Claude Opus 5.5 vs Sonnet 5.5 vs Haiku 5.5: Which to Use in the Claude App and Claude Code

Short answer: make Claude Sonnet 5.5 your everyday model, switch to Claude Opus 5.5 only when a task is long, messy or expensive to get wrong, and use Claude Haiku 5.5 for quick, clearly described jobs. The one thing most people miss: Claude Code starts on Opus 5.5 by default on Pro, Max, Ultra, Team and Enterprise, so unless you changed it, you are already running the most expensive of the three for every small edit. This guide explains the difference in plain words, shows exactly where to click (Claude app) and what to type (Claude Code), and lists the settings that waste your usage. Everything below was checked on October 31, 2026 against Anthropic’s model overview, pricing page, Claude Code model configuration docs, the Claude Help Center, and an independent test published by OpenRouter on October 30.

Key facts:

  • All three are the same generation. Anthropic released Opus 5.5 on September 22, Sonnet 5.5 on September 28 and Haiku 5.5 on October 7, 2026.
  • All three have a 1-million-token context window and up to 128K tokens of output. (Anthropic model overview)
  • Pay-per-use price per million tokens (input/output): Opus 5.5 $4/$20, Sonnet 5.5 $2/$10, Haiku 5.5 $0.10/$0.50 for prompts up to 100,000 tokens. (Anthropic pricing)
  • Free plan users get Sonnet and Haiku but not Opus. Pro, Max and Ultra include all three. (Claude pricing)
  • Claude Code defaults to Opus 5.5 on Pro, Max, Ultra, Team, Enterprise and Anthropic API accounts. (Claude Code docs)

The three models in one sentence each

Think of them as three sizes of the same brain, not three different products.

  • Claude Opus 5.5 is the big one. Anthropic describes it as built “for long-running agentic coding and knowledge work” and says it “remains clearly stronger at complex, open-ended work requiring sustained judgment.” Use it when you would hand the task to your most senior colleague. (Anthropic)
  • Claude Sonnet 5.5 is the middle one. Anthropic calls it “strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets,” and it generates output more than 30% faster than Sonnet 5. Use it for most work. (Anthropic)
  • Claude Haiku 5.5 is the small, fast one, aimed at “high-volume, latency-sensitive tasks such as classification, extraction, and routing.” Use it for quick, mechanical jobs you can describe precisely. (Anthropic model overview)

There is also a fourth, larger model, Claude Fable 5.1, for the hardest long-horizon work. It is not part of this comparison: on Pro it is paid with usage credits, and on Max and Ultra it is capped at half your weekly limit (Claude pricing). If you are deciding between Opus, Sonnet and Haiku, you do not need Fable.

How much better is Opus, really?

Less than the price gap suggests on everyday work. In Anthropic’s own launch table for Sonnet 5.5, the two are close on most tests, and Sonnet even wins one:

Test (Anthropic’s numbers)Sonnet 5.5Opus 5.5
Terminal-Bench 4.0 (agentic coding)70.6%66.4%
CursorBench 4.0 (coding in Cursor)55.5%57.8%
GDPval-AA v2.1 (real-world office work)18441846
Humanity’s Last Exam, with tools64.5%67.7%
OSWorld 2.1 (operating a computer)80.1%81.8%

Source: Introducing Claude Sonnet 5.5, Anthropic. Haiku 5.5 sits clearly lower on hard agentic coding: Anthropic reports 39.2% on Terminal-Bench 4.0 for Haiku 5.5, against Sonnet 5.5’s 70.6% (see our Haiku 5.5 launch notes).

OpenRouter’s independent test, published October 30, points the same way. It gave all three models 11 coding tasks with hidden tests plus one three-bug fix inside Claude Code. Every model passed every test. What separated them was cost: in the Claude Code bug-fix, the median session cost was $0.098 on Opus 5.5, $0.062 on Sonnet 5.5 and $0.0052 on Haiku 5.5. On its benchmark runs, Sonnet scored 92.4% on GPQA Diamond (science questions) against Opus’s 90.4% and Haiku’s 86.6%. (OpenRouter)

Two honest caveats. First, OpenRouter’s tasks were short and clearly specified, and OpenRouter itself says they “can’t rule out an advantage” for Opus on long, ambiguous work, which is exactly what Opus is built for. Second, OpenRouter sells access to all three models, so treat its “use Sonnet by default” advice as one well-documented data point, not the final word.

Anthropic’s own advice is split by audience, and that is worth knowing. Its developer docs say “start with Claude Opus 5.5 for most workloads” if you are unsure (model overview). Its Claude Code help article for everyday users says “Sonnet is the right choice for the large majority of coding work,” and warns that Opus “uses meaningfully more of your quota” (Claude Help Center). If you are on a subscription and keep hitting limits, the second piece of advice is the one that matters to you.

