Anthropic shipped Claude Code 2.1.94 this week, and the headline feature isn't the UI polish — it's Amazon Bedrock integration via Mantle and a default effort level bump that will hit your API bill before you finish reading this. Here's what changed, what's broken, and what you should do right now.

The official changelog confirms three changes worth your attention:

There's also a compact Slack `#channel` header with a clickable channel name — a small UX quality-of-life change that signals Anthropic is investing in Claude Code as a team-layer product, not just a developer CLI. And yes — Bedrock Bearer Token (ABSK) authentication is broken in 2.1.94. We'll get to that.

Everyone's talking about the effort level change. They're missing the point. Mantle is Anthropic's emerging orchestration layer that enables Claude Code to route requests across multiple LLM backends — including Amazon Bedrock — without forcing teams to rebuild their entire toolchain. Setting `CLAUDE_CODE_USE_MANTLE=1` opts your environment into this infrastructure. What does this actually mean for enterprise teams? It means Claude Code is positioning itself as a multi-provider agent orchestrator, not just a single-model coding assistant. AWS-native shops — the ones running workloads on Bedrock, storing data in S3, authenticating via IAM — can now route Claude inference through their existing cloud infrastructure instead of punching new holes in their security perimeter for direct Anthropic API calls. That's a significant enterprise unlock. Security teams at Fortune 500 companies have been blocking direct API keys to external AI providers for 18 months. Bedrock changes the conversation from "can we use this?" to "here's the procurement path we already have approved." The competitive implication is blunt: this is a direct shot at GitHub Copilot's enterprise lock-in. Copilot sits inside GitHub's ecosystem. Claude Code via Bedrock sits inside AWS's ecosystem. For the roughly 45% of enterprise software teams that are AWS-primary, this matters enormously.

The shift from medium to high effort as default isn't cosmetic. High effort means Claude runs more internal reasoning passes before returning a response — better code, more thorough analysis, longer latency, higher token consumption.

This is exactly why defaulting to high effort makes strategic sense for Anthropic. Better outputs justify the product. But better outputs cost more tokens, and at scale, that math matters.

If you're running Claude Code in automated pipelines — PR reviews, test generation, doc updates — audit your token usage this week. High effort in a loop is how you generate a $40K surprise invoice. The fix is simple: set `--effort medium` explicitly in your automation configs. Don't wait for your billing alert to tell you.

Let's be direct about the bug. Bedrock Bearer Token (ABSK) authentication is broken in 2.1.94. If your team authenticates to Bedrock via ABSK and you update to 2.1.94, you will hit authentication failures. The current workaround: downgrade to 2.1.92.

This is a real problem, and it's worth naming clearly: shipping broken Bedrock auth in the same release that announces Bedrock support is an embarrassing release coordination failure. Anthropic's release cadence has been aggressive — they've shipped double-digit patch versions in the first quarter of 2026 alone — and this is the cost of that velocity. Features ship; integration tests miss edge cases in production. There's also a separate reported memory leak affecting the Cursor extension, with some users hitting up to 37GB of RAM consumption. That's not a Claude Code bug specifically, but it lives in the same ecosystem and reflects the broader growing pains of rapidly iterated AI tooling. OOM cascades in a development environment are a productivity killer. If your team runs Claude Code inside Cursor, watch your memory metrics and keep `ulimit` guardrails in place.

Here's the honest comparison as of this release:

Cursor wins on stability right now. Copilot wins on enterprise GitHub integration. Claude Code wins on reasoning depth and — as of this release — AWS-native deployment for teams that need it. The teams that should move fastest on 2.1.94 are AWS-centric enterprises that have been blocked from Claude adoption by security constraints. The teams that should wait: anyone relying on ABSK authentication. Downgrade to 2.1.92 and wait for the patch.

The Mantle-powered Bedrock integration signals something larger than one version bump. Anthropic is building Claude Code to sit at the infrastructure layer of software engineering, not just the IDE layer. When Claude can route across AWS Bedrock, Google Vertex, and Azure OpenAI from a single interface, it stops being a coding assistant and starts being the coordination layer for how your engineering team interacts with AI at scale. This is exactly the kind of tooling that changes how you staff teams. Not because engineers become less necessary — they become more powerful. A senior engineer who can orchestrate multi-model pipelines through Claude Code, route inference through Bedrock for compliance, and hook everything into CI/CD via the Agent SDK can do what required a four-person team 18 months ago. That engineer is rare. They're getting rarer as demand accelerates. Traditional hiring platforms built for resume screening and keyword matching aren't finding them — because the signal isn't in a resume, it's in how they actually build with AI. The teams winning in this environment aren't hiring more engineers to compensate for tool complexity. They're hiring fewer, better engineers who make these tools compoundingly effective. 2.1.94 is one more reason the talent gap between AI-native engineers and everyone else is widening — and why finding those engineers is the actual competitive advantage. Watch for the ABSK patch. Watch your token costs. And start thinking about Bedrock as your on-ramp to the multi-cloud AI stack that's coming whether you plan for it or not.