Code with Claude SF 2026: What Anthropic Actually Shipped

Code with Claude San Francisco ran on May 6, 2026, the first day of a three-city tour that continues to London on May 19 and Tokyo on June 10.¹ The headline news landed in writing the same morning: Claude Code’s five-hour rate limits were doubled across Pro, Max, Team, and seat-based Enterprise plans, the peak-hours reduction on Pro and Max was removed, and Anthropic announced a 300+ megawatt compute partnership with SpaceX.²

TL;DR:

• Claude Code five-hour rate limits doubled for Pro, Max, Team, and seat-based Enterprise plans, effective May 6.²
• Peak-hours throttling on Claude Code dropped for Pro and Max accounts; Opus API rate limits raised “considerably.”²
• SpaceX Colossus 1 deal gives Anthropic “more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month.”²
• Ten financial-services agent templates shipped as Cowork and Code plugins, plus a Managed Agents cookbook.³
• Microsoft 365 add-ins for Excel, PowerPoint, and Word landed; Outlook is queued.³
• Anthropic acquired Vercept to push computer-use capabilities further.⁴
• Claude Code v2.1.126 through v2.1.131 shipped between May 1 and May 6, including --plugin-url, claude project purge, and a working skillOverrides setting.⁵

What Shipped For Claude Code

The capacity announcement came with three numbers worth keeping in front of you. Five-hour limits doubled on every paid Claude Code tier. Peak-hours reduction is gone for Pro and Max. API limits for Opus models went up by a factor Anthropic described as “considerably,” with a per-tier table on the announcement page.²

The compute behind those limits is the SpaceX deal. Anthropic gets “all of the compute capacity at their Colossus 1 data center,” which adds “more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month.”² Whether the GPU mix on Colossus 1 affects Opus pricing is not stated, but the timing matters: a rate-limit doubling shipped on the same day as the capacity disclosure is a coordinated announcement, not two coincidences.

The CLI itself moved fast across the same window. Five releases shipped between May 1 and May 6:⁵

The EnterWorktree change in v2.1.128 is the kind of fix that prevents real data loss: branching from origin/<default-branch> instead of local HEAD silently dropped unpushed commits. The gateway-model-discovery change in v2.1.129 is a quiet but consequential reversal: third-party gateway shops who saw model listings appear automatically in v2.1.126 through v2.1.128 now have to set CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1 to keep that behavior. If your team runs Bedrock, Vertex, or Foundry through ANTHROPIC_BASE_URL, the env-var flip belongs in your shared .envrc.

The --plugin-url flag is the single most important addition for plugin distribution. A URL-fetched zip lets you point a teammate at a one-liner instead of asking them to clone a repo and run a build. For people maintaining the kind of plugin packaging covered in the Claude Code quickstart, that flag changes how install instructions get written.

For hook authors, the more relevant fix is in v2.1.126: Bash(mkdir *) and Bash(touch *) allow rules now honor in-project paths, which had been silently failing before. Permission rules that worked outside a project tree but not inside it are a class of bug that costs hours to track down.

Financial Services Agent Templates

The day before SF, Anthropic shipped ten financial-services agent templates. The lineup, in Anthropic’s own words: “Pitch builder creates target lists, runs comparables, and drafts pitchbooks for client meetings; Meeting preparer assembles client and counterparty briefs ahead of calls; Earnings reviewer reads transcripts and filings, updates models, and flags thesis-relevant changes; Model builder creates and maintains financial models from filings, data feeds, and analyst inputs; Market researcher tracks sector and issuer developments…Valuation reviewer checks valuations against comparables…General ledger reconciler reconciles general ledger accounts and runs net asset value calculations against the books of record; Month-end closer runs the close checklist…Statement auditor reviews financial statements…KYC screener assembles entity files.”³

The distribution model is the part to read twice: “Each one ships as a plugin in Claude Cowork and Claude Code, and as a cookbook for Claude Managed Agents.”³ One artifact, three surfaces. The same agent definition is invocable from a developer terminal, from the consumer Cowork product, and as a managed cookbook recipe. For shops that have been writing one-off scripts to do month-end close work, the cookbook surface is the path that does not require maintaining a runtime.

Three other shipping items belong with the templates:

• Microsoft 365 add-ins. “Claude can work directly in Microsoft Excel, PowerPoint, Word, and Outlook via add-ins.”³ Outlook lands later. Excel is the one to watch, because the alternative path (writing scripts that emit .xlsx files) loses the audit trail that finance teams need.
• Data connectors. Eight new connector partners landed at once: “Dun & Bradstreet, Fiscal AI, Financial Modeling Prep, Guidepoint, IBISWorld, SS&C IntraLinks, Third Bridge, Verisk.”³ These are not casual integrations. They are the licensed data feeds that institutional finance already pays for.
• Moody’s MCP app. “Moody’s has launched an MCP app that brings proprietary credit ratings and data on more than 600 million public and private companies.”³ An MCP-distributed Moody’s feed is the single most credible signal that MCP is going to be the integration substrate for licensed financial data, not a hobbyist protocol.