Price and plan cheat sheet

If you pay per token (for example through a Claude Console API key, or a tool that asks for one), these are Anthropic’s list prices per million tokens:

Opus 5.5Sonnet 5.5Haiku 5.5
Input$4$2$0.10 (up to 100K-token prompts), $0.50 above
Output$20$10$0.50 (up to 100K), $2.50 above
Cache hits (re-reading the same context)$0.20$0.10$0.01 (up to 100K), $0.05 above
Speed (Anthropic’s label)ModerateFastFastest

Sources: Anthropic pricing, model overview.

Watch the Haiku footnote: its price jumps fivefold once a single prompt is longer than 100,000 tokens. In a long Claude Code session the whole conversation is re-sent on every turn, so a long Haiku session can drift past that line without you noticing.

If you use a subscription, you do not pay per token. Instead, the bigger model uses up your 5-hour and weekly limits faster. Anthropic’s Claude Code help page puts it this way: “Opus costs several times more per turn than Sonnet, and Sonnet more than Haiku. Spending Opus on routine work is the fastest way to drain a daily limit.” (Claude Help Center)

Which models each plan includes (Claude pricing):

PlanPriceOpusSonnetHaikuClaude Code
Free$0NoYesYesNo
Pro$20/month ($17/month billed annually)YesYesYesYes
Max$100/monthYesYesYesYes
Ultra$200/monthYesYesYesYes

Which one should you pick? A simple rule

Use this as a starting point, then adjust after a week of real use.

Use Haiku 5.5 when the job is small and you can describe exactly what you want: rename things, fix a typo across files, explain a short error message, sort or tag a list, pull names and dates out of a document, summarise a page, write a quick reply. Anthropic’s Claude Code guide suggests Haiku for “renames, log lines, regex explanations, boilerplate.” (Claude Help Center)

Use Sonnet 5.5 when it is normal work with a clear goal: build a feature you can describe, fix a known bug, write or edit a document, make slides or a spreadsheet, review a draft. This should be your default for most days.

Use Opus 5.5 when you are genuinely stuck or the change is wide: a bug nobody can explain, a big reorganisation of a project, an architecture or strategy decision, a long research or analysis task where a wrong answer costs you hours. Anthropic’s Claude Code help lists “hard debugging, cross-cutting refactors, architecture calls” as Opus territory.

When an answer is wrong, ask why before switching. Anthropic’s own guidance on choosing a model and effort level is a useful test: if Claude had all the context, clearly tried, and was still wrong, pick a bigger model. If it skipped a file, did not run the tests, or gave up halfway, raise the effort level instead (explained below). And if your request was vague, fix the request first; a bigger model does not fix missing context.

How to switch models in the Claude app (web, desktop, mobile)

  1. Open a chat at claude.ai, or in the Claude desktop or mobile app.
  2. Look next to the send button. The current model and effort level are shown there.
  3. Click the model name and pick Opus, Sonnet or Haiku. If you do not see the one you want, click More models.
  4. To change effort, click the model name, then Effort, and pick a level. The level marked Default is the one Anthropic recommends for that model.

You can switch at any point in a conversation; the change applies from Claude’s next reply. (Claude Help Center)

Things to know in the app:

  • Thinking cannot be turned off on Haiku 5.5, Sonnet 5.5 or Opus 5.5 in the Claude app. If you remember a “thinking” toggle from older models, that is why it behaves differently now. (Claude Help Center)
  • Haiku 4.5 is gone from the picker on Free, Pro and Max as of October 29, 2026. If an old chat was on Haiku 4.5, Claude now answers it with Haiku 5.5. Nothing was deleted. (Claude Help Center)
  • Lower effort stretches your usage. Anthropic says Low and Medium “work well for routine tasks and stretch your usage further,” while higher effort takes longer and uses more of your limit.
  • On an Enterprise plan, a missing model or effort level usually means your admin turned it off for your role, and an admin can also force every new chat to start on a set model. You can still switch inside a chat.

How to switch models in Claude Code

You do not need to edit any files. Everything below is typed into Claude Code itself.

Check what you are running now

Type /status. It shows your current model and account. If you have a status line configured, the model also appears there. (Claude Code docs)

Switch with /model

  • Type /model on its own to open a picker. Use the arrow keys to choose, then press Enter to switch and save it as your default for future sessions.
  • In the same picker, press s instead of Enter to switch for this session only, without changing your default.
  • Or type the name directly: /model sonnet, /model opus, or /model haiku. Typing it directly behaves like pressing Enter, so it becomes your new default.

On Anthropic’s own service, opus, sonnet and haiku point to Opus 5.5, Sonnet 5.5 and Haiku 5.5. If you use Claude Code through Amazon Bedrock, Google Cloud or Microsoft Foundry, the same words can point to older versions, so check /status after switching. (Claude Code docs) For Haiku 5.5 specifically, Anthropic’s docs show /model claude-haiku-5-5, which pins the exact version.

You can switch in the middle of a task without losing the conversation.

Let Claude Code plan with Opus and build with Sonnet

Type /model opusplan. In plan mode, Claude Code uses Opus for the thinking and design work; when it starts making changes, it switches to Sonnet automatically. Anthropic’s help pages call “plan with Opus and execute with Sonnet” a common pattern. (Claude Code docs, Claude Help Center) This is the easiest setup if you want Opus quality on the decisions without paying Opus prices on every edit.

