- Opus 5.5 is the model, released 22 September 2026. The API ID is
claude-opus-5-5 - Claude Code is the product: a terminal agent. It can call Opus 5.5, and also Sonnet or Haiku
- API list price: $4 input / $20 output per million tokens. Cache reads are $0.20
- A subscription is not invoiced from that card. Pro, Max, and Team use a rolling five-hour window plus a weekly limit, shared by chat and the terminal
- Context is 1 million tokens. Synchronous output caps at 128K. Thinking stays on. Default effort is
medium - Most day-to-day coding is fine on Sonnet. Keep Opus for hard refactors and overnight agents. Do not park a long process on a laptop that sleeps
Searching “Opus 5.5 or Claude Code” usually starts from the wrong premise. The two names are not on the same layer. One is weights and a rate card. The other is a shell that reads a repository, runs commands, and edits files. Treating them as two mutually exclusive coding models mis-counts both the bill and the quota.
This piece follows Anthropic documents that were still current on 25 September 2026: the launch note, the model page, subscription pricing, and the Claude Code usage article. Prices and limits follow the console and /model. If the docs move, the official page wins.
Two names: the model and the shell
Claude Opus 5.5 is the first model in the Claude 5.5 family, aimed at long-running agentic coding and knowledge work. It is on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. The synchronous ID is claude-opus-5-5; on Bedrock it is anthropic.claude-opus-5-5. Anthropic’s retirement commitment is not sooner than one year after 22 September 2026.
Claude Code is the terminal agent included in every paid plan. Anthropic states that it shares one usage pool with the web, desktop, and mobile apps. Inside Claude Code you pick a model. You do not buy a second model named Claude Code.
| Question | Opus 5.5 | Claude Code |
|---|---|---|
| What it is | A model | A terminal agent shell |
| What you buy | Per-token API, or a model tier inside a plan | A session that reads the repo, runs tools, edits files |
| Can it stand alone | Yes. API, Bedrock, Cursor can call it | No. It always calls some model |
| What the bill looks like | Input, output, cache write, cache read | A subscription window, or an API key you attach |
Earlier notes: Opus 4.8 as a coding agent and Fable 5 and Opus 4.8. Version 5.5 changes the rate card and the step count, not “buy another CLI.”
Price: a token rate card and a subscription window
API: $4 / $20, and cache reads are the bulk
Against Opus 5, input and output are each 20% lower: $4 and $20 per million tokens. A 5-minute cache write is $5, a 1-hour cache write is $8, and a cache read is $0.20, 60% below Opus 5’s $0.50. Anthropic’s own tests say default settings cost about 40% less than Opus 5 on typical workloads, because cache reads dominate agent bills and 5.5 uses fewer tokens on the same kind of task. Output is more than 30% faster than Opus 5.
| Per million tokens | Opus 5.5 | Opus 5 |
|---|---|---|
| Input | $4 | $5 |
| Output | $20 | $25 |
| Cache write (5 min / 1 hour) | $5 / $8 | $6.25 (5-minute tier on the launch table) |
| Cache read | $0.20 | $0.50 |
| Batch API | 50% off input and output: $2 / $10 | Batch discount at the then-current rate card |
| Fast mode | $8 / $40, up to about 2.5× speed | Not this 5.5 rate card |
The shortest cacheable prompt is 512 tokens. API keys and cloud accounts can set a one-hour cache; subscribers already had that lifetime. Forked sub-agents can keep reading the parent cache instead of buying the same repository context again. Anthropic’s three operating notes for long sessions are concrete: pick the model at the start and switch less often; compact before you step away; set the cache lifetime to one hour for long jobs. In Claude Code, /usage shows how much of the session is cache reads.
Fast mode is available in Claude Code and on the Claude Platform, listed at up to about 2.5× speed, at double the standard unit price. It buys latency, not a smarter second model.
Pro, Max, Team: not invoiced per token
Do not multiply the subscription price by the API rate card. The help center says Pro and Max usage is included in the monthly fee, so a dollar figure on a session is not an invoice. Team and Enterprise seat allowance is not metered in dollars either. $4 and $20 appear on Console and cloud-provider accounts.
- Pro: $20 a month, or about $17 a month on annual billing. Each five-hour window is at least 5× Free. Claude Code is included.
- Max 5x / 20x: $100 or $200 a month. Relative to Pro, each five-hour window is 5× or 20×, with a higher output limit.
- Team: a Standard seat is above Pro; Premium is 5× Standard. Seat prices on the pricing page are the source of truth. The commonly published cards are Standard $25 (about $20 annually) and Premium $125 (about $100 annually).
