Claude Opus 5.5: What Changed, Who Can Use It, and Whether to Try It on Your Next Task

Last updated: September 26, 2026

Claude Opus 5.5 is the newest model in Anthropic’s Opus line, released September 22, 2026. Anthropic says it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. It needs a paid Claude plan or paid API access.

Best for, not for, and when to wait

At a glance

01 What changed, and what Anthropic measured

Anthropic calls Opus 5.5 “a major step up from Opus 5” and says it leads its benchmarks for agentic coding, computer use and knowledge work, writes more clearly, and generates output “more than 30% faster.” Opus 5.5’s benchmark results use maximum effort unless noted, and the launch post warns that “benchmark margins have become a less reliable guide to real-world differences.”

On CursorBench 4.0, run by Cursor, Opus 5.5 at its default medium effort scored 52.5% for about $3 a task; Opus 5 at maximum effort scored 46.6% for $11.95. The Opus 5.5 cost is the system card’s estimate from Cursor’s token counts. FSR has not tested any of this.

02 Who can use it, and what it costs

The Free plan includes neither Opus nor Claude Code. In the US, before tax, Pro costs $20 a month or $200 a year up front (shown as $17 a month), and Max starts at $100 a month; prices vary by region. Pro does not include API use, which is billed separately.

Claude Code 2.1.280 or later defaults to Opus 5.5 on paid plans and the API unless your settings or organization say otherwise. Anthropic’s developer account calls it “the default for paid plans,” but the help pages do not say which model new app chats use; check the model picker.

The 40% describes whole tasks: Anthropic says Opus 5.5 “costs less per token than Opus 5 and uses fewer tokens per task.” But that compares default settings, medium effort against high, and at the same setting Opus 5.5 “tends to think more per turn,” which is billed as output. Cache reads, which Anthropic says make up most agentic and coding costs, fell 60%.

At the same token counts, price alone cuts this cache-heavy example by 40% and uncached work by 20%; real savings depend on how many input, output and thinking tokens Opus 5.5 uses.

03 How usage limits work now

On a Claude plan, the apps and Claude Code share a rolling five-hour limit and a weekly limit, with no fixed message count, so a Claude Code test spends chat allowance too. With an API key, Claude Code is billed per token instead. Higher effort, Anthropic says, means “you’ll reach your usage limits faster.”

Do not add these up. The 25% is what a 20% price cut buys if limits track price (1 ÷ 0.8 = 1.25), but Anthropic’s pages do not say how limits are counted; on Team and Enterprise seats, usage inside the allowance “isn’t metered in dollars,” and Team limits are per member. After Claude Code’s model-specific “You’ve hit your Opus limit,” switching model family with /model keeps you working; after a session or weekly limit, which covers all models, it does not.

04 If you have a limit reset

Anthropic gave “subscription users” a limit reset; its developer account named Pro, Max and Team, to use “any time until Oct 22,” with no time zone given. It appears in Settings > Usage on the web or desktop app, not in the mobile app or Claude Code. No card? Skip this section.

• Your card says whether it refills the five-hour or the weekly limit.
• It cannot be undone and does not move your weekly reset day.
• It does not refund extra usage already billed.
• Downgrading or canceling first forfeits it.
• It covers your whole account.

If no limit blocks you, keep it; it expires at the time on the offer. If the limit it refills blocks urgent work, using it lets you continue, unless the other limit blocks you too. If that limit resets soon anyway, weigh working now against saving it.

05 If you use Claude Code or the API

On Pro or Team Standard, if neither you nor your organization picked a model, Claude Code 2.1.280 and later runs Opus 5.5 instead of Sonnet 5. At API list prices, section 02’s coding session costs about 1.5 times as much ($0.549 against $0.368) and the same uncached work twice as much; Anthropic’s pages do not say how that maps to plan limits. Check /model and /effort to switch back; Anthropic says Sonnet 5.5 will follow “in the coming weeks.”

On Pro, Max and Team, Claude Code starts in auto mode by default, where a classifier reviews actions instead of you, and Plan mode still “runs shell commands to explore.” For a first test, switch to Manual (in the terminal, one Shift+Tab from auto), which asks before most edits, commands and network access unless your permission rules pre-approve them. Use a disposable copy with no production access: /rewind undoes Claude’s direct file edits, not shell-command changes.

On the API, swapping the model ID can break requests: disabling thinking, setting a thinking budget, forcing tool use, non-default sampling parameters and assistant prefill all return a 400 error, and Priority Tier is not supported. Check the migration guide against what your code sends, and keep your current model in production until the code passes in development; Opus 5 remains active, with retirement no earlier than July 24, 2027.

06 A small first test

Reading this needs no account or test. If you do test, set a usage budget first; on Pro or Max, check in Settings > Usage whether usage credits (prepaid, billed at API rates past your limits) and auto-reload are on. FSR suggests one exercise it has not run: turn a public or self-written spec of two to five pages into a brief that separates requirements, open questions and sources.

1. On Pro or Max, check “Help Improve our AI models” in Settings > Privacy before pasting anything; with it off, Anthropic says new chats stay out of training, except feedback you submit and chats flagged for safety review. Team and Enterprise data is not used for training by default, though feedback can be. Paste only what you may share.
2. Write down the improvement you want and the requirements the result must keep.
3. In a new chat, choose Opus 5.5 and leave effort at the menu’s Default.
4. Run the task once, logging any retry or effort change as a separate run, and watch Settings > Usage.
5. Check the result against the source for unsupported claims, missed requirements and citations you cannot trace.
6. Keep the original, save accepted text separately, and compare total effort, checking included, with your current method.