The benchmark number Anthropic cited alongside all of this: Claude Opus 4.7 “leads the industry on Vals AI’s Finance Agent benchmark, at 64.37%.”³ That number deserves the same scrutiny any benchmark gets. The interesting comparison point is whether 64.37% on a synthetic finance workload survives the contact with a real audit trail. The cookbook-on-Managed-Agents distribution model is what makes that a testable question: a firm can run the same agent against its own books and find out.

Computer Use Path: The Vercept Acquisition

Anthropic announced its Vercept acquisition on February 25, 2026, and the SF event is the first place to read the integration as a roadmap signal: “Today, we’re announcing that Anthropic has acquired Vercept to help us push those capabilities further.”⁴ Vercept’s stated thesis, as Anthropic frames it: “making AI genuinely useful for completing complex tasks requires solving hard perception and interaction problems.”⁴

The handoff is explicit. “Vercept will wind down its external product in the coming weeks and join Anthropic in pushing the frontiers of computer use.”⁴ Translation: the Vercept product is going away, and the team is being absorbed into the computer-use roadmap. Anthropic’s framing is that the team has “spent years thinking carefully about how AI systems can see and act within the same software humans use every day,” and that “expertise maps directly onto some of the hardest problems we’re working on at Anthropic.”⁴

What that means for builders depends on what computer use ships next. The current Claude Code desktop remote control surface is one of the cleanest places to watch the integration land, because remote-control and computer-use share the same problem space: see the screen, act on the UI, recover from failure.

What Apple Ecosystem Developers Should Take From This

If you build for iOS, three of the May 6 announcements matter more than the rest.

The rate-limit doubling changes how Xcode plus Claude Code work day to day. The previous five-hour ceiling was the constraint that pushed Apple Silicon developers to batch their work into deliberate sessions. With the ceiling doubled and peak-hours throttling gone for Pro and Max, you can run a longer build/run/test loop without pausing. If you have been running Codex on iOS work as a hedge against rate limits, the hedge gets less necessary.

The MCP-as-data-substrate signal from Moody’s matters because Apple’s own surface for structured data on iOS, App Intents, is converging with the same idea. An MCP app distributing 600M company records is the same shape as an App Intents donation flow distributing structured user state to Spotlight. The protocols are different. The bet that structured tool/data definitions get distributed independently of the UI is the same. If you are building an iOS app that wants to be queryable by Claude, the MCP path is now the one with industrial-grade examples to copy.

The Vercept acquisition is the one to watch quietly. Apple platforms have always been the hardest target for cross-platform automation, because UIKit and SwiftUI accessibility trees are not the same shape as the DOM. A team that “spent years thinking carefully about how AI systems can see and act within the same software humans use every day” gets handed a problem with two halves: the desktop browsers Vercept already understood, and the mobile and Mac apps the Anthropic computer-use product will keep getting asked to drive. The first surface where that work shows up will be Claude Code’s desktop remote control. The second is anyone’s guess.

FAQ

Did the May 6 Claude Code rate-limit changes apply to API users?

API users got a separate change. Anthropic raised “API rate limits considerably for Claude Opus models,” with the per-tier table published on the announcement page.² The five-hour-window doubling is specific to Claude Code on Pro, Max, Team, and seat-based Enterprise plans.

Are the financial agent templates available outside the finance vertical?

The templates are positioned for finance workflows, but they ship as Cowork and Code plugins plus a Managed Agents cookbook.³ The plugin format is not vertical-locked. A non-finance team could fork the General ledger reconciler template’s structure (close checklist, journal entries, close reports) for any reconciliation problem.

What changed for plugin authors in v2.1.129?

Three things. The new --plugin-url <url> flag lets you distribute a plugin zip via URL rather than a git clone. The skillOverrides setting now actually works (off, user-invocable-only, name-only). And themes and monitors should now sit under "experimental": { ... } in the manifest; top-level still works but claude plugin validate will warn.⁵

What does the SpaceX Colossus 1 deal mean for Opus pricing?

Anthropic did not publish pricing changes alongside the capacity announcement. The disclosed numbers are 300+ megawatts and “over 220,000 NVIDIA GPUs” coming online “within the month.”² The rate-limit doubling shipping the same day suggests the new capacity is being routed to Pro/Max/Team/Enterprise headroom first.

Is --dangerously-skip-permissions now safe to run inside .claude/?

Safer is the right word, not safe. As of v2.1.126, the flag bypasses prompts for writes to .claude/, .git/, .vscode/, and shell config files. “Catastrophic removal commands still prompt as a safety net.”⁵ The flag now honors what its name has always claimed; the name is still accurate.

References