What happened
Anthropic has launched Claude Opus 5.5, the first model in a new family, and the pitch is unusually blunt: near-flagship results for noticeably less money. The company says its new AI text generator matches Claude Fable 5.1 "on most tasks" while costing about 40 percent less to run than Claude Opus 5 — and it produces output more than 30 percent faster. The launch landed on September 22, 2026, with Claude Sonnet 5.5 and Claude Haiku 5.5 promised "in the coming weeks" with similar gains in performance, efficiency, and safety.
Anthropic frames the release as an answer to three customer complaints: cost, efficiency, and communication quality. Financial services, law, and software development teams are named as the loudest voices.
The scores Anthropic is pointing at
"Agentic coding" sounds intimidating, but it just means the model doing multi-step work by itself — opening a terminal, editing files, checking whether the code still runs. On Terminal-Bench 4.0, Opus 5.5 scores 66.4 percent. For comparison: Fable 5.1 gets 55.8 percent, Opus 5 gets 52.3, OpenAI's GPT-6 Astra gets 57.9, and GPT-5.6 Sol gets 37.3. On FrontierCode v1.1, Opus 5.5 hits 54.4 percent (Fable 5.1: 50.3; Astra: 53.3). On CursorBench 4.0 it finishes ahead of both Fable 5.1 (51.8 percent) and GPT-5.6 Sol (41.7 percent).
Knowledge-work and reasoning scores follow the same shape: Opus 5.5 tops the GDPval-AA v2.1 knowledge-work index, ahead of both Fable 5.1 and GPT-6 Astra, and it scores 67.7 percent on Humanity's Last Exam with tools enabled. There's also a best-in-class result on OSWorld 2.0, a "computer use" test — though Anthropic labels that figure partial.
The price cut, in plain terms
A token is a small chunk of text, roughly three-quarters of a word. You pay for the tokens you send in (your prompt, your documents) and the tokens the model sends back (the answer). Cache reads are the cheap part: if you keep asking about the same long document, the model can reuse what it already processed instead of re-reading the whole thing.
Anthropic's new prices per million tokens:
• Input: $4, down from $5.
• Output: $20, down from $25.
• Cache reads: $0.20, down from $0.50 — a 60 percent cut.
• Cache writes: $5, down from $6.25.
That's a 20 percent cut on token prices. But the bigger number is total operating cost: Anthropic says it drops by about 40 percent versus Opus 5, because the model also burns fewer tokens per answer. Subscribers get more room too — five-hour usage limits rise 20 percent, and with the lower token use Anthropic says those limits stretch about 25 percent further overall. You can also bank a limit reset for when you really need it.
The company makes a cost-per-task case as well. On FrontierCode, it says Opus 5.5 beats GPT-6 Astra at roughly 20 percent of the cost per task; on Terminal-Bench 4.0, it claims parity with Astra at 40 percent of the cost; on CursorBench, it beats GPT-5.6 Sol by 11 points at one-third of the cost. The backdrop is obvious: pressure from OpenAI and, above all, from cheap Chinese models that cost a fraction as much.
Less "Claudish," more like a person
If you've used Claude, you know the tell: three bullets of throat-clearing before the actual answer, and phrasing that sounds like a consulting deck. People call it "Claudish," and it's been a fair criticism. Anthropic says Opus 5.5 leads with the most important information, uses less jargon, and follows writing instructions more closely. Early testers described it as clearer and easier to understand — which matters a lot when you hand a model a long task and check back later.
It's also the first Opus model to ship with safeguards for cyber-related requests.
What it means for you
Leaderboard numbers don't pay your bills. Cheaper tokens and fewer wasted sentences do. Here's where this actually shows up.
At home: chores stop turning into projects
Long, tedious asks are exactly where cheap tokens matter. "Read these twelve school emails and put the dates in one list" or "build me a seven-day meal plan around what's already in my fridge" used to feel wasteful to run again and again. At $4 input and $20 output per million tokens, you can afford the back-and-forth — and the 30 percent faster output means less staring at a spinner.
At work: the first draft stops sounding like a press release
The "Claudish" problem was never that the model got facts wrong. It's that you had to rewrite everything to sound human. If Opus 5.5 follows style instructions more closely and puts the point first, you delete fewer paragraphs. Multiply that by three reports a week and it's a real chunk of your afternoon back.