Every plan has a rolling five-hour window; paid plans add a weekly limit. Length, model, and features change how much fits in a window. There is no fixed message count. At the cap you wait, upgrade, or turn on usage credits and continue at standard API rates. The 22 September launch also raised five-hour limits on Pro, Max, and Team, and gave subscribers one rate-limit reset they can save and spend when they choose. /clear (alias /reset) clears the conversation. It does not spend that reset, and it does not refill the five-hour or weekly allowance.
Tokens: context, thinking, and effort
The model page fits on one line: 1 million tokens of context, 128K synchronous max output, reliable knowledge cutoff and training cutoff both June 2026. On the current tokenizer, 1 million tokens is about 555,000 English words. The Message Batches API can go to 300K output tokens with the header output-300k-2026-03-24. Input is text and images; output is text.
Adaptive thinking on Opus 5.5 is always on and cannot be disabled. Depth is the effort parameter. The API default is medium. Thinking tokens are billed as output. The help center therefore warns that leaving Opus on for renames and log lines is the fastest way to drain a daily allowance. Most published benchmarks are at max effort; the Terminal-Bench 4.0 high score for Opus 5.5 uses xhigh. FrontierCode at default medium (54.6%) and the max-table 54.4% are not the same setting. Quote the effort, or you will describe the cheaper default as an unconditional top score.
Three behavior changes carry over from Opus 5 and match Fable 5.1: thinking cannot be turned off; forced tool choice returns an error; thinking blocks are tied to the model and the conversation. On the Claude API and Google Cloud, the older computer_20251124 computer-use tool is no longer accepted. Text between tool calls lands in thinking blocks, and at the default display that text is empty. A client that streams inter-tool text as a progress bar goes quiet between calls until display is set.
Coding: fewer steps, not only a higher score
The launch post puts Opus 5.5 in front on agentic coding, and also says that at this level benchmark gaps are a weaker guide to real differences. In Anthropic’s own use, the gap to Fable 5.1 is narrower than the scores suggest. The figures below are from the official table, with the stated conditions. Runs used production safeguards. When a safeguard intervened, cybersecurity tasks were finished by Opus 4.8, and biology or frontier-model-development tasks by Opus 5, which lowers 5.5’s score on those items.
| Benchmark | Opus 5.5 | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% (xhigh, SE ±2.6) | Fable 5.1 55.8%; Opus 5 52.3%; GPT-6 Astra 57.9% (OpenAI, high) |
| FrontierCode v1.1 main | 54.4% on the max table; 54.6% at default medium | GPT-6 Astra top score 53.3%. At default effort Anthropic puts the cost near one fifth |
| CursorBench 4.0 | 57.8% on the max table; 52.5% at medium | Medium still above Fable 5.1 max (51.8%) and Opus 5 max (46.6%) |
| OSWorld 2.0 | 81.8% partial | Fable 5.1 80.7%; Opus 5 74.0% |
The number worth remembering at purchase time is the step count. In early testing, an audit and fix of a 200,000-line codebase took under three hours on 5.5; Opus 5 took over 20 hours and 2.5× the tokens. An internal translation of HAProxy from C to Rust: both passed nearly all regression tests; 5.5 took 9.5 hours, Fable 5.1 took 12, at 51% lower cost. On a public command-line benchmark cited from the Kiro side, 5.5 made about 40% fewer calls and used about half the tokens versus Opus 5, while solving more tasks. GitHub’s tests in Copilot CLI and VS Code also reported terminal tasks in less than half the steps.
Those are vendor and early-customer tests, not an SLA for your repository. The useful inference is one sentence: on the same Opus tier, 5.5’s advantage is fewer round trips and cheaper cache reads. If the task is single-file completion, the leaderboard does not help. Sonnet 5 ($2 input / $10 output per million, default effort high) is usually the better fit.
Agent workflow: how the model sits inside Claude Code
The help center split is still clear: Sonnet for most coding; Opus for cross-cutting refactors, hard debugging, and architecture; Haiku for lookups, simple edits, and high-volume scripts. Opus spends several times the allowance of Sonnet per turn. The recommended path is to plan with Opus and execute with Sonnet, which is the built-in /model opusplan. Switching with /model mid-session does not clear the conversation, so the execution model still sees the plan.
Two things not to infer from an alias:
- The
opusalias points at the current default Opus, which after launch is Opus 5.5. It is not a permanent version pin. To lock behavior, use the full model ID and pin it in managed config. - Sub-agents inherit the session model. If you change the default from Sonnet to Opus, a night of subtasks bills as Opus. The help center still says not to leave Opus as the default. Third-party notes on Claude Code 2.1.280 say Pro and Team Standard now open on Opus. Do not guess. Open
/modelon the machine in front of you.
