Code with Claude SF 2026: The Day Anthropic Declared Platform Intent

Table of Contents

The number Dario Amodei opened with was not a product feature. It was a statement of position.

Anthropic had projected 10x revenue growth in Q1 2026. They got 80x. Annualized revenue run rate, as of April, sits at $30 billion — up from $87 million in January 2024. API volume is up 17x year-on-year. Claude Code, launched publicly in mid-2025, hit $1 billion in annualized revenue faster than any developer tool in history and now represents more than half of Anthropic’s enterprise revenue.

These are not incremental metrics. They are the backdrop for understanding why the six announcements at Code with Claude SF were not a product roadmap — they were a strategy declaration.

The Infrastructure Bet That Unlocks Everything Else
#

Every launch at the event was downstream of one deal: Anthropic signed an agreement with SpaceX to use all available capacity at the Colossus 1 data center in Memphis — more than 300 megawatts, over 220,000 NVIDIA GPUs, coming online within the month. The facility was originally built by xAI.

This is the third major compute commitment Anthropic has announced in 2026, joining agreements with Amazon (up to 5 GW, 1 GW online by end of year) and Google/Broadcom (5 GW, coming online in 2027). The SpaceX deal fills the near-term gap — it is capacity that is available now, not in 12 to 18 months.

The practical effect landed immediately. Effective May 6, Claude Code’s five-hour rate limits were doubled across all paid plans: Pro, Max, Team, and seat-based Enterprise. Peak-hour throttling for Pro and Max accounts was removed entirely. Opus API rate limits were raised “considerably.” For Claude Code’s heaviest users — teams running multi-agent workflows against large codebases — these changes are not incremental. The constraints that force you to structure work around rate windows are gone.

The orbital compute angle is not a joke. Anthropic confirmed it is exploring developing multiple gigawatts of compute capacity in space with SpaceX. Whether that happens or not, the directional bet is clear: Anthropic is building infrastructure at a scale that makes near-term capacity constraints structurally impossible, not just unlikely.

Code Review GA: The Product Launch of the Event
#

If the SpaceX deal was the infrastructure headline, Code Review GA was the product headline — and the one with the clearest near-term impact for engineering teams.

Code Review dispatches a team of agents to review a pull request in parallel. The architecture has three phases:

Bug-finding agents work in parallel, scanning the PR for logic errors, security issues, and implementation mistakes
Verification agents check each candidate finding — filtering false positives before they reach the developer
Ranking surfaces issues by severity, so the output is prioritized rather than a flat list

The result lands on the PR as two outputs: a single high-signal overview comment summarizing what was found, and in-line comments on specific lines. The intent is one PR comment worth reading, not 47 marginal notes.

What it costs
#

Each review averages $15–25, scaling with PR size and codebase complexity. Large PRs (over 1,000 lines changed) get findings 84% of the time, averaging 7.5 issues. Small PRs (under 50 lines) get findings 31% of the time, averaging 0.5 issues. Reviews take approximately 20 minutes.

This is billed separately via Anthropic’s “extra usage” mechanism — it does not count against your plan’s included usage, and it does not replace /ultrareview (which remains available to Pro and Max users, with three free reviews included). Code Review is available on Team and Enterprise subscriptions.

The “every Anthropic team uses it” claim
#

Anthropic said Code Review is used by every internal engineering team. That claim matters not as a credibility signal but as a calibration: the tool was built under real production conditions before GA, against real codebases at real velocity. The false-positive filter in the verification step reflects what happens when you run a review tool against ten thousand PRs and learn what wastes engineers’ time.

Managed Agents: Three Features, One Coherent System
#

The Managed Agents updates (covered in depth here) shipped as three coordinated features:

Dreaming — a scheduled background process that reviews past sessions, synthesizes patterns, and curates the memory store so agents improve over time without model retraining. Harvey, the legal AI company, saw task completion rates increase 6× after deploying it.

Outcomes — a rubric-based evaluation layer. You define what success looks like before the agent starts; a separate grader (in its own context window) evaluates the output and sends the agent back for another pass if criteria are not met. Webhook notification fires on completion.

Multiagent Orchestration — a coordinator/specialist model. A lead agent decomposes a task and delegates to up to 20 specialist agents running in parallel on a shared filesystem. Each specialist can have its own model, prompt, and tool set.

These three features close the three main gaps between “capable prototype” and “production autonomous system”: agents that learn from history, agents that can self-verify output quality, and agents that can scale horizontally without architectural changes.

The Mercado Libre Number
#

The enterprise adoption data point Anthropic chose to lead with was Mercado Libre. The Latin American e-commerce and fintech platform — with 23,000 engineers — is targeting 90% autonomous coding by Q3 2026.

Ninety percent autonomous is not a target about replacing engineers. It is a target about what proportion of code shipped per quarter is primarily AI-generated, with humans in the review and direction role rather than the implementation role. For a 23,000-person engineering organization, the scale of the organizational transformation implied by that number is hard to overstate.

Mercado Libre is not a software company in the traditional sense — they are the infrastructure for digital commerce and payments across Latin America. If they hit 90%, the story becomes: agentic coding has crossed into industries where software is critical infrastructure, not just tech companies eating the world from the outside.

Anthropic also mentioned Shopify and other enterprise customers in the same tier, but did not publish specifics.

What Code with Claude SF Actually Was
#

The mistake in reading this event as a product launch is that the individual announcements — while significant — are not the point individually. The point is the architecture they describe together.

Compute (SpaceX deal) → removes the capacity ceiling
Rate limits → translates compute into user capacity
Code Review GA → production-grade autonomous code review, billed per-use
Managed Agents → production-grade autonomous agents that improve over time
80x growth → reveals the demand already exists at this scale
Mercado Libre → demonstrates what the demand looks like at the enterprise end

These are not six separate features. They are one system: a vertically integrated platform for autonomous software development, from the GPUs at the bottom to the PR review comment at the top.

Cursor’s response to this event was essentially silence. OpenAI’s Code Review equivalent (Codex Code Review) is still in limited beta with manual trigger only. GitHub Copilot’s equivalent is a rules-based heuristic layer, not a multi-agent verification loop.

The window where IDE-first AI tools could argue parity with terminal-native agentic platforms is closing. The Code with Claude SF announcements did not close it — they made it visible.

The next event is Code with Claude London on May 19–20, the same day as Google I/O 2026. The scheduling was not accidental. Anthropic will be in direct air competition with whatever Gemini 4 announcement Google makes that day. It will be worth watching both.

Sources:

The Infrastructure Bet That Unlocks Everything Else#

Code Review GA: The Product Launch of the Event#

What it costs#

The “every Anthropic team uses it” claim#

Managed Agents: Three Features, One Coherent System#

The Mercado Libre Number#

What Code with Claude SF Actually Was#

Related