TL;DR: Anthropic shipped five Claude Code versions between April 14–21, delivering a 67% session resume performance gain, a cost-cutting prompt caching flag, and a significant architectural shift to native binaries. Meanwhile, Codex is quietly building out SSH remote connections. Update to 2.1.116 today. Here's everything that matters.

Five versions in seven days is not normal release cadence. It signals Anthropic is in execution mode on Claude Code, treating it as a product that competes on iteration speed, not just model quality. These updates collectively touch cost, performance, security, and developer experience. Here's what actually matters, ranked by impact.

The headline number: Claude Code 2.1.116 delivers up to a 67% performance improvement on `/resume` for sessions exceeding 40MB. If you're running long-context coding sessions, which any serious AI-augmented engineering workflow will accumulate quickly, this is the most immediately felt improvement in the batch. Large sessions are the reality of agentic coding. A multi-file refactor, a debugging session that spans an hour, or an agent loop running across a codebase all generate substantial context. Before 2.1.116, resuming those sessions had become a friction point. That friction is now largely gone. This is the update to prioritize. Action: Run `npm update -g @anthropic-ai/claude-code` or your equivalent update path. There is no reason to stay on earlier versions.

Version 2.1.108 introduced the `ENABLE_PROMPT_CACHING_1H` environment variable, enabling a 1-hour prompt cache TTL across Anthropic API, Bedrock, Vertex, and Foundry. The economics here are significant: prompt caching can reduce API costs by up to 90% on repeated queries within the cache window. For enterprise teams running Claude Code at scale, this is a lever worth pulling immediately. The 1-hour TTL is particularly well-suited to focused coding sessions where context (your system prompt, project files, coding standards) gets reused repeatedly across a work block.

Platform support covers all major deployment targets: native API, AWS Bedrock, Google Vertex, and Foundry. If your organization routes through any of these, evaluate this now.

Most coverage will focus on the `/tui` fullscreen mode added in 2.1.110 (flicker-free rendering, same-conversation context, genuinely useful). But the more strategically important development is the observability layer Claude Code is assembling. The `/recap` feature gives engineers a way to re-establish context when returning to a session, surfacing a structured summary rather than requiring a manual scroll through conversation history. That sounds like a minor convenience. It is not. In an agentic workflow where a session might have executed dozens of tool calls across hundreds of lines of output, `/recap` is what makes long sessions navigable. Combined with `/cost` breakdowns and PreCompact hooks (which fire before context is compacted, letting you inject logic or summaries), Anthropic is building something Cursor and VS Code extensions have not: a model of the agent loop as a measurable, tunable system rather than a black box. This matters for engineering leaders because it changes how you manage AI usage at the team level. When you can see exactly what your agents are costing, when you can hook into compaction events, and when your engineers can recap context cleanly, AI-assisted development becomes auditable. Auditability is what converts skeptical CTOs into buyers.

Version 2.1.113 changed how Claude Code's CLI spawns its runtime: instead of executing bundled JavaScript, it now spawns a native binary delivered via per-platform optional dependencies (macOS, Linux, Windows). This is not a user-facing feature. It is an infrastructure decision with real security and operational implications. The case for it: native binaries reduce the JavaScript-in-a-process attack surface that has made Node-based CLIs a supply-chain target. Bundled JS is easier to tamper with or inject into; a platform-specific binary is harder to poison mid-delivery. The case for scrutiny: your security team now needs to verify and pin binaries for each platform rather than auditing a single JavaScript bundle. Binary distribution across three platforms introduces deployment complexity, especially in locked-down enterprise environments with strict allowlisting policies. The same version also added `sandbox.network.deniedDomains` for explicit domain blocking within the Claude Code sandbox. For teams running Claude Code in security-sensitive environments, this is a meaningful control. For security teams: Review your binary allowlisting policies. Map which platforms your Claude Code deployments run on and establish a verification workflow for the new binary artifacts before your next major deployment.

Version 2.1.109 added the ability for the model to discover and invoke built-in Slack slash commands. This is part of Anthropic's broader push on tool-use and agentic integration. For teams that have wired Claude Code into Slack workflows, this expands what the agent can do without explicit configuration. The practical ceiling here is still being mapped, but teams using Claude Code in DevOps notification pipelines or incident response workflows should explore what becomes newly automatable.

Codex 26.415 began rolling out SSH remote connections in alpha and added support for multiple terminals plus system tray integration. The SSH alpha is the development worth watching. Remote connections bring Codex into territory that Claude Code has not yet publicly committed to: the ability to connect a coding agent to a remote development environment rather than only your local machine. For distributed teams running development on cloud instances, this is a genuine workflow unlock. Multiple terminal support and system tray integration are table-stakes UX improvements, but they signal OpenAI treating Codex as a persistent developer environment rather than a session-based tool. That framing matters for enterprise adoption.

Claude Code is iterating faster on infrastructure and observability. Codex is moving faster on remote/distributed workflow features. Neither has the full picture yet, which is exactly why this space remains competitive.

Here is your concrete action list, prioritized:

Five Claude Code versions in one week is a signal, not just a changelog. Anthropic is treating Claude Code as infrastructure that needs to win on reliability, cost efficiency, and security as much as raw capability. The observability layer (/recap, /cost, PreCompact hooks) is the most underappreciated development in this batch. The teams that instrument their AI development workflows today will have compounding advantages in 12 months. The engineering orgs winning with AI in 2026 are not the ones that adopted the most tools. They are the ones that made those tools measurable. Claude Code is becoming measurable. That is worth paying attention to.