Claude has two new models. Opus 5.5 is the one for work that needs careful thought, and it is already on for paid plans. Sonnet 5.5 is the quicker one for everyday tasks. Both sit in the model menu next to the send button, and you can swap at any point in a chat.
What changed
- Opus 5.5, 22 September. Anthropic's announcement says it "communicates more naturally than prior models", puts the most important information up front, and writes more than 30% faster than Opus 5. Its developer account says it is now the default on paid plans.
- Sonnet 5.5, 28 September. Anthropic's launch page calls it a faster, lower-cost complement to Opus 5.5, strongest at "well-scoped everyday tasks" and at making documents, slides and spreadsheets. It runs 30% faster than Sonnet 5 and costs up to 30% less per task.
- Haiku 5.5 is next, due in the coming weeks according to the same page.
- More usage on paid plans. The Opus announcement raises the five-hour usage limit on Pro, Max, Team and seat-based Enterprise. It names no change to plan prices.
Which one to pick
The model menu sits next to the send button, and you can change it at any point in a chat. Here is how Anthropic itself divides the work, as of 30 September 2026.
Opus 5.5
- Built for: work that needs judgment across many steps. Anthropic says it "remains clearly stronger at complex, open-ended work requiring sustained judgment".
- Everyday examples: a plan you will revise several times, a difficult reply, a document where the order of ideas matters.
- Default on: paid plans, according to Anthropic's developer account.
- Verdict: keep it if you are on a paid plan and most of your chats are the thinking kind.
Sonnet 5.5
- Built for: well-scoped everyday tasks, documents, slides and spreadsheets.
- Everyday examples: a quick rewrite, a summary, a table, a first draft you will finish yourself.
- Cost: at Medium effort, the default in the Claude apps, Anthropic says it beats Sonnet 5's best score for less than a tenth of the cost per task. That is Anthropic's own per-task measure, and your plan's bill is unchanged.
- Verdict: pick it for quick jobs, and whenever you are watching your usage limit.
One setting matters as much as the model. Anthropic's help page (read 30 September 2026) says Low and Medium effort "work well for routine tasks and stretch your usage further", and that on Opus 5.5 thinking is always on, whatever the effort.
What testers found
- Plainer writing, measured. Every ran its test prompts on 22 September and found Opus 5.5's prose averaged a reading grade of 6.95, a full grade below Opus 5. It also revised a proposal through 12 rounds of pushback.
- It sticks to your rules. Alex McFarland, a writer, wrote on 24 September that it "follows skills more precisely than any model I've tested. It doesn't drift", and added that a lot of his verdict "comes down to vibes".
- It still buries the point. Every's headline weakness. Given ten minutes to write a run of show, the minute-by-minute schedule for an event, it built handouts and training material instead and ran out of time with no schedule.
- It can eat your allowance. Two of Every's testers hit their weekly limits during the test, and one run used up its whole allowance.
The prompt: run my task and put the point first
Replace each bracketed line under THE TASK, paste your material, and send it to whichever model is on. To compare the two models, run it once on each and paste the first answer into the last line the second time.
You are doing one real piece of my work, and I am going to judge you on it. I am testing whether a new model is worth leaving on, so I want the result I asked for, at the length I asked for, with the point in the first two lines. Extras I did not ask for count against you, and so does anything invented. THE TASK (I fill this in once) - What I want done: [describe it plainly, e.g. "reply to this email from a supplier who missed a deadline", "turn these notes into a one-page plan for the team", "rewrite this paragraph so my mum could follow it"] - Who it is for: [e.g. "my manager", "a customer who is already annoyed", "me, as notes"] - What a good result looks like: [e.g. "they know what I decided and what I need from them by Friday", "one page, no jargon"] - How long it can be: [e.g. "under 150 words", "one page", "five bullet points"] - What it must not do: [e.g. "no apology at the start", "no new dates I did not give you", "do not mention the price"] - What matters most if you run short: [e.g. "the reply itself, more than any extras", or "nothing, take your time"] - The material: [paste the email, notes, draft or facts here. If a fact is not here, it does not exist] - Another model's answer to the same task (optional): [paste it here if you want the two compared] BEFORE YOU START 1. Say in one line what you understand the task to be and who it is for. If a line above is blank or still in brackets and you need it, ask me once and wait. If you can do without it, say what you assumed. 2. The deliverable comes first. Do not prepare supporting material, examples or a better version of my brief before the thing I asked for exists. 3. If the task needs a fact, a name, a date or a number I did not give you, write [NOT STATED] in its place. Never fill a gap from general knowledge. STEP 1: PUT THE POINT FIRST Your first two lines carry the answer: the decision, the ask, or the one thing the reader must know. Everything else comes after it. If the task is a piece of writing, the first sentence of that writing does the same job. STEP 2: DO THE TASK Produce exactly what I asked for, at the length I set, for the reader I named. One deliverable. If you think an extra would help, do not make it; describe it in one line at the end and I will ask if I want it. STEP 3: CUT IT Read what you wrote. Remove every sentence that repeats an earlier one, explains a sentence that was already clear, or warms up before saying something. If you are over the length, cut until you are under it, starting with the least useful sentence. STEP 4: CHECK IT AGAINST MY BRIEF Go down "what a good result looks like" and "what it must not do" line by line. Mark each one MET or NOT MET. If anything is NOT MET, fix it before you show me. STEP 5: IF I PASTED ANOTHER MODEL'S ANSWER, COMPARE THEM Judge both answers, yours included, on four things: is the point in the first two lines, is it within the length, did it invent anything, did it follow "must not do". Give a one-line verdict per answer, then tell me which to keep and why in no more than three lines. Do not favour your own answer, and do not praise either. RULES - Never invent a fact, a figure, a name or a quotation. [NOT STATED] is always the right move. - Keep your own messages short. The deliverable is what I came for. - If you are not the model I picked for this chat, say so at the top of your reply. - Do not tell me the task was interesting, and do not summarise what you did at the end. CHECK YOURSELF BEFORE YOU ANSWER 1. Is the point in the first two lines? 2. Is it within the length I set? 3. Did anything come from outside the material I gave you? 4. Did you make anything I did not ask for? 5. Name the one thing in this result I should check with my own eyes.
The honest bit
- The reviews are early. Both were written within two days of the launch, and one reviewer says his verdict comes down to feel. Give it a month of real use before you treat any of it as settled.
- Menus move. Two models landed in the menu in one week. If yours does not match this page, the help page linked above is the current word.
- The prompt fixes the habit, once. A model that buries the point will do it again in the next chat. Keep the prompt in a Project or a rules file if it helps.
Try it on one task today
Take the last thing you asked Claude to write, run the prompt on Opus 5.5 and again on Sonnet 5.5 with the first answer pasted in, and read the verdict. That tells you more about which to pick than any review, this one included.
A few quick questions
Do I have to switch anything?
No. Anthropic's developer account says Opus 5.5 is the default on paid plans, and the model menu next to the send button still lists the older models if you want one back. Open the menu once to see what yours is set to.
Why did Claude say it switched models?
Anthropic's help page says Opus 5.5 runs safety checks on every request, including files and search results it reads. Some cybersecurity requests fall back to Opus 4.8, and some biology or AI-development requests to Opus 5. You see a notice, and you can switch back from the menu.