Stop if the output invents facts or misses a must-have requirement, or if a notice says Claude switched models; that reply came from an older model, not Opus 5.5. If your work must stay in the EU, check your contract: the Claude API documents US-only or global inference with US data storage, and cloud providers set their own regions.

07 FAQ

08 Methodology and limits

Starter Guide, Tier C. FSR ran no tests of Opus 5.5, bought nothing and used no limit reset, so nothing here measures quality, speed or cost on real work. The cost examples are fixed-token calculations, not bills, and section 06 is a proposal.

FSR read Anthropic’s launch post, system card, pricing page, help and privacy articles, and Claude Code and API documentation on September 24, 2026 (Asia/Tokyo). FSR’s editor also checked the launch post, both developer-account posts and the pricing page in a browser; from Japan, the pricing page showed higher dollar prices than the US figures used here. Where this briefing says a page is silent, that covers only the pages listed.

FSR researched and drafted this briefing with AI assistants, including Anthropic’s Claude, and placed sources beside the claims they support; agreement between AI tools is not independent verification. This page has no affiliate or paid links.

09 Verdict

The case rests on Anthropic’s results and lower prices; any saving depends on your caching and effort.

10 Corrections and contact

FSR BRIEFING PACKET - Claude Opus 5.5: What Changed, Who Can Use It, and Whether to Try It on Your Next Task
Publisher: Future Stack Reviews (future-stack-reviews.com)
Format: Starter Guide. Evidence depth: Tier C. Public documents and pages only. FSR ran no tests of Opus 5.5 and used no limit reset.
Read: 24 SEP 2026 Asia/Tokyo. Section numbers refer to that article.
Disclosure: no affiliate or paid links. FSR used AI assistants, including Claude, to research and draft the article.

WHAT IT IS
Claude Opus 5.5, released by Anthropic on Sep 22, 2026. Anthropic says it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. It needs a paid Claude plan or paid API access; the Free plan includes neither Opus nor Claude Code.

KEY NUMBERS (do not add them together)
- API list price: $4 input and $20 output per million tokens (Opus 5: $5 and $25). Cache reads $0.20 (Opus 5: $0.50).
- 40%: Anthropic's claim for typical workloads at default settings. Default effort is medium on Opus 5.5 and high on Opus 5. Not measured by FSR.
- 25%: Anthropic's developer account says limits go 25% further because the price is lower. Anthropic's pages do not say how limits are counted.
- +20%: five-hour session limits "in Claude Code", per the same account.
- Against Sonnet 5, the old Claude Code default on Pro and Team Standard: input and output cost twice as much, cache reads the same. Section 05.

WHAT DEPENDS ON YOUR ROUTE
- Apps and Claude Code signed in with a Claude plan share five-hour and weekly limits; with an API key, Claude Code is billed per token. The help pages do not say which model new chats use; check the model picker.
- Claude Code 2.1.280 or later: default model Opus 5.5 at medium effort, unless your settings or organization say otherwise. Pro, Max and Team start in auto mode by default; for a first test, switch to Manual (in the terminal, Shift+Tab once from auto) and use a disposable copy with no production access. /rewind does not undo changes made by shell commands.
- API: disabling thinking, forcing tool use, non-default sampling and prefill return a 400 error. Keep your current model in production until your code passes in development. Opus 5 remains active, with retirement no earlier than Jul 24, 2027.
- Data: the pricing page lists model training as "Opt-out" on Free, Pro and Max, and "None by default" on Team and Enterprise; feedback you send is an exception. The Claude API documents US-only or global inference and US data storage; cloud providers set their own regions.
- Limit reset (offered to Pro, Max and Team users, until Oct 22): one use, no undo; the card says whether it refills the five-hour or the weekly limit.
- Task type: Anthropic says most cybersecurity tasks are re-routed to Opus 4.8, and some biology and frontier AI development requests go to Opus 5. A notice shows when the model switched.

OPEN QUESTIONS - check your own account or ask your admin
1. Which model new chats open with on your plan.
2. How your plan counts Opus 5.5 usage against other models.
3. What your reset card refills, and its exact expiry time.
4. Whether your organization allows Opus 5.5, and which data terms apply.
5. Whether usage credits or auto-reload are on for your account.

FSR VERDICT
Worth one small, budgeted test if you have access to Opus 5.5 and a checkable task your current tool handles poorly. Otherwise keep your current tool, and wait on any route whose terms you cannot confirm.

LIMITS OF THIS PACKET
Prices, limits and defaults change; verify on Anthropic's pages before you act. FSR ran no tests, so nothing here measures quality, speed or cost on your work. Where this packet and the live pages disagree, the live pages win.

Use this briefing to help me decide whether to test Claude Opus 5.5 on my next task. Ask me about my plan, my tools, the task and how I would check the output before you answer.

Starter Guide. Read September 24, 2026 (Asia/Tokyo). Tier C: no tests, no purchase, no limit reset used. No affiliate or paid links. Prices, limits and defaults change; recheck Anthropic’s pricing and help pages before relying on them.