Set the effort level

Type /effort to open a slider, or /effort low, /effort medium or /effort high to set it directly. You can also adjust it with the left and right arrow keys inside the /model picker. In Claude Code, medium is the default on all three 5.5 models. Anthropic suggests low for quick back-and-forth, high for bug fixes where edge cases matter, and warns that max “may show diminishing returns and is prone to overthinking.” /effort auto clears your saved choice. (Claude Code docs)

Update Claude Code if a model will not load

Each model needs a minimum version: Opus 5.5 needs v2.1.280 or later, Sonnet 5.5 needs v2.1.284 or later, and Anthropic says to use v2.1.293 or later with Haiku 5.5. If switching fails with a “does not support this model” message, run claude update in your terminal (or ask Claude Code to run that command for you), then restart it. (Claude Code docs)

Watch out for helper agents

When Claude Code hands work to a helper (a “subagent”) that does not have its own model set, the helper uses whatever model your session is on at that moment. So if you switch to Opus and then ask for a big research or test run, all of that helper work runs on Opus too. If you have custom helpers for routine jobs, ask Claude Code to “set the model for my subagent to haiku” so they stay cheap. (Claude Code docs)

Three setups that work

1. The budget setup (Pro plan, often hitting limits). Run /model sonnet once so it becomes your default. Use /model and press s to jump to Opus for one hard problem, then go back. Keep effort at the default. Clear between unrelated tasks with /clear. More usage-saving habits are in our guide to reducing Claude Code token overhead.

2. The quality-first setup (Max or Ultra). Run /model opusplan. You get Opus for planning and Sonnet for the hands-on edits. Raise effort to high only for bug hunts in an existing project.

3. The writing and office setup (Claude app). Leave Sonnet 5.5 selected for documents, slides and spreadsheets, which Anthropic names as Sonnet’s strength. Switch to Opus 5.5 for one long analysis or a decision memo, and to Haiku 5.5 when you just need a quick summary or rewrite.

Common mistakes

  • Never changing the Claude Code default. Claude Code starts on Opus 5.5 on most plans. If you only ever do small edits, you are paying the top rate for them.
  • Cranking effort to max “to be safe”. Higher effort makes Claude read more files, run more checks and produce more text before replying. That is useful for hard bugs and wasteful for a rename.
  • Using Haiku for very long sessions. Over 100,000 tokens per prompt, Haiku’s per-token price rises fivefold on pay-per-use accounts, and a long session gets there faster than you think.
  • Trusting the cost number when using a third-party gateway. OpenRouter found Claude Code’s own cost estimate was 30 to 60 times too high for Haiku 5.5 when routed through OpenRouter, because Claude Code did not recognise its model name. Check the gateway’s own billing page instead. (OpenRouter)
  • Switching models when the real problem is the prompt. If Claude is guessing, give it the file, the error message, or the goal first.

Troubleshooting

I do not see Opus in the Claude app. Opus is not on the Free plan. On Pro and above, click the model name and check More models. On Enterprise, your admin may have turned it off for your role.

Claude Code keeps going back to Opus. A project setting or an organisation default can override your personal choice on the next launch. Run /status to see what is active and ask your admin if a company default is set. If you pressed s in the picker, the switch was for that session only.

“/model haiku” works but the session feels short of memory. Anthropic says Haiku 5.5 runs with the full 1M-token window on its own service with no extra setting. If you run Claude Code through OpenRouter, OpenRouter reports that Claude Code 2.1.307 did not recognise its Haiku 5.5 model name and assumed only 200,000 tokens; its fix is to add [1m] to the end of the Haiku model name in your OpenRouter setup.

I ran out of usage anyway. See our step-by-step guide on what to do when Claude Code says your usage limit is reached.

FAQ

Is Sonnet 5.5 better than Opus 5.5? On some of Anthropic’s own tests, yes (Terminal-Bench 4.0: 70.6% vs 66.4%). On most others Opus is slightly ahead, and Anthropic says Opus is clearly stronger on complex, open-ended work. For everyday tasks the gap is small and Sonnet costs half as much per token.

Is Haiku 5.5 good enough for coding? For small, well-described changes, yes; it passed every short coding test in OpenRouter’s run. For hard, multi-step agent work it scores far lower than Sonnet on Anthropic’s Terminal-Bench 4.0, so keep it for the simple jobs.

Which model does Claude Code use if I never change anything? Opus 5.5 on Pro, Max, Ultra, Team, Enterprise and Anthropic API accounts, unless your organisation set a different default. (Claude Code docs)

Can I switch models mid-conversation without losing context? Yes, in both the Claude app and Claude Code. The new model picks up from the next reply.

Do I need to know the version numbers? No. In Claude Code, opus, sonnet and haiku always point to the latest version Anthropic recommends for your provider. Use a full name like claude-opus-5-5 only if you want to stay on one exact version.

For more on running Claude Code day to day, see our Claude Code daily-driver guide and how to set Claude Code permissions and auto mode. More tutorials are on the 7minai homepage.

Sources