prerequisite for adoption in regulated industries.

The Part That Doesn't Fit the Narrative

Buried in Lisa's otherwise optimistic model overview was this: "Models are still imperfect. Claude absolutely still has verbal ticks, can be stumped by viral common sense questions, and sometimes does more than you asked for."

"Does more than you asked for" is polite language for a class of failure that becomes much more consequential as agents gain longer task horizons and higher autonomy. When a model that's supposed to keep your project on track this week decides on its own that "on track" means something different than you meant—and acts on that interpretation for hours—the blast radius is not the same as a chatbot that gives a weird answer.

This isn't a reason to dismiss what Anthropic is building. The Binty example—where Claude API integration took 20 days off the process of licensing a foster family—is a reminder that the stakes cut both ways. Efficiency gains in human services translate into real outcomes for real people. That's worth taking seriously.

But so is the question of what it means to hand progressively more judgment to a system that still, by Anthropic's own account, sometimes misreads the assignment.

The developers in that London room are the ones who'll make those calls. Anthropic can build the model. They can build the platform. The part where capability meets consequence—that's still a human problem.

Dev Kapoor covers open source software and developer communities for Buzzrag.