On September 22, Anthropic released Claude Opus 5.5, which it calls "the first model in our new Claude 5.5 family" (https://www.anthropic.com/claude-opus-5-5). The launch post reached 1793 points on Hacker News, and for the rest of the week the site kept returning to Claude: an independent analysis of the model, a Claude Code change about AGENTS.md, and a post from Anthropic's own engineers on how they made claude.ai faster.
If you build web apps and use an AI model while you do it, you are probably already on this one without choosing it. The same day, Claude Code 2.1.280 made Opus 5.5 the default Opus model and moved Pro and Team Standard plans from Sonnet to Opus by default (https://code.claude.com/docs/en/changelog). So it is worth knowing what actually changed.
What changed
Anthropic's headline is cost and speed. It says Opus 5.5 costs 40% less to run than Opus 5 and generates output more than 30% faster (https://www.anthropic.com/claude-opus-5-5).
The price list backs up part of that directly:
- Input: $4 per million tokens, down from $5 on Opus 5.
- Output: $20 per million tokens, down from $25.
- Cache reads: $0.20 per million tokens, down from $0.50.
- Cache writes: $5 per million tokens, down from $6.25.
There is also a fast mode at $8 per million input tokens and $40 per million output. The context window is 1M tokens (https://code.claude.com/docs/en/changelog). It is on the Claude Platform as claude-opus-5-5, and on Amazon Web Services, Google Cloud and Microsoft Azure.
Anthropic also says Sonnet 5.5 and Haiku 5.5 "will follow in the coming weeks". No date beyond that.
How it scores on the work we care about
Anthropic published a benchmark table against Opus 5 and OpenAI's GPT-6 Astra. The rows closest to day-to-day coding:
- Terminal-Bench 4.0: 66.4%, against 52.3% for Opus 5 and 57.9% for GPT-6 Astra.
- CursorBench 4.0: 57.8%, against 46.6% for Opus 5.
- FrontierCode v1.1: 54.4%, against 48.0% for Opus 5 and 53.3% for GPT-6 Astra.
- OSWorld 2.0, which measures using a computer's interface: 81.8%, against 74.0% for Opus 5.
These are the vendor's own numbers. The independent one comes from Artificial Analysis, which gives Opus 5.5 a score of 58 on its Intelligence Index, ranked 1 of 211 models (https://artificialanalysis.ai/models/claude-opus-5-5). The same page ranks it 94 of 211 on cost efficiency, and that gap is the most useful thing on it.
The catch: price times verbosity
A lower price per token only lowers your bill if the model does not use more tokens. Artificial Analysis found that running its index, Opus 5.5 generated 260M tokens, against a median of 88M. It calls that "very verbose".
So two things are true at once. Each token is cheaper than it was on Opus 5. And this model spends a lot of tokens. What you pay on a real task is the product of the two, and nobody has published that number for frontend work yet. Anthropic's "40% less to run" is its own claim, measured on its own workloads.
For a web developer, where this bites is agent loops. An agent that re-reads a large Next.js codebase on every step is mostly paying for input, and input and cache reads are where the price fell furthest. An agent that writes long explanations of what it did is paying for output, which is where verbosity costs.
What the safety notes mean in practice
Two lines in the announcement matter more to someone letting a model touch their repo than the benchmarks do. Anthropic says Opus 5.5 is "much less likely than recent models to take hard-to-reverse actions", and "more resistant than Opus 5 to prompt injection". It also reports that the model attempted to get around its boundaries 85% less often than Opus 5 in its automated audit.
If you run an agent that reads web pages, issues or pull request comments, prompt injection is the attack you are actually exposed to. Those are claims, not guarantees; your guards still have to assume the model can be talked into something.
The frontend story hiding in the same week
The post I would send any React developer this week is not the launch. It is "Once Claude can measure something, it can make it faster", where three Anthropic engineers describe making claude.ai roughly 3x faster in a two-week sprint in August (https://claude.dev/blog/how-we-made-claude-ai-faster/).
One detail first, because it is easy to misread: the model doing the work there was "Claude Tag", which the authors describe as an internal research model comparable to Opus 5.5. It was not Opus 5.5 itself.
The numbers are p75 timings, August 13 against August 27:
- A fresh load of claude.ai on the web went from 3,085 ms to 550 ms.
- Loading a conversation on the web went from 1,557 ms to 646 ms.
- Sending a message on the web went from 180 ms to 59 ms.
Over 3,000 changes were merged, nearly 200 feature flags were added, and sidebar re-renders were cut by 90%.
The method is the lesson, and none of it needs a frontier model. They measured before touching anything: instruction counts, React commits, style recalculations, DOM mutations, wall-clock time. They checked that the lab number moved with the latency users actually felt. Then they shipped each change behind a flag and read field data before the next one. One fix I liked: they baked a static version of the message composer into the HTML, so people can start typing while React is still initialising.
That is ordinary frontend performance work, done at a pace a person could not sustain. The model made the loop faster. The measurement made the loop possible.
What is not known yet
- Whether the coding benchmark gains show up on everyday frontend tasks, like form state, CSS bugs or a slow list, is not measured anywhere public.
- What a real task costs, once verbosity is counted, is still mostly Anthropic's word.
- When Sonnet 5.5 and Haiku 5.5 arrive, and whether they are the better default for cheap, frequent calls.
My take
I run my own content and outreach worker on Claude Code, scheduled through the day, and the thing I learned there is that tokens per run matter more than price per token. My notes record 93–139K tokens per call when each phase starts in a fresh session, against 965K for one session left running all day. No price cut would have saved as much as that one change did.
So my read on Opus 5.5 is that the cheaper cache reads are the real news for anyone running agents, and the verbosity finding is the real warning. The part of this week I will actually use is the claude.ai post: before I let any model optimise a React page for me, I want the benchmark it is optimising against, and proof that the benchmark tracks what a user feels.
