By Assaf Biton
Everyone calls a one-release-per-day pace "noise you can ignore." I read all 49 of them so you do not have to. Between v2.1.98 (April 10, 2026) and v2.1.146 (May 21, 2026),
Anthropic's terminal coding agent shipped 49 versions in six weeks, adding Agent View, /goal, Routines, and /ultrareview while Opus 4.7 became the default model
Best for: Senior developers who want to delegate long autonomous tasks and review results, DevOps teams integrating AI into CI pipelines for automated test fixing
TL;DR: In six weeks Anthropic shipped 49 Claude Code releases, all on the v2.1.x line (no v2.2 yet).
Claude Opus 4.7 went generally available on April 16, 2026, and became the default coding model on Max and Team Premium during the April 13 to 17 rollout window, with a new xhigh effort level
Best for: Knowledge workers doing deep document analysis and long-form writing, Developers who need precise instruction-following for complex multi-step tasks
xhigh effort level between high and max. Agent View (claude agents, May 11) gives you one screen for every background session. /goal keeps Claude running turn after turn until a condition holds. Routines run cloud agents while your laptop is closed, and /ultrareview spins up a fleet of bug-hunting reviewers, now $5 to $20 per run since the free-run window closed on May 5.
Why the May 2026 update wave matters
April was about autonomy: Auto Mode, more hooks, the first cloud runtimes. If you missed it, our Claude Code April 2026 recap covers that sprint. May is about scale. The question shifted from "can Claude run one task without me approving every step?" to "how do I watch ten tasks at once without ten terminals open?" The unit of work changed from a terminal to a session you can leave and come back to.
Opus 4.7 is the new default coding model
Claude Opus 4.7 went generally available on April 16, 2026, and the default rollout in Claude Code happened during the Week 16 window of April 13 to 17 (v2.1.105 to v2.1.113), when it became the default on Max and Team Premium. Everywhere else you pick it from /model. The practical change is a new effort level called xhigh that sits between high and max, applied automatically the first time you switch to 4.7. The full ladder reads low, medium, high, xhigh, max, and running /effort with no argument now opens an arrow-key slider. Fast mode moved to 4.7 on May 11 at the same price as before.
Pro Tip: On Max, set /model opus then /effort xhigh for agentic coding. That is the documented recommendation for most coding tasks. Keep high for quick edits, where xhigh burns tokens you do not need to spend.
Agent View: one screen for every background session
This is the biggest workflow shift, and it shipped on May 11 (v2.1.139) as a research preview. Run claude agents and you get a full-screen dashboard listing every Claude Code session on your machine, grouped by state: Working, Needs input, Ready for review, Completed. Each row is a real conversation that runs with no terminal attached, hosted by a per-user supervisor process that survives auto-update and machine sleep. Each session also isolates itself in its own git worktree before editing, so ten parallel sessions read the same checkout but never stomp on each other's writes.
The clever part is the peek loop. Press Space on a row and a side panel shows the session's last output or the question it is stuck on. Type a reply, press Enter, and it goes to that session without you leaving the dashboard. So you steer a dozen agents from one screen instead of tab-hopping across windows.
Watch the meter. Each background session burns your subscription quota independently, so running ten agents in parallel uses quota roughly ten times as fast as one. The redesigned /usage screen now breaks down exactly what is driving your limits: parallel sessions, subagents, cache misses, and long context, each with a percentage and a tip.
/goal and the new autonomous loops
/goal <condition> sets a completion condition for the session. After every turn, a fast model (Haiku by default) checks whether the condition holds. If not, Claude starts another turn using the evaluator's reason as guidance. If yes, the goal clears. So you can write /goal npm test exits 0 and walk away.
Here is the catch most people miss. The evaluator does not call tools. It only judges what Claude has already surfaced in the conversation. So "the database is healthy" will not work as a goal unless you also ask Claude to check the database and report back. Conditions have to be things Claude's own output can demonstrate.
Two related changes set this up about five weeks earlier, back in v2.1.98 on April 9, 2026. The Monitor tool streams background watcher events straight into the chat, killing the old "Bash sleep loop" trick. And /loop got self-pacing the same day, so you can leave off the interval and let Claude schedule the next tick itself.
Cloud features: Routines, Ultraplan, and Ultrareview
Routines are saved Claude Code configs (prompt, repos, connectors) that run on Anthropic's cloud on a schedule, a GitHub event, or an API call. The point that separates them from Remote Control: they keep working when your laptop is closed. By default a routine can only push to branches prefixed claude/, a guardrail against a scheduled job touching your main branch.
Ultraplan hands a planning task to a cloud session while your terminal stays free, then lets you comment on the draft in the browser. You launch it with /ultraplan and pick whether the plan runs on the web or comes back to your terminal.
Ultrareview runs a fleet of reviewer agents in a sandbox and reproduces each bug before reporting it. Both Ultraplan and Ultrareview need a claude.ai login, and neither works on Bedrock, Vertex, or Foundry.
The free Ultrareview runs are gone. Pro and Max accounts got three free /ultrareview runs at launch, but that one-time allotment expired on May 5, 2026. After that, every run bills to usage credits at roughly $5 to $20 depending on the diff size, and your account needs usage credits turned on or the launch is blocked. Check with /usage-credits first.
The quiet wins worth knowing
A few smaller changes punch above their weight. claude --from-pr 1234 resumes the exact session that opened a pull request, from the PR alone. Native per-platform binaries replaced bundled JavaScript on April 17, so claude no longer spins up Node and starts faster. Windows users no longer need Git Bash either. Small wins, but they remove daily friction.
What's next for Claude Code
The cadence has not slowed. Since this update wave closed on May 21, Anthropic released Claude Opus 4.8 on May 28 and made it the new default for serious coding shortly after. The version stream is still on the v2.1.x line as of mid-June, now in the v2.1.18x range (latest is v2.1.183), with no v2.2 announced. So expect Opus 4.8 to inherit everything described here, including xhigh and Agent View.
Claude Code May 2026 update questions
What is the current default model in Claude Code?
During the May 2026 wave, Claude Opus 4.7 was the default on Max and Team Premium (since April 13). As of late May, Opus 4.8 took over as the new default after its May 28 release, so a fresh install today starts on 4.8.
Is Agent View in Claude Code generally available?
No. Agent View (claude agents) shipped May 11 as a research preview in v2.1.139, not a stable release. It works, but it can change. You need v2.1.139 or later to run it.
How much does /ultrareview cost now?
The three free runs were a one-time launch allotment that expired May 5, 2026. Since then, each run bills to usage credits at roughly $5 to $20 depending on the diff size, and your account must have usage credits enabled.
What to do this week
If you are on Max: switch to /model opus with /effort xhigh and start your day from claude agents instead of a plain terminal. That single habit change is the whole point of the May wave.
If you run long tasks: reach for /goal with a condition your output can prove, and let it run while you do something else.
Before you pay for Ultrareview: the free runs ended May 5, so budget $5 to $20 per run and reserve it for pull requests that actually matter.
The honest read: most of these 49 releases are noise for a casual user. But Agent View, Opus 4.7, and /goal together change what one person can run at once. If you adopt one thing, make it the claude agents habit. For deeper context on the model, see our Opus 4.7 launch coverage and the AI news feed.