For small business: agency-level copy on a coffee budget
Product descriptions, support replies, proposal drafts, follow-up emails — this is the unglamorous work that eats a founder's day. Opus 5.5 is priced at a fifth of what you'd spend per task with the more expensive model Anthropic is comparing against, and cache reads at $0.20 per million tokens mean reusing your brand guide or your FAQ over and over is nearly free.
For studying: a tutor that actually explains things
The Humanity's Last Exam score of 67.7 percent with tools is a proxy for something practical: reasoning across subjects. Ask it to explain a concept three ways, quiz you on it, then grade your answer. Because the model leads with the key point, you spend your time learning instead of untangling the explanation. If you'd rather experiment with free tools before paying for anything, MyKreaTool is a good place to see what free AI tools can already handle.
For creative work: charts, captions, voice
Visual chart recognition now leads on Chartography with tools, ahead of both Fable 5.1 and Opus 5. That's genuinely useful if you work with dashboards or exported data and want plain-English summaries. And on the writing side, "follows writing instructions more closely" is the line that matters if you've ever spent twenty minutes telling a model to stop starting sentences with "Moreover."
For a side hustle: more output, thinner bill
Freelance writers, transcribers, and two-person agencies live and die on margin. Faster output at a lower per-token price means you can take on the project you'd normally skip. The cost-per-task claims matter here far more than the leaderboard — a task that costs a fifth as much is a task you can actually quote on.
How to try it right now
1. Start free. Open Claude in a browser on Anthropic's consumer app and check the model picker. If Opus 5.5 is listed, you're set; if not, you're on the default model for your plan, and you can still test prompts to see how the writing feels.
2. Give it a style sample. Paste a couple of paragraphs of your own writing and say "match this tone, no bullet points, lead with the answer." It's the fastest way to judge whether the "Claudish" fix is real for your work.
3. Ask for the answer first. Add "put the conclusion in the first sentence, then the reasoning" to your standing instructions. Old habits are hard to break — even for a new model.
4. Upgrade when limits bite. Paid plans get five-hour limits that are 20 percent higher, and the cheaper token burn stretches them about 25 percent further. You can also save a limit reset for a deadline day.
5. For developers, go to the API. You'll pay $4 per million input tokens and $20 per million output tokens. Turn on prompt caching early: cache reads at $0.20 per million versus $0.50 add up fast if you keep feeding in the same documentation or codebase.
6. Run your own bake-off. Take one real task, run it through Opus 5.5 and whatever you use now, and count how many edits each output needs. That number beats any benchmark table.
Upsides and what changes
The good stuff is easy to list. Token prices are down 20 percent, cache reads are down 60 percent, total running cost is roughly 40 percent lower than Opus 5's, output is over 30 percent faster, and subscriber limits go 20 percent higher (stretching about 25 percent further in practice). Quality is flat-to-better across almost every row in Anthropic's table, and the writing is supposed to need less cleanup.
The strategic shift matters more than any single number. For a couple of years the story was "pay more, get more." Opus 5.5 is Anthropic admitting that most buyers just want good enough at a price that doesn't hurt — a direct response to OpenAI and to low-cost Chinese models. Sonnet 5.5 and Haiku 5.5 arriving in the coming weeks with the same treatment suggests this becomes the new baseline across the whole lineup.
Limitations
Take the benchmark table for what it is: vendor-selected. GPT-6 Astra still beats Opus 5.5 on AutomationBench (41.4 percent to 40.0) and on Terminal-Bench-Science 0.1 (64.6 to 58.7), several cells read N/A, and the OSWorld result is explicitly partial — so "wins on most tasks" is Anthropic's framing, not a clean sweep. The "less Claudish" promise rests on early testers, who are rarely a random sample, and writing taste is subjective: you may still spend your evening deleting "It's worth noting that." Cheap is also not the same as free — $4 and $20 per million tokens add up on heavy use, and light Chinese models cost a fraction of that. Faster and cheaper makes it tempting to stop checking facts. Don't.
Conclusion
Claude Opus 5.5 is a pricing story dressed up as a model launch, and that's exactly why it's worth your attention: it's the moment "good enough, much cheaper" became the default sales pitch in frontier AI — plus a genuine attempt to kill the corporate robot voice.
Here's your one action for today: open Claude, paste a short sample of your own writing, and ask it to rewrite one paragraph in your voice with the main point first. Two minutes tells you whether the "Claudish" fix is real — and whether that next 40 percent in savings lands in your budget or someone else's.
Comments 0