Claude Opus 4.8 support shipped in 2026.5.28. To use it, point your model config at the Opus 4.8 id through your Anthropic auth, or set it per session with /model. The model loads through the same provider path as earlier Claude releases, so if Opus 4.6 or 4.7 already worked for you, the upgrade is a one-line change.
That’s the short version. The longer answer is about when Opus 4.8 earns its price tag in an always-on agent, because running a flagship model on every heartbeat is the fastest way to a surprising API bill. This guide covers the setup, the config shapes that actually work, and a routing pattern that keeps Opus 4.8 on the turns where it matters.
Table of contents
- What Claude Opus 4.8 is
- Setting Opus 4.8 as your default model
- Per-session and per-task overrides
- When Opus 4.8 is worth it (and when it isn’t)
- Cost control patterns
- Troubleshooting
- FAQ
What Claude Opus 4.8 is
Opus 4.8 is Anthropic’s newest model in the Opus line, following 4.6 and 4.7. It targets the same workloads the Opus tier has always handled well: long multi-step reasoning, code, and tasks where a wrong answer costs more than the extra tokens. For an agent that plans across tools, sessions and channels, that reasoning depth is the whole reason to reach for an Opus model instead of a mid-tier one.
added 4.8 in the 2026.5.28 release, in the same batch as Fal Krea image schemas, NVIDIA featured model catalogs and MiniMax streaming music responses. The provider plumbing matters here: because resolves Anthropic model ids through its provider layer, support for a new Opus version is mostly a catalog and id change, not a new integration. Bare direct Anthropic model ids resolve without a cached catalog, which is why the upgrade path is so short.
If you’re new to how picks and routes models, the complete guide and how works cover the runtime model before you start tuning it.
Setting Opus 4.8 as your default model
Two things have to be true: your Anthropic credentials are configured, and your model setting points at the Opus 4.8 id.
If you already run any Claude model in , the credential half is done. To switch the default, set the model in your config and restart the agent:
# set the default model for the agent
/model claude-opus-4-8If you authenticate through a Claude CLI OAuth profile rather than a raw API key, loads those OAuth overlays for your auth profile, so the same /model change applies. The 2026.5.27 release also rewrote legacy api_key auth profiles to a canonical form, which means older configs keep working after an upgrade without a manual migration.
A few setup notes that save time:
- Use the exact model id exposes. Bare Anthropic ids resolve directly, but a typo silently falls back or errors at first turn.
- Restart after changing the default so heartbeats and cron jobs pick up the new model, not just your next interactive message.
- Check
doctorif a turn fails immediately after switching. Recent releases made restart guidance actionable and bound OAuth/token requests, so a stuck auth state surfaces faster than it used to.
Per-session and per-task overrides
Setting Opus 4.8 as the global default is usually the wrong move for an always-on agent. The better pattern is a cheap default with Opus 4.8 as an override for the turns that need it. supports per-session model overrides, so you can switch mid-conversation:
# daily driver stays cheap
/model deepseek-v3
# escalate a hard task to Opus 4.8 for this session
/model claude-opus-4-8This is the same hybrid approach we walk through in Best Cheap LLMs for : a budget model for routine messaging and heartbeats, a flagship model only when the task justifies it. Sub-agents can carry the stronger model while the main loop stays cheap, so a single complex request doesn’t reprice your entire day.
Here’s the decision in table form.
| Workload | Suggested model | Why |
|---|---|---|
| Heartbeats, status checks | Cheap or local model | High volume, low stakes |
| Inbox triage, routine replies | Mid-tier model | Good enough, much cheaper |
| Multi-step planning, code, research | Claude Opus 4.8 | Reasoning depth pays off |
| Privacy-sensitive bulk work | Local model via Ollama | Zero API cost |
When Opus 4.8 is worth it (and when it isn’t)
Opus 4.8 is worth it when the cost of a wrong or shallow answer is high: refactoring code, planning a multi-tool workflow, or research where the agent has to hold a lot of context and reason over it. On X, early users testing 4.8 are mostly comparing it against 4.6 and 4.7 for exactly these harder tasks rather than for everyday chat, which matches what we’d expect from the Opus tier.
It is not worth it for the bulk of an agent’s day. A typical active setup runs 100 to 200 turns daily across heartbeats, channel messages and scheduled jobs. Most of those turns are simple. Paying flagship prices for “acknowledge this message” or “check if anything changed” burns budget with no quality gain. That’s the trap the override pattern above is built to avoid.
A reasonable rule: if you could not tell the difference between a cheap model’s answer and Opus 4.8’s answer on a given turn, the cheap model should handle that turn.
Cost control patterns
Once Opus 4.8 is an override rather than a default, three habits keep spend predictable:
- Route heartbeats to a cheaper model. Heartbeat-heavy configs are the single biggest source of silent cost. Keep them on a budget or local model.
- Escalate inside sub-agents. Let the main loop stay cheap and spawn an Opus 4.8 sub-agent for the hard part of a task, so the expensive model only touches the expensive work.
- Watch the turn count, not just the per-token price. Volume times price is the bill. A slightly pricier model on 10 turns beats a cheap model on 500.
We cover the full playbook, including caching and prompt trimming, in How to reduce your API costs. The model choice is one lever; turn volume is the bigger one.
Troubleshooting
The model id won’t resolve. Confirm you’re using the exact id exposes for Opus 4.8 and that your Anthropic auth is active. Bare direct Anthropic ids resolve without a cached catalog, so a resolution failure usually points at auth, not the catalog.
It worked interactively but heartbeats still use the old model. You changed the session model but not the default, or you didn’t restart. Set the default in config and restart so background runs adopt it.
Auth broke after an upgrade. Legacy api_key profiles are rewritten to canonical form on recent releases, and Claude CLI OAuth overlays load for auth profiles. If a profile is stuck, run doctor for the now-actionable restart guidance.
FAQ
Which version added Claude Opus 4.8? Support landed in 2026.5.28, alongside new image, video and music provider updates.
Do I need a new integration to use Opus 4.8? No. resolves Anthropic model ids through its provider layer, so a new Opus version is a model-id change, not a new integration. If an earlier Claude model worked, 4.8 works with a one-line switch.
Should I make Opus 4.8 my default model? Usually not for an always-on agent. Use a cheap default and override to Opus 4.8 per session or per sub-agent for hard tasks. Running a flagship model on every heartbeat is the main way agent costs balloon.
Can I use Opus 4.8 through a Claude subscription instead of an API key?
loads Claude CLI OAuth overlays for auth profiles, so OAuth-based auth works in addition to raw API keys. Configure the profile, then point /model at the Opus 4.8 id.
How do I keep costs down while still using Opus 4.8? Route heartbeats and routine turns to cheaper or local models, escalate only hard tasks, and track total turn volume. See our cost reduction guide for the full approach.
Opus 4.8 is a one-line upgrade in , but the real decision is routing, not installation. Set a cheap default, keep Opus 4.8 for the turns where reasoning depth changes the outcome, and your agent gets smarter on the hard work without getting expensive on the easy work.
Sources: