Claude Opus 5.5 for Ad Creative: What It Costs, What It Changes, and What It Still Cannot Make

TL;DR

Anthropic released Claude Opus 5.5 on September 22, 2026. The headline for advertisers is the bill: 4 dollars per million input tokens and 20 dollars per million output, a 20% cut against Opus 5, and Anthropic puts the all-in saving at about 40% on typical workloads once cheaper cache reads are counted. It is also more than 30% faster at generating output. It reads images and writes text, and it does not output images or video. So it runs your angle development and your performance reads for meaningfully less than the last model did, and it still cannot produce the 40 on-brand, correctly sized variants your Advantage+ campaign needs on Monday. Pair it with a production tool and the loop closes.

Claude Opus 5.5 reached the Claude Platform, Claude Code, and the Pro, Max and Team plans on release day, alongside Amazon Web Services, Google Cloud and Microsoft Azure. Opus 5 stays available for teams mid-migration. It arrives two months after Opus 5, which is a fast cadence even by this year's standards.

For a media buyer, two lines from the release do the work: the price went down and the speed went up. Everything else is a better version of what Opus 5 already did for you.

The specs that affect your workflow

The context window matters for one specific job: dropping a full ad-account export into a single prompt and asking what moved. A 1 million token window holds months of daily creative-level data without you pre-summarizing it first.

What a run costs

Opus 5.5 bills at 4 dollars per million input tokens and 20 dollars per million output tokens. Two details change the bill more than the headline rate does:

Run the numbers on a real task. Hand Opus 5.5 a 200,000 token export of creative-level performance and ask for a written read plus 30 new angles, and you use about 30,000 output tokens. That is 0.80 dollars of input and 0.60 dollars of output, so 1.40 dollars for the round. The same request on Opus 5 ran 1.75 dollars. Now re-run it weekly against the same cached brief and the input side drops to pennies, because cache reads land at 0.20 dollars per million rather than 4.

That last detail is the one worth building a habit around. The weekly performance read is the task teams skip because it feels like it costs something. At cache-read rates it does not.

Fast mode runs the same model at up to 2.5x the speed for double the price, in Claude Code and on the Claude Platform. For creative work it is rarely worth it: nobody is waiting on an angle list in real time. Save it for interactive sessions where latency is actually the problem.

Analysis is cheap at these rates. Creative volume is where the money goes, and no token price changes that.

What actually improved

Anthropic's published numbers put the gains in long, multi-step work rather than single-prompt tasks, which is the shape of creative analysis:

One number is worth pausing on. Opus 5.5 scores 66.4% on Terminal-Bench 4.0 against 55.8% for Claude Fable 5.1, Anthropic's flagship, on the same benchmark. Fable 5.1 costs 10 and 50 dollars per million tokens. Opus 5.5 costs 4 and 20. For the long multi-step work that creative analysis actually is, the cheaper model now leads, which reverses the advice that held when Opus 5 shipped.

The price cut is the part that is new

Opus 5 was already good enough at the thinking half of ad creative. What a 20% per-token cut, a roughly 40% lower all-in bill, and faster output change is how freely you use it. Asking for 40 angles instead of 20 stops being a decision. Re-running last quarter's export because you had a new hypothesis stops being a decision. The model got better; the hesitation went away.

The jobs worth pointing it at first:

• Angle development. Give it your product, audience, and objections and ask for 30 to 40 distinct messaging angles, each with a hook and a one-line rationale. Not ten rewordings of one.
• Performance reads. Drop in the creative-level breakdown for the last 14 days and ask which hooks held hook rate and which decayed.
• Research synthesis. Feed it reviews, support tickets, and competitor copy, and have it cluster the objections worth answering in creative.
• Scripts and hooks. First-three-second hooks, UGC-style scripts, and headline sets you can hand straight to production.

Each of those is a task you currently pay a human to do on a Tuesday morning. None of them puts a new creative in the auction.

Where Opus 5.5 stops

Opus 5.5 is a text-output reasoning model with image understanding. It does not output images, video, or audio. That limit has held across every Opus release and it did not move with this one:

• A description is not a variant set. Advantage+ and Performance Max want many on-brand assets sized for every placement, refreshed on a cadence. Copy decks do not add up to that.
• Brand consistency is a production problem. Palette, logo placement, type, and safe areas have to hold across every asset. A model that writes cannot enforce them across 200 renders.
• The volume math is unchanged. Starving the platform of fresh assets is still the fastest route to creative fatigue, and a cheaper, faster reasoning model does not render one frame of the fix.
• It does not size for placements. Feed, Stories, Reels, and Shorts each want different crops and safe zones.

The workflow that uses both

1. Brief once. Product, audience, offer, brand rules, and the objection you want to overcome. Keep it in the cache and reuse it.
2. Ask for angles, not copy. Expand that brief into 30 to 40 distinct angles with a hook and a one-line rationale each, then rank them by fit to your audience.
3. Produce the variants in a creative tool. Feed the winning angles into Tadka and get a grid of audience-tuned image and video variants that keep your palette, fonts, and logo in place.
4. Ship, read, repeat. Run the results back through Opus 5.5 weekly against the cached brief, and let the read set the next round of angles.

The loop keeps a reasoning model on the parts that are reading and deciding, and keeps a production tool on the part that is making 200 assets that look like your brand.

Which tool for which job

• Use Claude Opus 5.5 when the task is thinking: performance analysis, research, angle generation, scripts, and copy.
• Use a creative-production tool when the task is output: many on-brand variants, correct sizes for every placement, and refreshes that beat fatigue.
• Use both, in that order, on every campaign where you need strategy and volume, which covers most performance work.

What to do this week

• Move any Opus 5 workload straight over. Same workflow, same prompts, 20% less per token and about 40% less all-in.
• If you were paying Fable 5.1 rates for creative analysis, re-test on Opus 5.5 first. It now scores higher on the multi-step benchmark at well under half the price.
• Put your standing brief in the prompt cache before you run anything repeatedly, so the weekly read bills at 0.20 dollars per million rather than 4.
• Ask for 30 angles from your best-performing month, not 3, then cut the list yourself.
• Route production to a tool built for creative volume rather than writing copy decks and hoping. Our roundup of the best AI ad creative tools covers the options.

Sources: Anthropic: Introducing Claude Opus 5.5, Anthropic: Introducing Claude Opus 5, Anthropic: Models overview, Anthropic: Pricing

Opus 5.5 gives you the read and the angles. Tadka turns those angles into a grid of on-brand, audience-tuned ad creatives across every size your campaigns run, so the platform never runs short on variety. Try it in the studio, or set the loop up with the volume playbook.