Claude Code 2.1.314 Ships as Copilot Code Review Adds Max Tier

Two release trains, one market

Anthropic and GitHub spent the last week of October 2026 shipping updates to AI code review at the same time, and they are building toward different ideas of what the reviewer's job is. Anthropic's Claude Code moved to version 2.1.314 on October 30. The release added an option to lock writes to working directories, routing of gateway traffic by identity-provider group, and automatic compaction for long chats that have gone idle.26 Two days earlier, GitHub put a new "Max" effort level for Copilot code review into public preview. It also made Copilot approvals generally available, which means Copilot's sign-off can now count toward a repository's required reviews.3

The two announcements differ in kind. Claude Code's changelog is a steady stream of reliability work, much of it aimed at its /code-review workflow. GitHub's release is a policy change, because it lets an AI's approval stand in for a human one. My read is that GitHub's move matters more over time, even though Anthropic currently ships at a faster pace.

What changed in Claude Code

The newest GitHub release notes for Claude Code are mostly fixes. Idle compaction now covers VS Code and the Code tab in Claude Desktop. If a conversation of 200K or more tokens sits idle and its prompt cache expires, it gets summarized, and the summary is used on the next message.21 Sessions that kept failing with "Invalid effort level" when an admin capped a model's effort now drop to a lower effort, retry, and say which level they used.21 Headless mode had been exiting silently with code 0 when startup failed. It now prints the error and exits with code 1.21

Several fixes touch review directly. A finished /code-review, or any other skill running in a sub-agent, no longer runs a second time when a cloud or background session restarts.21 Version 2.1.307 added a command to stop ongoing Code Review subscriptions.22 It also added a claude_code.safety_event OpenTelemetry event, which records tool calls that a permission rule, auto mode, safety check, hook or sandbox acted on.28 Version 2.1.304 brought new Code Review controls and security checks for Managed Agents.22 Earlier, 2.1.287 fixed review comments and their "Why this was flagged" explanations being cut off mid-sentence.25

The cadence is part of the story. Third-party trackers list releases arriving every few days through August, September and October.26 Some of them are large: one tracker describes 2.1.290 as a 190-bullet batch of fixes, and 2.1.292 as closing six sandbox and permission holes at once.27 That same history shows the review feature taking shape. /code-review moved to the background in July, and /review was folded back into it in August.27

My reading is that Anthropic treats Claude Code as infrastructure for agents. Review is one skill running on top of sub-agents, sandboxes and permission hooks, and most of the engineering goes into making that base trustworthy. A lot of the recent work is about governance, including mod-level permission deciders, gateway routing and safety telemetry.2926 That points to enterprise buyers who want audit trails before they let agents run on their own.

GitHub's bet: the AI can approve

GitHub took a more direct route. Max runs several independent review agents on models from different providers. An aggregation agent then reconciles their findings and posts one combined review.29 GitHub frames it as the level for high-risk changes such as authentication logic or shared libraries. It also says Max takes longer and uses more AI credits than Lite or Balanced.3

Approvals are the bigger change. They are off by default. Admins can turn them on at the enterprise, organization or repository level and limit which file paths Copilot may approve. If new commits arrive after an approval, it is dismissed, just as a human reviewer's would be.3 The feature has been through a staged rollout. It entered preview on September 1, and comment auto-resolution followed on September 11.7 On October 2, review requests became available through the REST and GraphQL APIs with a per-request effort level. On September 28, Balanced replaced Lite as the default.1

GitHub also changed how Copilot is sold. Enterprise customers can now use all Copilot features, including code review, without buying a separate subscription.10 Separately, the company launched Copilot Projects, which spreads work across parallel agents and returns pull requests for humans to review.9 Put together, these moves look like a feedback loop. More agent-written code means more pull requests, and GitHub's answer is to sell more review, partly done by agents, to handle them. Its Universe keynote said much the same thing: more code means more demand for review.10

Where the coverage disagrees

Cost is the clearest disagreement. GitHub's documentation estimates $0.05 to $1 in AI credits per Lite review and $0.25 to $5 per Balanced review.5 InfoWorld reports GitHub's average figures as about $0.25 for Lite, $1.50 for Balanced and $10 for Max.9 Other reporting says GitHub has not published a credit range for Max at all.2 The figures don't strictly contradict each other, since averages can sit inside ranges. Still, buyers should treat the $10 Max figure as provisional until GitHub's own documentation catches up. One analysis noted that the documentation had not yet been updated to describe Max as available.2

Quality claims are even less settled. GitHub says Max posted the highest F1 score on ReviewBench for high-severity issues.10 It has not released the score, the competing systems' scores or the test set, so nobody can check the claim yet.2 The independent benchmark that is public shows a tight field. Martian's Code Review Bench, checked October 4, put Copilot at 61.0% F1. That was within about a point of Greptile, CodeRabbit, Macroscope and Claude, with Claude at 60.2%.8 A later snapshot listed Claude at 60.9 and CodeRabbit at 61.8.14 With scores that close, small changes in the test set can reshuffle the order. One October comparison ranked Cursor Bugbot first on high-severity bugs.12

Anthropic's own numbers come from Anthropic. It says 84% of pull requests over 1,000 lines get findings and fewer than 1% of findings are marked incorrect.13 Benchmarks rarely capture the price gap, though. Claude's managed review averages roughly $15 to $25 per review, billed separately from the subscription.1319 Copilot's Balanced default costs far less per run.5

Approver versus adviser

This is the main split in the market. Claude Code Review reportedly reached general availability for Team and Enterprise on October 15. It is described as never approving or blocking a merge.12 CodeRabbit, Greptile and Qodo also only comment.12 Codex can submit approve or request-changes verdicts from ChatGPT desktop.12 Copilot's approval now counts toward branch rules, though.3

GitHub's own responsible-use guidance still says Copilot may miss issues in large or complex changes and should be backed up by careful human review.7 An analyst quoted by InfoWorld agreed. He said multiple agents improve the odds of catching problems but guarantee nothing, and that high-risk changes should keep human approval.9 The safeguards GitHub built suggest it expects cautious, selective adoption: approvals are off by default, limited by path, and dismissed on new commits.2

Practitioners' experience is mixed. One team took mandatory Copilot review off a repository in July because the comments were noisy and advisory. It made Codex its main release reviewer and kept CodeRabbit as an option.4 Comparison guides often recommend running two layers: a marketplace reviewer by default, plus a narrow custom Claude check for domain rules.1120

The takeaway

I read this as a split rather than a race. GitHub is making the reviewer part of the merge gate and charging by depth. Anthropic is making the agent runtime under review safer and easier to audit, and it stops short of giving the AI a vote. Teams already standardized on GitHub will find Copilot's approvals and API hooks the easiest path. They should audit which repositories quietly moved to Balanced, and the higher bills that came with it.5 Teams that already run Claude Code have a reviewer that keeps improving almost weekly. It costs more per review and leaves the final call to a person.1213