Claude Opus 5.5 Guide: Use It Today [Hands-On]

Long agent runs on the last Opus tier burned budget fast. The model talked past the finish line or stalled on scope. Claude Opus 5.5 shipped September 22, 2026 – first of the Claude 5.5 family – aimed at that exact pain: multi-hour coding and knowledge work with clearer progress, higher five-hour limits on Pro/Max/Team, and typical token workloads about 40% cheaper than Opus 5 (Anthropic launch figures).

This is not a benchmark dump. It’s how you switch today in the apps, Claude Code, or the API, fire one real multi-step job, and skip the silent failures that show up in hour one.

What actually changes in daily work

1M-token context. 128K max output on regular Messages. Adaptive thinking you cannot turn off. Default effort: medium. Model ID: claude-opus-5-5. Knowledge cutoff June 2026 – all per Anthropic’s model overview as of release.

$4 per million input. $20 per million output. Cache reads at $0.20 per million – the line that dominates overnight agents. Cache writes $5 (5 m). Output generation lands more than 30% faster in their tests. Fast mode (research preview in Claude Code / Platform) can hit about 2.5× speed at $8 / $40 if you want snappier interactive turns.

Picture a senior engineer who plans before talking, finishes a sprawling migration without constant nudges, and leads with the answer instead of a TED talk. You still draw the finish line. You stop babysitting “think carefully” prompts.

First useful session – pick one path

One path only until the model feels normal.

Path A – Claude.ai / desktop

1. Open a chat on Pro, Max, Team, or Enterprise (free usually won’t show full Opus).
2. Model picker beside send → Claude Opus 5.5.
3. Effort → leave medium. High/xhigh only after medium stalls on hard reasoning.
4. Paste one complete brief: task + done criteria + stop rules. Send. No drip-fed micro-steps.

Path B – Claude Code

claude --model claude-opus-5-5
# or inside a session:
/model claude-opus-5-5
/effort medium

Permanent default if you want it: claude config set model claude-opus-5-5. Interactive wait-on-each-reply? Try /fast – higher token price, quicker turns.

Path C – API shape

client.messages.create(
 model="claude-opus-5-5",
 max_tokens=16000,
 # adaptive thinking is always on - do not send type: disabled
 messages=[{
 "role": "user",
 "content": "Migrate payment endpoints to the new client.nDone means: every endpoint uses the new client, old client deleted, tests pass.nStop and ask only if a test fails for a reason you can't explain."
 }]
)

Pass thinking blocks back unmodified in tool loops. Append-only history is the safe habit – some newer accounts get ugly errors when you rewrite around those blocks.

Pro tip: Strip every “think step by step / think carefully” line from system prompts and CLAUDE.md. The model already thinks before each reply; those phrases mostly add latency in Anthropic’s own chat testing.

Long Claude Code run? Short rule in CLAUDE.md: keep going when no input is needed; stop only before destructive actions or true blockers. Have it maintain a live checklist file (TASKS.md works) so progress survives context summarization.

Ever watched an overnight agent chew a month’s budget while you slept badly? That anxiety is exactly why the cache-read rate matters more than a pretty input sticker price – and why a crisp definition of done beats another clever system prompt.

Pitfalls that waste day one

• Disabling thinking. HTTP 400. Depth only moves with effort (low → max). Medium is the intentional default – start there.
• Forced tool_choice (any/tool style). Errors out. Auto-style selection only; use structured outputs when you need schema-valid JSON. Old Opus 5 tool loops that forced a tool will break on migrate.
• Quiet tool loops. Progress text between tools now sits inside thinking blocks. Default display can hide it, so streaming UIs look frozen until you surface summaries/updates.
• Safety fallbacks. Bio/cyber/frontier classifiers can swap you to an older model mid-chat. Trigger can be an earlier file, memory, or search hit – not just the last message. Apps show a “Switched to…” notice; fresh chat or model picker. Code: /model or Esc Esc. Don’t demand raw internal chain-of-thought dumps – that path refuses or flags.
• Vague “done.” Long runs invent extra work. Name the acceptance tests or the artifact.

Early community notes: self-review looks stronger than older Opus, still not gospel. Read the diff. Run the app before you merge.

Against the obvious alternatives

If the bill is mostly cache reads on overnight agents, $0.20 cache reads are what actually moves the bill – not a modest sticker drop on raw input alone. Full pricing tables: Anthropic’s Opus 5.5 page.

Available as of launch on Claude Pro/Max/Team/Enterprise apps, the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry – seat and region still decide what the picker shows.

FAQ

Is Claude Opus 5.5 on the free plan?

No. Full Opus sits on Pro, Max, Team, Enterprise, plus API and the major clouds. If the picker is empty, it’s the plan – not a bug.

What effort level should I use first?

You’re renaming symbols across a service and fixing call sites. Medium is overkill; low is enough. You’re untangling a gnarly auth migration and medium keeps looping the same dead end – then high. Reserve xhigh/max for work you’d previously throw at the top capability tier, and only after a quick side-by-side on your own use. Higher effort burns limits faster even when per-token rates look friendlier.

Will my Opus 5 API code break?

Plan on yes until you retest. The mental model shift is the trap: people copy old flags that disabled thinking or forced a tool, then chase 400s that look like auth failures. Thinking blocks also changed shape for between-tool progress, so a streaming UI that assumed plain text goes quiet and operators blame “hangs.”

Swap the model ID, drop disable-thinking flags, stop forcing tools, pass thinking blocks through untouched, retest the stream. Migration notes plus the Opus 5.5 usage playbook are the checklist – one golden path in staging before production traffic.

Open Claude or your terminal. Select claude-opus-5-5. Paste one complete task with a crisp definition of done. Let medium effort run. That’s it – one complete task, medium effort.