/usage is the plan window. /cost is the dollar figure for an API-key session. Subscribers are not invoiced for that cost figure. When the allowance is gone, an Enterprise seat can temporarily attach an API key if the organization allows it. An individual waits for the window, switches to a lighter model, or turns on usage credits. When a Codex weekly limit is exhausted, see seven ways to keep moving.
The same model choice applies when Claude Code is a video-editing shell. Natural-language editing is covered in video-use with Claude Code and Codex. The long jobs in that piece are the same billing problem as an overnight migration here.
How to choose
| Your week | Bill | Model |
|---|---|---|
| Interactive features, tests, known bugs | Pro at $20/month is usually enough to start | Sonnet. Switch to Opus when stuck, or use opusplan |
| Hours of cross-repo agent work, and you want a capped monthly fee | Max 5x or 20x | Opus 5.5 for the hard stretch, Sonnet for execution |
| CI, batch, replayable overnight jobs that must hit a ledger | API key. Use Batch at half price when the job can wait | Pin claude-opus-5-5, start effort at medium, cache lifetime 1 hour |
| Short tasks, small context, speed first | Haiku on the plan, or Sonnet 5 on the API | Do not pay Fast mode’s double rate for Opus |
| Team seats that share a pool with web chat | Team Standard or Premium | Restrict who can change the default to Opus |
There is no option where Opus 5.5 replaces Claude Code, or the reverse. The accurate sentence is: whether this Claude Code session should be Opus 5.5, and whether that usage draws a subscription window or an API key.
Where a long session should run
The model page describes Opus 5.5 as built for coding that runs a long time. Those sessions usually fail because a laptop lid closes, an index is wiped, or DerivedData and the repo cache disappear overnight. After the allowance resets you still have to feed the million-token context again, and the $0.20 cache-read advantage is gone.
The stabler split: keep the daytime machine for interaction and devices; put the repo, simulators, and the Claude Code process on a Mac that does not sleep, held up with tmux or launchd. Models turn over quickly; the scarce piece is the execution node is that argument. A light single-repo task can start on 16GB. Parallel builds plus sub-agents are calmer on 24GB and a larger disk.
Do not tie an overnight agent to the meeting laptop
Opus 5.5’s cache and shorter step count only count while the process is still alive. Put the repo on a separate Mac so closing the lid does not zero the five-hour window.
FAQ
Are Claude Opus 5.5 and Claude Code the same thing?
No. Opus 5.5 is the model released on 22 September 2026, API ID claude-opus-5-5. Claude Code is the terminal coding agent. It can call Opus, Sonnet, or Haiku. On a paid plan they often show up together, but the bill is not the same.
What is the Opus 5.5 API price?
Per million tokens: input $4, output $20. A 5-minute cache write is $5, a 1-hour cache write is $8, a cache read is $0.20. The Batch API is half price on input and output. Fast mode is $8 input and $40 output, up to about 2.5× speed.
Does Claude Code charge per token?
On Pro, Max, Team, or Enterprise, usage counts against the plan’s rolling five-hour window and weekly limit, not the $4/$20 card. Web, desktop, mobile, and Claude Code share one pool. A Console API key is billed per token. /cost on a session is API spend, not a subscription invoice.
What are the context and output limits?
The context window is 1 million tokens. Synchronous Messages API output caps at 128K tokens. The Message Batches API can reach 300K output tokens with the beta header output-300k-2026-03-24. The reliable knowledge cutoff is June 2026. Adaptive thinking is on by default and cannot be turned off. Default effort is medium.
For everyday coding, Opus or Sonnet?
Anthropic’s help center still recommends Sonnet for most coding, and Opus for cross-file refactors, hard debugging, and architecture. Opus spends meaningfully more of the allowance per turn. A common pattern is /model opusplan: plan with Opus, execute with Sonnet. The machine default is whatever /model shows.
Why put a long agent on a separate Mac?
A long Claude Code session depends on a live process, a cache, and a repo that stays open. Laptop sleep interrupts the job. Put the repo and the terminal on a Cloud Mac that does not sleep. After SSH drops, the process can continue, and the allowance is spent on the model instead of re-reading context.
Closing
Claude Code is where you work. Opus 5.5 is the most expensive tier inside it, and the one that pulls long tasks to a close. On the API, remember 4 / 20 / 0.20. On a subscription, remember the five-hour window and “do not leave Opus as the default without a reason.” The score table says it can take fewer steps. Those steps stick only if the machine stays awake.