Claude Opus 5.5 was released on September 22, 2026, and it is the model Anthropic suggests starting with for most work. It carries a 1M token context window, as every current Claude model does, and on a design job that single number is the whole story. It is a consistency story rather than a quality one.
Nothing about its taste is different. Point it at "a premium landing page" with no values attached and you get the same Inter, the same axis gray, the same translateY(20px) at 300ms that every flagship produces from an empty slot. Point it at a page of specification and a long build and it becomes the most useful model in the lineup.
The short call: reach for Opus 5.5 when the existing code and the full brief have to stay in the window together. For a single specified section, a cheaper tier returns the same page. The specification is what Thanor sells, and it is the part that moves the result.
What landed on September 22#
The release cut the API price below Opus 5's, to $4 in and $20 out per million tokens at the time of writing, made thinking always on and sharpened how the model reads charts and screenshots, according to Anthropic's notes on what changed. The 1M token window is not new to Opus, since Opus 5 had it in July, but for web work it is the number that changes the shape of a session rather than the shape of the CSS.
What fits in one conversation:
- A complete token file: surfaces, ramp, scale, radii, curves, six states per control.
- The real copy for eight sections, not a summary of it.
- The components already built, pasted as code rather than described.
- The refusal list, restated at the bottom where it stays in effect.
The practical result is that the footer still uses the 620ms reveal you named for the hero. On a 200K window, long builds drift: the model has stopped seeing the early decision and fills the slot again from the median, which is why a page built in six messages often reads as six pages. Long one-page sites are where that consistency shows, and each of these keeps one mechanism running from the first section to the last:
A cut-out bottle travels through the sections as you scroll, and with motion off the page still carries every word.
Open the prompt →A hero film of glossy nails on hot pink, then a one-page site whose hero, card row and studio column each pin as you scroll.
Open the prompt →The scroll is the round: every pinned section is a fight the visitor wins by scrolling, down to a scoreboard footer.
Open the prompt →On the public votes, Opus 5.5 is the highest-placed Claude model on Design Arena's Website board at the time of writing, seventh at 1324 and 43 points behind Meta's Muse Wonder, and first on its Overall Agentic board at 1459. Design Arena, explained covers what each board measures. Claude Opus 5 is real and superseded, so new work should not be built on it; check a model id against Anthropic's own list before a pipeline depends on it.
The window is also the reason to stop summarizing. Pasting the real component code costs a few thousand tokens and removes an entire class of near-miss, where the model builds something adjacent to the thing you described in prose and you spend twenty minutes reconciling the two by hand.
The specification Opus 5.5 executes cleanly#
Six slots, and this model will hold all six to the bottom of a long file.
| Slot | Value to state | Why this one |
|---|---|---|
| Ground | #0B0B0F page, #14141A raised | kills the gradient that fills unnamed depth |
| Neutral | #8A8577, biased toward amber | a gray on the axis is the clearest tell |
| Accent | #D8492B, one element per section | restraint is never a default |
| Scale | 17px base, 1.26 ratio, five steps | stops four near-identical sizes |
| Reveal | cubic-bezier(0.22, 1, 0.36, 1), 620ms | one curve, named, reused |
| Measure | 64ch, line height 1.62 | density is what reads as expensive |
Add a 90ms stagger capped at four items and the motion language is finished. Six rows plus one stagger rule is under eighty words, and it survives eight sections intact in a 1M window.
The row that earns its place most is the accent. Naming the single element per section allowed to carry #D8492B is the difference between a page with a focal point and a page painted in a brand color, and it works identically in every model in this cluster.
Restate the six rows at the bottom of a long prompt as well as the top. It costs sixty words, it reads like redundancy, and on an eight section build it is the difference between a footer that obeys the brief and a footer that reinvents it. These two endings were specified as carefully as any hero:
A closing section: a brass automaton among glowing glass reeds at night, with four lines of type set over the film.
Open the prompt →A studio footer where the cursor picks the film frame in which a knitted creature's pupils already point at you.
Open the prompt →A million tokens changes consistency, not taste#
Three things improve when the window is large enough to hold the whole job.
- Late sections obey early decisions. The price block still uses the ramp value from the hero, because the hero is still visible to the model.
- Existing code gets matched rather than replaced. Paste the component as code and the new section joins it. Describe it in prose and you get a near-miss that has to be reconciled by hand.
- The refusal list keeps working. "No gradient behind text" stated once at the top of a long session tends to lapse. Stated at the top and restated at the bottom of a
1Mcontext, it holds.
