On September 22, 2026 at 16:38 UTC, Anthropic published Claude Code v2.1.280 with three changes that affect every Claude Code operator and a fourth that affects anyone running agent stacks on Pro or Team Standard plans.
The headline changes, verbatim from the v2.1.280 release page:
1. Opus 5.5 becomes the default Opus model. Added Claude Opus 5.5 (claude-opus-5-5), now the default Opus model — 1M context, $4/$20 per Mtok with $0.20/Mtok cache reads. 2. Plan-tier default change. Changed the default model on Pro and Team Standard plans from Sonnet to Opus, matching Max, Team Premium, and Enterprise. 3. Symlink-write security gap closed. Fixed writes through a symlinked path being judged by their in-tree spelling: the prompt names where the write lands, and acceptEdits, allow rules and auto mode no longer approve one landing outside. 4. New env var. Added CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH to change the 2,048-character cap on MCP tool descriptions and server instructions for every MCP server in the session.
Two additional auto-mode reliability fixes ship in the same release (denial-on-decline; backoff-then-stop on no-answer after ten retries), and two prompt-cache preservation fixes (host-app model-switch cache miss; resumed fork subagent tool-list rebuild).
Opus 5.5 pricing. Anthropic's pricing page now shows Claude Opus 5.5 at $4 input, $20 output per MTok, with a 0.05x cache multiplier (footnote 2: "Cache hits and refreshes on Claude Opus 5.5 are priced at 0.05x the base input price") — $0.20/MTok cache hits and refreshes. That is strictly cheaper than Claude Opus 5 ($5/$25 with $0.50 cache at the standard 0.1x multiplier) and Claude Opus 4.x ($5/$25/$0.50). Same 1M context window as Fable 5.1 / Mythos 5.1.
| Model | Input $/MTok | Output $/MTok | Cache hits $/MTok | Cache multiplier |
|---|---|---|---|---|
| Claude Opus 4.7/4.8 | 5.00 | 25.00 | 0.50 | 0.1x |
| Claude Opus 5 | 5.00 | 25.00 | 0.50 | 0.1x |
| Claude Opus 5.5 (new) | 4.00 | 20.00 | 0.20 | 0.05x |
| Claude Fable 5.1 | 10.00 | 50.00 | 0.25 | 0.025x |
Symlink-write fix. Until v2.1.280, a Claude Code session with acceptEdits, an allow rule, or auto mode enabled could be steered into writing a file outside the approved working directory by a model that used a symlink whose in-tree spelling matched an approved path. The new behavior: the prompt names the actual destination of the write (where the symlink resolves), and acceptEdits / allow rules / auto mode no longer approve a write that would land outside the working tree. This is a sandbox-escape-class fix that operators running untrusted-agent workloads against acceptEdits or auto mode should treat as a required upgrade.
CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH. New environment variable; previously the 2,048-character cap on MCP tool descriptions and server instructions was hardcoded. Operators with MCP servers that publish long descriptions (tool-callable APIs whose tool schemas include lengthy parameter docs) can now raise the cap per session via env var.
Plan-tier default change. Pro and Team Standard plans now default to Opus. Max, Team Premium, and Enterprise were already on Opus defaults per the v2.1.280 release-note wording. Operators who explicitly selected Sonnet or any other model on those plan tiers retain their selection; the change affects the default for new sessions and unset sessions.
Four operational questions land on Claude Code operators with this release:
1. AcceptEdits and auto mode upgrade. If you run Claude Code with acceptEdits, an allow rule for a directory tree, or auto mode on, update to v2.1.280. The symlink-write fix closes a real exploit class; staying on v2.1.279 means a model that follows a symlink could write outside the approved tree without triggering approval. 2. Cost re-model for Opus 5.5. At $4/$20/$0.20 cache, Opus 5.5 is the cheapest Opus-tier model Anthropic has shipped. Agent stacks that route Opus workloads on max effort need to recalculate spend; a single-turn Opus 5.5 max-effort task with 200K input + 50K output at the standard cache-hit rate is approximately $1.80, vs $1.95 on Opus 5 — a small per-turn saving but material at fleet scale. 3. MCP server description cap. Operators with MCP servers shipping longer-than-2,048-character tool descriptions were previously truncated silently. If your MCP server hit that cap, raising it via CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH may surface parameter details that were previously invisible to the model. 4. Auto mode reliability. The "retry forever on safety-check decline" and "deny forever on safety-check no-answer" fixes both materially affect long-running auto-mode sessions. The turn now stops after ten no-answer retries with a message; previously the loop would consume tokens indefinitely.
This is a documentation comparison. Primary sources verified verbatim at fetch 2026-09-22T20:09–20:10 UTC:
No first-hand test of the symlink-write fix, the auto-mode retry/backoff behavior, or the new default-model change was executed in this sweep.
CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH. No documentation note on maximum safe value.The symlink-write fix is the headline of v2.1.280 for security-conscious operators; it should be treated as a required upgrade, not a feature drop. Opus 5.5 as the new default on Pro and Team Standard plans changes the cost shape of the most common Claude Code deployments; operators on those plans will see slightly different per-token spend immediately on the upgrade. CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH is a long-requested operational knob that should have shipped sooner. The auto-mode retry and no-answer fixes close two long-running reliability gaps.
Today: Upgrade any Claude Code install running acceptEdits, allow rules, or auto mode to v2.1.280. Re-pull the pricing model for Opus-class workloads; Opus 5.5 is the new default on Pro and Team Standard plans.
This week: Audit any MCP server deployments where tool descriptions exceeded 2,048 characters. If you previously truncated silently, raise the cap via CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH and verify the model now sees the full description.
Skip if not in scope: If you run Claude Code on Sonnet with explicit model pinning, the plan-tier default change does not affect your sessions. If you do not run auto mode or acceptEdits, the symlink-write fix is a defense-in-depth update rather than a behavior change for your workflow.