Claude API Migration: Treating API Version 2026-10-21 And The New SDK Majors As A Change, Not A Chore

A Claude API migration has three parts that move at once this month: the API version your requests send, the SDK major version that sends it, and the licence the SDK ships under. Pin all three, test the new combination on your own cases, get legal to look at the licence change, and switch with a named approver. The gate below makes that the only path to production.

What did Anthropic release on 21 October?

Anthropic's release notes for 21 October 2026 list API version 2026-10-21 and new SDK versions that send it by default: Python SDK 2.0.0, TypeScript SDK 1.0.0, C# SDK 13.0.0, Go SDK 2.0.0, Java SDK 3.0.0, PHP SDK 1.0.0, Ruby SDK 2.0.0 and ant CLI 2.0.0. They state that the new SDK and CLI versions also have breaking changes that do not depend on the API version, and that they are licensed under Apache 2.0 instead of the MIT License. A migration guide sits alongside the notes.

The Python SDK moved quickly after that: 2.1.0 on 21 October and 2.2.0 and 2.3.0 on 23 October, according to its GitHub releases. Read the migration guide for the specific changes that affect your code before you upgrade.

What does a safe Claude API migration look like?

How do you gate the SDK upgrade in CI?

"""Gate an Anthropic SDK major upgrade behind a change record.

On 21 October 2026 Anthropic released API version 2026-10-21 and new SDK majors that send it by default
(Python SDK 2.0.0, TypeScript SDK 1.0.0). The release notes say the new SDKs also have breaking changes
that do not depend on the API version, and are licensed under Apache 2.0 instead of MIT.
Python 3.14 standard library only. Reads a requirements file, does not install anything.
"""
from __future__ import annotations

import re

REQUIREMENTS = """
anthropic==2.3.0
httpx2>=2.10
"""
CURRENT = {"anthropic": "1.14.0"}  # what production runs today
CHANGE = {"id": "CHG-3120", "eval_gate": "passed", "licence_review": "pending", "api_version": "2026-10-21"}


def pins(text: str) -> dict[str, str]:
    out = {}
    for line in text.splitlines():
        m = re.match(r"^([A-Za-z0-9_.-]+)==([0-9][^\s;]*)", line.strip())
        if m:
            out[m.group(1).lower()] = m.group(2)
    return out


def check(text: str) -> list[str]:
    found, problems = pins(text), []
    if "anthropic" not in found:
        problems.append("anthropic is not pinned to an exact version")
        return problems
    new_major = int(found["anthropic"].split(".")[0])
    old_major = int(CURRENT["anthropic"].split(".")[0])
    if new_major > old_major:
        if CHANGE.get("eval_gate") != "passed":
            problems.append("major upgrade without a passed evaluation")
        if CHANGE.get("licence_review") != "approved":
            problems.append("licence changed from MIT to Apache 2.0: legal review not approved")
        if CHANGE.get("api_version") != "2026-10-21":
            problems.append("change record does not name the API version the new SDK sends")
    return problems


if __name__ == "__main__":
    issues = check(REQUIREMENTS)
    print(f"{CHANGE['id']}: {'ready' if not issues else 'blocked'}")
    for i in issues:
        print("  -", i)
    CHANGE["licence_review"] = "approved"
    print(f"{CHANGE['id']} after review: {'ready' if not check(REQUIREMENTS) else 'blocked'}")

The requirements pin anthropic 2.3.0 while production runs 1.14.0, so the gate sees a major bump. The change record shows a passed evaluation and names API version 2026-10-21, but the licence review is still pending, so the change is blocked. Once legal approves, it is ready. Run it as a required CI check on every pull request that touches dependencies.

What breaks in production?

• **Unpinned dependencies.** A range such as anthropic>=1 picks up 2.x on the next build. Pin exact versions.
• **Mixed versions across services.** Two services on different SDK majors send different API versions. Upgrade together or record why not.
• **Licence surprises.** Procurement and legal track licences. A change from MIT to Apache 2.0 is usually simple to accept, but it should be recorded.
• **Behaviour drift.** A new API version can change responses without errors. Compare outputs on golden cases, as in our LLM evaluation playbook.

What are the alternatives to upgrading now?

Our blog on AI model testing explains to non-technical colleagues why upgrades need testing, and our playbook on Claude email triage shows a tool loop that will need re-testing.

Where should a team start?

List every service that imports an Anthropic SDK, record its version, and open one change record for the move to 2026-10-21. Our AI Automation London team runs upgrades like this for the agents we build, such as the daily mismatch report, and a 14-day Proof Run includes the evaluation gate.