Claude Opus 5 Support Lands on OpenClaw Main, but Not in npm Yet

OpenClaw merged full Claude Opus 5 support into main on 24 July. The work adds the model across direct Anthropic, Claude CLI, Vertex, Bedrock and related provider routes, then promotes Opus 5 for new Anthropic and Claude CLI setups.

It is a meaningful model-contract change. It is also not in an official npm release at verification time.

The current npm tags remain:

• latest: 2026.7.1-2
• beta: 2026.7.2-beta.4

Both Opus 5 pull requests merged after beta.4 was published at 06:11 UTC on 24 July. A merge on the default branch is not evidence that either installed package contains the change.

The operator decision

Why hand-configuring the name is not enough

PR #113391 identifies a concrete compatibility problem in older OpenClaw model contracts: an unknown claude-* identifier could be treated like a pre-4.6 model. For Opus 5, that meant OpenClaw could construct requests using incompatible thinking or sampling parameters.

The merged implementation does more than add a picker entry. It adds:

• adaptive thinking by default with the supported effort range;
• removal of sampling parameters that Opus 5 rejects;
• assistant-prefill stripping;
• refusal-aware streaming behaviour;
• model-bound thinking replay;
• catalog metadata for a 1M-token context window and 128k output; and
• provider-specific request and catalog handling.

Those are runtime contracts. Adding a custom model string to an older package does not backport them.

Defaults change only for new setup paths

PR #113392 completes the rollout by advancing the generic opus alias, new Anthropic onboarding, the Claude CLI default and media-understanding defaults to Opus 5. It also adds first-run and migration coverage.

The rollout pull request says existing explicit Opus 4.8 configurations continue to work. A follow-up documentation and test merge, PR #113413, makes the other side of that contract unambiguous: the bare opus family alias is rolling. That distinction matters:

• a fresh environment created after the supporting release may select Opus 5;
• an existing environment pinned to Opus 4.8 should not be silently rewritten merely because Opus 5 exists; and
• an environment configured with bare opus is expected to move to the current Opus generation during an upgrade; and
• fleets that rely on implicit defaults can diverge across installation dates.

The added regression coverage separates bare opus, versioned opus-5, pinned opus-4.8, and retired model references. It records that a retired claude-opus-4-5 reference upgrades to Opus 5, while an explicit claude-opus-4-8 reference remains untouched.

Interpretation: model defaults and family aliases are now part of deployment reproducibility. If two nodes must behave and cost the same, use an explicit versioned model identifier rather than relying on whichever family alias or onboarding default a release supplies.

A same-day Claude CLI regression was fixed on main

Advancing the packaged Claude CLI default briefly broke route detection for configurations pinned to older seeded Claude models. The affected main code probed only the new default model ID, so a valid route such as anthropic/claude-opus-4-8 could stop advertising Claude session creation. Session adoption could also stamp the packaged default instead of the model the operator had routed.

PR #113424 merged at 23:04 UTC, before this update was verified. It changes route detection to probe the seeded Claude CLI model IDs and reuse the model actually routed by the operator. Upstream reports the focused Anthropic suite passing after the fix.

This was a default-branch regression and fix, not evidence of an affected npm package: both changes landed after the current beta.4 package. Operators building arbitrary main commits should use a revision at or after merge commit 866012f5be1921fe0b725f8d84944b09196aef12; packaged users should still wait for release notes that include the complete Opus 5 rollout and follow-up fix.

Cost and capacity need a fresh acceptance test

The merged catalog records direct Anthropic pricing of $5 per million input tokens and $25 per million output tokens, with separate cache rates. Those values are upstream metadata sourced from Anthropic, not a forecast of an operator’s actual bill. Cloud-provider pricing and region handling can differ.

The larger advertised context and output limits also do not make every workload sensible at maximum size. Before production adoption:

1. run representative prompts through the exact provider route;
2. confirm thinking and output budgets do not exceed your cost controls;
3. test a refusal and a tool-use turn, not just plain text;
4. inspect fallback routing so a failed Opus 5 request cannot silently move to an unintended model; and
5. compare a fresh setup with an upgraded existing setup to catch default drift.

Evidence and limits

Anthropic’s live model overview lists Claude Opus 5. OpenClaw PR #113391 merged at 20:47 UTC and the completion rollout, PR #113392, merged at 21:55 UTC on 24 July. The Claude CLI route fix merged at 23:04 UTC, and the rolling-alias documentation and regression coverage merged at 23:14 UTC. Upstream reports focused contract suites, a direct-Anthropic live smoke test and exact-head CI for the rollout.

OpenClaw Academy verified the primary pull-request records, merge chronology, Anthropic’s model listing and npm distribution tags at 23:20 UTC. We did not build main, send a paid Opus 5 request or reproduce the cloud-provider routes. The behavioural details above are therefore attributed source analysis until a package ships and is independently exercised.

Bottom line

Opus 5 support is real on OpenClaw main, and it is broader than a catalog update. The safe move for packaged users is to wait for a release that names the two merges—not to force the new model identifier into beta.4 or stable.

When that release arrives, pin defaults where consistency matters and test the exact provider path against cost, thinking, fallback and refusal behaviour before rolling it across a fleet.