🧭 Code with Claude 2026: Checkpoints, Auto Mode, and a Deliberate No-New-Model Keynote

Anthropic's second annual developer conference, Code with Claude 2026, opened in San Francisco on May 6. Unlike many developer conferences, the keynote did not launch a new model — a deliberate choice the company framed as a signal that the focus is now on getting existing models reliably into production rather than chasing benchmark headlines. The two centerpiece announcements were Claude Code Checkpoints and Claude Code Auto Mode, both shipping same-day.

Checkpoints: save state, rewind with confidence

Checkpoints automatically save your code state before every change Claude makes. If a multi-step session takes a wrong turn, you can rewind to any prior state using Esc Esc (double-tap Escape) or the /rewind command, then choose to restore the code, the conversation, or both. The feature was one of the most-requested by Claude Code users, who had previously relied on git commits as manual checkpoints — a workflow that required constant context-switching. Checkpoints operate below the git layer, so you can rewind to mid-session states that were never committed.

# During a Claude Code session:
/rewind               # opens checkpoint browser
                      # choose: restore code only, conversation only, or both
# Or press Esc twice to jump to the most recent checkpoint

Auto Mode: safer autonomous execution

Auto Mode is a new permission model for Claude Code that sits between fully interactive (every action requires approval) and --dangerously-skip-permissions (no approvals at all). In Auto Mode, a secondary model-based classifier evaluates each proposed action and decides whether it requires human confirmation. The classifier uses session context — accumulated denials, action type, file sensitivity — to calibrate its decisions. A backstop rule kicks in if the session accumulates 3 consecutive denials or 20 total, halting the agent and escalating to the human. This makes it safe to leave Claude Code running on long tasks without babysitting every decision.

Conference structure: three cities, extended days

The 2026 edition spans three stops: San Francisco (May 6), London (May 19), and Tokyo (June 10). Each main conference day is followed by an "Extended" day aimed at independent developers and early-stage founders, with a focus on hands-on workshops rather than keynotes.

🧭 Anthropic Signs SpaceX Colossus 1 Deal — 220,000 GPUs and Doubled Pro/Max Rate Limits

On the same day as the conference keynote, Anthropic published a brief announcement: the company has signed a deal to use the entire compute capacity of SpaceX's Colossus 1 data center in Memphis, Tennessee. The deal gives Anthropic access to more than 300 megawatts of capacity and over 220,000 NVIDIA GPUs within the month. SpaceX and xAI merged earlier in 2026; xAI retains its larger Colossus 2 cluster while Anthropic takes Colossus 1. The compute is being added to meet demand that Anthropic's CEO Dario Amodei described as "80x growth" in a single quarter.

Immediate effect: rate limits doubled

• Pro and Max plans — rate limits doubled; peak-hour reductions removed. Pro and Max subscribers should see noticeably improved throughput during US daytime hours starting this week.
• Team and Enterprise plans — seat-based rate limits doubled.
• API (Opus tier) — Opus token limits raised.

Space compute interest

As part of the agreement, Anthropic "expressed interest" in working with SpaceX to develop multiple gigawatts of compute capacity in orbit — satellite-based inference infrastructure that would reduce terrestrial data center land and power constraints. This remains exploratory and has no committed timeline, but it is the most concrete public statement either company has made about space-based AI compute.