Claude Code 2.1.321 shipped on 9 November 2026. This version makes Claude Fable 5.2 the default Fable model and fixes a flaw that could send MCP OAuth refresh tokens to another host. If your agentAI agentAn AI program that performs work steps by itself, for example reading a message, drafting a reply, and recording the result.Open the glossary uses MCPMCPA standard that lets an AI program use your tools and data under clear permissions.Open the glossary servers with OAuth sign-in, or you saved Fable 5.1 as your model, upgrade and check your model today.
ClaudeClaude ProThe paid subscription for the Claude AI assistant from Anthropic. It unlocks connectors to outside tools.Open the glossary Code is the coding agent from Anthropic. It runs in the terminal, the desktop app, IDEs, and the cloud. MCP (Model Context Protocol) is a standard that connects AI to other applications. OAuth is a sign-in method that gives the agent an access token for a service, without storing your password. A refresh token is the key that keeps the agent signed in without a new sign-in from you.
What changed in Claude Code 2.1.321?
Claude Code 2.1.321 contains 118 changes, according to the official release notes on GitHub, published on 9 November 2026 at 19:22 UTC. The split is 3 new features, 74 bug fixes, 15 improvements, 23 behavior changes, and 3 text updates.
These changes have the most impact on teams that run agents:
- Fable 5.2 becomes the default Fable model. Model ID
claude-fable-5-2, 1M token context, $8 input and $40 output per 1 million tokens, cache reads at $0.20. - Saved Fable 5.1 picks move to the
fablealias. Pro and Max plans move at once. Team and Enterprise plans move after the organization allows it. - The MCP OAuth flaw is fixed. Before, secrets such as refresh tokens could go to another host when an MCP server sent a redirect.
- 3 Bash permission flaws are fixed. The next section gives the details.
- New option
claude remote-control --settings <file>. This option applies 1 settings file to every session that Remote Control starts. - Mods that Claude writes turn on without a prompt in auto and bypass modes. Tool calls from mods also show no permission prompt. Deny rules and organization policy still apply.
The previous release, 2.1.319, fixed 6 other permission flaws. Read the summary in our Claude Code 2.1.319 article on permission flaws and auto mode.
Why does the MCP OAuth flaw matter for your agent?
This flaw could leak a refresh token to the host that an MCP server redirected to. The holder of that token could use your access to that service until you revoke the token.
The 2.1.321 release notes describe this fix in 1 line. On 10 November 2026, we found no GitHub advisory (GHSA) and no CVSS score for this flaw on the Claude Code advisory page. So no official risk level exists yet.
This table shows the 4 security fixes in 2.1.321:
| Flaw | Behavior before 2.1.321 | Who is affected |
|---|---|---|
| MCP OAuth | Refresh tokens could go to another host when the server sent a redirect | Users of remote MCP servers with OAuth sign-in |
| Bash sandbox | In 1 case, a read-only Bash command ran outside the sandbox without a permission prompt | Users of sandbox mode |
| Working folder | Permission checks judged a file write from the wrong folder and did not prompt | Users of per-folder permission rules |
| cp and mv | Deny rules and protected-file prompts did not apply to the file that cp or mv into a folder writes | Users of deny rules for secret files |
The highest risk is on third-party MCP servers that you do not control. Your own server, such as the ones we build with our MCP server development service, is easier to audit because you know the target of each redirect.
Fable 5.2 is the default: what is the impact on cost and limits?
Fable 5.2 costs 20% less per token than Fable 5.1, but its price is still 2 times the price of Opus 5.5.
| Model | Input per 1 million tokens | Output per 1 million tokens | Cache read | Context |
|---|---|---|---|---|
| Claude Fable 5.2 | $8 | $40 | $0.20 | 1M |
| Claude Fable 5.1 | $10 | $50 | $0.25 | 1M |
| Claude Opus 5.5 | $4 | $20 | $0.20 | 1M |
Sources: Claude models overview and Fable 5.1 page, read on 10 November 2026.
Illustration for 1 task with 200,000 input tokens and 20,000 output tokens, without cache: Fable 5.1 costs $3.00, Fable 5.2 costs $2.40, and Opus 5.5 costs $1.20.
In the Anthropic announcement of 9 November 2026, Fable 5.2 scores 69.9% on FrontierSWE v2. Fable 5.1 scores 56.3% and Opus 5.5 scores 62.3%. Anthropic also states that false safeguard blocks fell by about 55% compared with 5.1. Anthropic still names Opus 5.5 as the recommended daily model for most coding work.
Plan limits also change. The same announcement states that Fable can use up to 50% of weekly limits on Pro plans and standard Team seats. The Claude Code model documentation states that on some plans, Fable uses usage credits. The /model menu then shows the label "Requires usage credits" on the Fable row.
