💡 300k Token Outputs Are Now Live on the Message Batches API

Anthropic has extended the maximum output length on the Message Batches API to 300,000 tokens for Claude Opus 4.6 and Claude Sonnet 4.6. This is a ten-fold increase over the previous 32k cap and opens up a class of long-form generation jobs that previously required stitching together multiple API calls or chunking output manually. To access it, add the output-300k-2026-03-24 beta header to your batch request.

What you can do with 300k output tokens

• Book-length drafts: A single batch task can now produce a full technical manual, legal brief, or multi-chapter document without mid-generation interruption
• Exhaustive structured extraction: Extract every entity, relationship, and annotation from a large corpus in one pass — no pagination logic required
• Large code scaffolds: Generate a complete multi-file project skeleton — routes, models, tests, config — in a single batch job
• Long reasoning chains: Use extended thinking alongside extended output for tasks requiring both deep deliberation and voluminous output

Request snippet

POST /v1/messages/batches
anthropic-beta: output-300k-2026-03-24
anthropic-version: 2023-06-01
Content-Type: application/json

{
  "requests": [
    {
      "custom_id": "long-doc-gen-001",
      "params": {
        "model": "claude-opus-4-6-20260130",
        "max_tokens": 300000,
        "messages": [{ "role": "user", "content": "Write a complete..." }]
      }
    }
  ]
}

💡 Data Residency Controls: Pin Your Model Inference to the US

Anthropic has shipped data residency controls for the Claude API — a new inference_geo parameter that lets you specify where model inference runs on a per-request basis. At launch, two options are available: "us" (US-only inference) and "global" (default, no geographic constraint). US-only inference is priced at 1.1× the standard rate for models released after February 1, 2026, including Opus 4.6 and Sonnet 4.6.

Who needs this

• US federal and state agencies subject to FedRAMP or ITAR requirements that prohibit data leaving US jurisdiction
• Financial services firms with regulatory guidance or internal policies restricting data processing to domestic infrastructure
• Healthcare organisations navigating HIPAA Business Associate Agreements that may carry geographic constraints
• Legal teams handling matter data under firm policies or client instructions requiring domestic processing

API usage

POST /v1/messages
Content-Type: application/json
anthropic-version: 2023-06-01

{
  "model": "claude-sonnet-4-6-20260131",
  "max_tokens": 1024,
  "inference_geo": "us",
  "messages": [{ "role": "user", "content": "Summarise this contract..." }]
}

💡 Claude Code's April Wave: MCP Persistence, --bare Flag & Channel Relays

The April 2026 Claude Code releases bring a cluster of quality-of-life improvements aimed at developers building on top of Claude Code or running it in automated pipelines. Three stand out as particularly significant for power users.

MCP tool result persistence override

MCP tool results have historically been truncated when they exceed Claude Code's internal buffer size — a problem for tools that return large payloads like database schemas, API specs, or full file trees. The new _meta["anthropic/maxResultSizeChars"] annotation on tool results lets individual MCP servers declare a higher persistence limit (up to 500,000 characters), bypassing the default cap for that result alone. If your MCP server regularly returns large structured payloads, annotate the result and the data will pass through without truncation.

The --bare flag for scripted automation

Running Claude Code in scripts or CI pipelines previously incurred overhead from hooks, the Language Server Protocol sync, plugin sync, and skill directory walks — none of which matter in non-interactive contexts. The new --bare flag skips all of these when combined with -p (print mode), giving you a leaner, faster execution path for scripted claude -p "…" --bare invocations.

--channels permission relay

Unattended Claude Code sessions on remote servers have long faced a problem: when a tool needs user approval, there's no one at the terminal to grant it. The new --channels flag enables channel servers that declare the permission capability to forward tool approval prompts to your phone or another device — turning any networked device into an approval terminal for remote Claude Code sessions.

Other notable additions

• disableSkillShellExecution setting — prevents inline shell execution in skills and custom slash commands, useful for teams that want to audit all shell calls through the standard Bash tool permission flow
• Multi-line deep links — claude-cli://open?q= now accepts encoded newlines, enabling richer pre-filled prompts from external tools
• Plugin bin/ executables — plugins can now ship their own executables under bin/ and invoke them as bare commands from the Bash tool, reducing the need for users to install dependencies separately
• /plan <description> — the plan command now accepts an optional inline description (e.g. /plan fix the auth bug) that enters plan mode and immediately seeds the first step
• Simplified effort levels — the max effort level is removed; levels are now low / medium / high with new symbols ○ ◐ ● replacing the previous indicator icons