Code with Claude SF Recap: SpaceX Compute, Doubled Limits, and Claude Agents That 'Dream'

Anthropic’s Code with Claude conference ran in San Francisco today, May 6. It was the first leg of the expanded 2026 tour, with London on May 19 and Tokyo on June 10 still to come. Despite weeks of speculation about a Jupiter-v1-p model reveal, no new model was announced. The keynote was about compute, agents, and shipping Claude Code at scale.

The Compute Story: SpaceX, Colossus 1, and Doubled Limits

The day-one announcement was a deal with SpaceX for access to all of the compute capacity at xAI’s Colossus 1 data center in Memphis. That brings Anthropic over 300 megawatts of new capacity, with more than 220,000 NVIDIA GPUs (a mix of H100, H200, and GB200), online within the month.

What that means for paying users, effective today:

Claude Code’s five-hour rate limits doubled for Pro, Max, Team, and seat-based Enterprise plans.
Peak-hour reductions removed for Claude Code on Pro and Max.
Opus API rate limits raised across the platform.

Anthropic also said it had “expressed interest” in working with SpaceX on multi-gigawatt compute capacity in space, which generated more headlines than the actual MOUs deserve. (Anthropic news post, CNBC coverage, Engadget).

The framing in the keynote, led by Chief Product Officer Ami Vora, was that API volume is up 17x year-on-year, and the Colossus deal is one of several capacity moves alongside existing Amazon (5 GW), Google/Broadcom (5 GW), Microsoft/NVIDIA ($30B Azure), and Fluidstack ($50B) commitments.

Claude Managed Agents: Orchestration, Outcomes, Dreaming

Dianne Penn (Head of Product, Research) and Angela Jiang (Head of Product, Claude Platform) introduced three additions to Claude Managed Agents:

Multi-agent orchestration. Coordinate fleets of specialized agents on a single task, with explicit delegation and aggregation patterns. The keynote demo used a hypothetical lunar drone product with Commander, Detector, and Navigator agents working in parallel.
Outcomes. Define rubric-style success criteria and let Claude iterate against them until the task is graded as done. It is the same idea as a test loop, applied to fuzzier deliverables.
Dreaming (research preview). Agents can be configured to review their own past sessions on a schedule, extract patterns, fix recurring failure modes, and reorganize their memory store. Reviews can run automatically or be gated for manual approval. (Testing Catalog writeup, InfoWorld)

Penn’s framing: high-quality memory plus multi-agent coordination is how you get “context windows that feel infinite” without actually growing the context window.

There was also an “advisor strategy” pitch from Katelyn Lesse (Head of Engineering, Claude Platform) and Jiang: have Sonnet call Opus only for the hard parts, and they claim “frontier model quality at 5x lower cost” on internal benchmarks.

Claude Code: Three Surfaces, One SDK, and Routines

Cat Wu (Head of Product, Claude Code) and Boris Cherny (Head of Claude Code) covered the Claude Code track. The frame is that Claude Code is now three surfaces, all built on the Claude Agent SDK:

CLI (the original).
IDE extension (VS Code and JetBrains).
Desktop app (the multi-session redesign that shipped in April).

Features called out as already in production at Anthropic and shipping more broadly:

Code Review for PRs.
Remote Agents running in the cloud, not your laptop.
CI auto-fix that opens a corrective PR when a build breaks.
Security Reviews automated against a configurable policy.
Routines, scheduled agent runs that handle work like dependency bumps, doc regeneration, or release notes overnight.

Wu cited Mercado Libre, where roughly 23,000 engineers are targeting 90% autonomous coding by Q3. That number deserves the usual caveats: “autonomous” is doing a lot of work, and customer-led case studies on a vendor stage are not third-party evidence. Still, the direction matches what Netflix’s session covered next, on moving engineers up a “Claude Code maturity ladder” from chat-style use to delegated background work.

Claude Design

A smaller announcement that landed well: Claude Design, a tool from Anthropic Labs aimed at visual and product design workflows. The keynote attributed it to Opus 4.7’s reportedly strong taste for visual design. It is the second Anthropic Labs project to graduate this year after Claude Mythos.

What Was Not Announced

No Sonnet 4.8. The March npm source map leak and the Jupiter-v1-p red teaming reports had primed the room for a model reveal. None came. Vora opened by saying the day was about “getting more from the models you already have,” which read as a deliberate reset.
No pricing changes. The doubled limits are a quiet upgrade inside existing tiers.
No on-prem story for managed agents. Still cloud only.

Watch the Sessions

The opening keynote and main-stage talks were livestreamed. Breakout sessions are being posted on demand by the end of the week at claude.com/code-with-claude. Simon Willison’s live blog is the best running narrative if you want a session-by-session feel for the day.

The extended SF event runs tomorrow, May 7, focused on independent developers and early-stage founders.

Sources: Anthropic: Higher limits and SpaceX deal, CNBC: Anthropic, SpaceX announce compute deal, Engadget: Anthropic doubling Claude Code rate limits, Simon Willison live blog, Testing Catalog: Dreaming for Claude Managed Agents, InfoWorld: Anthropic rolls out Claude Managed Agents, Code with Claude SF session list