For older pricing and a migration path, read our Claude Opus 5.5 pricing and migration guide.
Which early bug reports should you know about?
In the first 24 hours, users opened 3 types of GitHub bug reports that name Fable 5.2 or Claude Code 2.1.321.
- The desktop app cuts the Fable 5.2 context to 200K. The Code tab in the desktop app still bundles Claude Code 2.1.317. That version does not know Fable 5.2, so sessions auto-compact at about 170,000 tokens, not near 1M. Source: issue #108640.
- Normal questions get refused. Users report "reasoning_extraction" errors and safeguard flags on normal requests. Sources: issue #108651, #108682, and #108708.
- The 5-hour limit runs out faster. 1 report uses Fable 5.2 (issue #108697). 1 other report uses Sonnet 5.5 and hit the limit at about 1/6 of the output tokens of earlier windows (issue #108661).
All issues above are still open on 10 November 2026. Anthropic has not given a cause. The limit reports are not proven to come from version 2.1.321.
How does 2.1.321 feel in daily use?
In daily use, the largest change is in model choice, not in the interface.
What works well:
- Users of the
fablealias move to Fable 5.2 with no settings change. - Teams can apply 1 shared settings file to every Remote Control session with the
--settingsoption. - Long sessions are more stable. The release notes list a fix for high memory use in long sessions and a fix for replies that stopped when a TLS connection failed.
What causes friction:
- Saved Fable 5.1 picks move to Fable 5.2 automatically on Pro and Max. Cost and limit use can change without notice.
- The desktop app lags one version behind, so Fable 5.2 there does not use the 1M context yet.
- Mods now turn on without a prompt in auto and bypass modes. Teams with strict security rules must check the list of mods.
Example scenario: a marketing team uses Claude Code to pull ad reports every night through an MCP server with OAuth sign-in. In older versions, a redirect from that server could carry the refresh token to another host. In 2.1.321, the token stays with the original host. The team updates, restarts the scheduled agent, and signs in to the MCP server again if in doubt.
This release matters most to 3 groups: developers who use the sandbox and deny rules, business owners who run scheduled agents, and marketers who connect ad accounts through MCP.
Our checks on 10 November 2026
We checked the release channels on the npm registry and the versions on 2 of our work machines on 10 November 2026 at 09:30 WIB.
| Item checked | Result |
|---|---|
| latest channel | 2.1.321 |
| stable channel | 2.1.311 |
| next channel | 2.1.322, release notes not yet on GitHub |
| Work machine 1 | 2.1.321 through auto-update |
| Work machine 2 | 2.1.299 |
Users on the stable channel are still on 2.1.311. They do not have the MCP OAuth fix yet. The model documentation also states that Fable 5.2 needs Claude Code 2.1.321 or later. Our second machine shows that 1 machine can fall far behind, so check every machine that runs an agent.
How do you upgrade to 2.1.321 safely?
A safe upgrade has 7 steps, and the most important step is to restart all sessions that still run.
- Check the version on each machine with
claude --version. - Run
claude update. Users on the stable channel wait until 2.1.321 reaches that channel, or move to the latest channel. - Restart all sessions, including
claude remote-controland scheduled agents. - Open
/modeland check the active model. To stay on Fable 5.1, run/model claude-fable-5-1. - Look for the "Requires usage credits" label on the Fable row before you use Fable 5.2 for long tasks.
- For third-party MCP servers with OAuth, revoke access in the provider app if in doubt. Then sign in again.
- Use the CLI for long Fable 5.2 sessions until the desktop app bundles 2.1.321.
Conclusion: what is your next action?
Upgrade to 2.1.321 today if your agent uses MCP servers with OAuth, the sandbox, or deny rules. Test Fable 5.2 on 1 real task before you make it your main model, and keep Opus 5.5 for daily work when cost matters. If your team needs help to install and maintain agents with safe permission rules, see our AI Agent Ops Install service.
Questions and answers
Do I need to upgrade to Claude Code 2.1.321 now?
Yes, if your agent uses MCP servers with OAuth sign-in, sandbox mode, or deny rules. This version fixes the refresh token leak and 3 Bash permission flaws.
Is Fable 5.2 more expensive than Fable 5.1?
No. Fable 5.2 costs $8 input and $40 output per 1 million tokens, 20% less than Fable 5.1. Its price is still 2 times the price of Opus 5.5.
How do I go back to Fable 5.1?
Run /model claude-fable-5-1 in a Claude Code session. You can also set the ANTHROPIC_DEFAULT_FABLE_MODEL variable so that the fable alias stays on Fable 5.1.
Why does Fable 5.2 compact early in the desktop app?
The Code tab in the desktop app still bundles Claude Code 2.1.317, so it treats Fable 5.2 as a 200K context model. Use the CLI for long sessions until the desktop app updates.
