Anthropic Ends Claude Subscription Support for OpenClaw and Third-Party Harnesses
Starting today at 12pm PT (3pm ET), Claude subscription credits no longer cover OpenClaw or any other third-party agentic harness. Boris Cherny, Head of Claude Code at Anthropic, confirmed the change: subscriptions were never designed for the usage patterns of external tools, and compute capacity needs to be prioritized for customers using Anthropic’s core products. The move cuts off the workaround that let power users run Claude at subscription pricing through community-built interfaces.
Anthropic is softening the blow with a one-time credit equal to the user’s monthly subscription cost, redeemable by April 17, and pre-purchase bundle discounts of up to 30% on Extra Usage. Users can also switch to a separate Anthropic API key to continue using third-party tools at standard API pricing. The timing is notable — this lands as the Peter Steinberger (OpenClaw creator) departure to OpenAI is still fresh. Hacker News is not happy. Expect more discussion over the weekend.
1M Context Window Beta Retiring April 30 for Claude Sonnet 4.5 and Claude Sonnet 4
Anthropic announced that the context-1m-2025-08-07 beta header will stop working for Claude Sonnet 4.5 and Claude Sonnet 4 on April 30, 2026. After that date the header has no effect. Developers using these models for long-context tasks need to migrate now. The recommended path: move to Claude Sonnet 4.6 or Claude Opus 4.6, both of which support the full 1M token context window at standard pricing with no beta header required.
The retirement signals that 1M context is no longer experimental — it’s a stable production feature on the newer models. If you’re still running Sonnet 4.5 or the original Claude Sonnet 4 in long-context workflows, the April 30 deadline is now your migration pressure point. Sonnet 4.6 is the same price tier; Opus 4.6 offers more capability at higher cost.
Message Batches API Hits 300k Output Tokens; Web Search and Web Fetch Go GA
Anthropic pushed several meaningful API upgrades this week. The Message Batches API max output cap is now 300k tokens for Claude Opus 4.6 and Sonnet 4.6 — include the output-300k-2026-03-24 beta header to unlock it. That’s a significant jump for batch workflows generating long-form content, structured data exports, or large code files. The Models API also got smarter: GET /v1/models now returns max_input_tokens, max_tokens, and a full capabilities object, so you can query model limits programmatically instead of hardcoding them.
On the tools side, the web search and web fetch tools are now generally available — no beta header required. Both support dynamic filtering, which uses sandboxed code execution to filter results before they hit the context window. That makes them faster and cheaper for high-volume use cases. API code execution is now free when combined with web search or web fetch calls, removing a previous cost barrier for tool-chained pipelines.
Claude Code Adds /powerup Lessons, New Hook Events, and Stability Fixes
The latest Claude Code update lands several developer-facing improvements. /powerup is a new interactive lesson system that teaches Claude Code features with animated demos — useful for onboarding new team members or exploring capabilities you haven’t used yet. Hook events got two additions: a defer permission decision for PreToolUse hooks (lets headless sessions pause at a tool call and resume with -p --resume), and a PermissionDenied hook that fires after auto-mode classifier denials with the ability to return {retry: true}.
Performance work includes replacing the WASM yoga-layout engine with a pure TypeScript implementation, which improves scroll performance in large transcripts. A bug causing the rate-limit options dialog to loop and crash long sessions is fixed. --resume now avoids a full prompt-cache miss for sessions with deferred tools, MCP servers, or custom agents. New env vars: CLAUDE_CODE_PLUGIN_KEEP_MARKETPLACE_ON_FAILURE for offline environments and CLAUDE_CODE_NO_FLICKER=1 for flicker-free alt-screen rendering.
Teams Shifting to Multi-Model API Platforms After Claude Code Leak
The March 31 Claude Code source leak — which exposed ~512,000 lines of TypeScript across 1,900 files via an accidentally bundled source map — is reshaping how some engineering teams think about model provider dependency. Several firms are accelerating moves to unified multi-model API platforms that route across Claude, Gemini, OpenAI, and open-source models through a single endpoint. The pitch: if one provider has an incident, you’re not fully exposed.
This is partly vendor-hedging theater and partly a genuine architectural conversation. Claude’s ecosystem has grown to 6,000+ integrations and 75+ connectors, which means more attack surface when something goes wrong. The irony is that Anthropic’s MCP (Model Context Protocol) was designed to make Claude more extensible — but extensibility cuts both ways. The teams responding most seriously to the leak aren’t dumping Claude; they’re adding resilience layers around it.
Hacker News Reacts to OpenClaw Cutoff: Pricing Trust Takes a Hit
The Hacker News thread on the OpenClaw subscription change is tracking strongly. The dominant sentiment: this feels like a bait-and-switch. Users built workflows and business tools on the assumption that Claude subscription pricing was stable infrastructure. Anthropic’s framing of the change as a compute management issue isn’t landing well with a community that reads product decisions closely. The one-time credit offer is seen as a short-term patch, not a solution.
The underlying tension is real: Claude subscriptions were never API access, but the line between “using Claude through a third-party interface” and “using the Claude API” was blurry enough that many users didn’t see this coming. Anthropic is now drawing that line clearly. Developers who built on OpenClaw-style access patterns are the ones feeling it most. Expect this to sharpen the API-vs-subscription conversation for the rest of Q2.
The OpenClaw Move Is About Control, Not Just Compute
Anthropic framed the OpenClaw cutoff as a resource management issue. But the timing and the precision of the change tell a different story. This isn’t a blanket usage cap — it’s a policy that specifically targets third-party harnesses using subscription credits to access Claude outside of Anthropic’s own products. That’s a distribution decision, not a capacity decision.
Read alongside the Claude Code source leak, the Accenture partnership, and the $100M Partner Network commitment, a pattern emerges. Anthropic is tightening its grip on how Claude reaches end users. The company wants Claude consumed through Claude.ai, Claude Code, Cowork, and vetted enterprise partners — not through an open ecosystem of third-party tools that Anthropic doesn’t control, can’t monitor, and doesn’t profit from at the same margin. The API remains open. But the subscription tier is increasingly reserved for Anthropic’s own surface area. That’s a meaningful strategic signal for anyone building on Claude outside of official channels.