What does not improve is judgment about anything you left open. Color relationships, spacing rhythm, the feel of a duration: none of it is being generated, so none of it gets better with capacity. A larger window remembers your decisions longer; it does not make decisions on your behalf.
That distinction is the one the whole launch cycle blurs, and it is why a token file beats a tier upgrade on frontend work. The test takes four minutes: paste last month's brief into this month's model and check whether the values came back. When they do, the release changed your build speed and not your design.
What to keep out of the window#
A 1M context invites the wrong habit, which is pasting everything. Irrelevant context does not sit harmlessly; it competes with the constraints you care about.
Leave these out:
- Brand documents written in adjectives. A twelve-page voice guide saying "bold yet approachable" adds tokens and no decisions. Extract the three values it implies and paste those.
- Competitor screenshots described in prose. A description of somebody else's page pulls the output toward that page's median, not yours.
- Every file in the repository. Paste the four components the new section touches. The other ninety dilute the instruction, and in Claude Code the model can read the rest when it needs to.
- Previous rejected attempts. Leaving three discarded versions in the window keeps their decisions alive.
What to keep in, every time, is short: the token file, the refusal list, the section order, the copy for the section being built, and the code it has to sit next to. That is usually under four thousand tokens of the million available, and the ratio is the point.
A long window is a tool for holding the right context, not all context. Used that way, Claude Opus 5.5 is the model to build a whole site in, and the same discipline applies to interface work.
One habit makes the difference permanent: keep the brief in a file on disk rather than in the conversation. A chat history is not a document you can diff, hand to a colleague or paste into a different model next month, and every one of those three things is worth more than the convenience of scrolling up.
The verdict on Opus 5.5#
Use Claude Opus 5.5 for any build where consistency across a long file is the risk: a full marketing site, a design system rollout, a section that has to match forty existing components. The 1M window is what lets the whole job fit in one pass.
Do not expect it to change the look of anything you left unspecified:
- Unspecified brief,
Claude Opus 5.5: the category median, held consistently. - Specified brief,
Claude Haiku 5.5: a designed section, one at a time. - Specified brief,
Claude Opus 5.5: a designed site, eight sections that agree.
Only the third of those is worth building a business on, and two thirds of it is the document you paste in. Claude Fable 5.2, Anthropic's top model since November 9, is the sibling to reach for when a job needs more reasoning than Opus gives it, and it takes the same brief unchanged. That interchangeability is the point: a brief that works in both is an asset you keep through the next three releases, while the model id on its first line is a detail you edit in a second.
Thanor writes that document: exact values per design, from the palette and the easing curves to the section order and the media the page uses, for personal and client work. Free accounts open the designs marked free, and the plans open the rest of the library. In Claude, the Thanor MCP pulls a design's full art direction into the session without the clipboard.
Open Vesperion and copy its prompt into Opus 5.5. It is free with an account, and it fills the window with one brief instead of with everything you own.
Questions this raises
What is new in Claude Opus 5.5 for design work?
Released September 22, 2026, it costs less than Claude Opus 5, keeps thinking always on and reads charts and screenshots more precisely, according to Anthropic. Its 1M token context is not new, since Opus 5 had one too, but it is what matters on a design job: the token file, the copy and the existing components all stay in view, so the last section still obeys the curve named in the first.
Is Claude Opus 5.5 better than Claude Fable for websites?
They serve different jobs. Anthropic suggests starting with Opus 5.5 for most work and moving to Claude Fable 5.2, its top model since November 9, for demanding reasoning and long agentic runs. Both have a 1M token context, and both take the same design brief unchanged.
Does a bigger context window make a better looking page?
It makes a more consistent one. Nothing about the color sense or the type judgment changes, because those are copied from the brief. What changes is that section eight stops contradicting section one.
How long should a brief for Opus 5.5 be?
One page of values, not fifty pages of context. A `1M` window is permission to include the real component code and the real copy, not an invitation to paste every document you own. Long irrelevant context dilutes the constraints that matter.
Which Opus release should I actually use?
Claude Opus 5.5, released September 22, 2026, and the model Anthropic's own documentation suggests starting with. Claude Opus 5 is real but now a legacy model, so new work should not be built on it. Check any model id against the provider's own list before you pin a pipeline to it.
Part of
Every model, on design
Each new model put on the same job, designing a website, and judged on what it actually produces. Which one to reach for, and the part of the result that does not depend on the model at all.