The price is the story
Claude Opus 5.5 is out, and the most interesting number on the announcement page is not a benchmark score. Input tokens now cost $4 per million, down from $5 on Opus 5. Output fell from $25 to $20 per million. Cache reads dropped hardest, from $0.50 to $0.20 per million. Anthropic says those rate cuts plus a model that burns fewer tokens per task net out to roughly a 40 percent drop in cost per task against Opus 5.
The rate cut alone is maybe 20 percent depending on your mix. The other 20 has to come from token efficiency, which is the part worth watching. If a model hits the same answer in fewer tokens, your bill drops without anyone touching a price list. That is a pricing lever providers can pull quietly, and it only shows up in invoices, not in press releases. Watch your own spend after a model swap, not the sticker rates.
A rewrite of how the model writes
Anthropic spent much of the announcement on prose quality, which tells you something about what users complained about. The page says Opus 5.5 puts the most important information up front, leans less on jargon and idiosyncratic phrases, and follows writing rules you give it. There is a side-by-side code comparison on the page too, showing the new model producing a tighter interval check than its predecessor on the same task.
I find this a slightly funny turn of events. Model vendors spent two years teaching assistants to sound a certain way, users spent two years complaining that everything reads like a press release, and now writing style is a headline feature of a flagship model. Whatever you think of the framing, it acknowledges the real complaint: output length and output tone are usability problems, not cosmetic ones. Less filler means fewer tokens billed and less time skimming.
The code review claim, with an asterisk
The strongest customer quote on the page concerns code review. A customer reports that Opus 5.5 caught 72 percent of known bugs in their code reviews even at its lowest effort setting, against 56 percent for Opus 5 running at high effort, with fewer false alarms and much shorter output. Another quote describes financial workflow coverage of 86.6 percent against expert rubrics, up from 60.3 percent for Opus 5. One quant team even caught the model noticing an off-by-one error in its own evaluation instructions and correcting for it.
All of this is vendor-reported. The customers are quoted, not named in a way you can go verify, and the measurements ran on Anthropic’s harness. That does not make the claims wrong, it makes them unverified. The pattern to watch is the effort setting doing the heavy lifting: a cheaper configuration beating the previous model’s expensive one suggests the effort controls now carry real weight, which is worth testing on your own workloads before believing or dismissing it.
Benchmarks and the fine print
The only public benchmark on the page is Terminal-Bench-Science 0.1, and Anthropic includes the caveats voluntarily. Their own reproduction puts Opus 5 at 29.0 percent against a public leaderboard figure of 30.0 percent, within stated error bars of roughly 3.5 to 5 points per model. The page also says outright that its scores are not directly comparable to other vendors’ published setups, because harness conditions differ.
Credit where due: most launch posts bury that sentence, if they include it at all. The honest read is that the page offers almost no cross-vendor comparison, and the numbers it does show are single-party measurements. If you are deciding between frontier models this week, the announcement gives you pricing and direction, not a ranking. Run your own eval set.
What to actually do
If you run cache-heavy workloads, the cache read price cut is the biggest single win here, a 60 percent reduction per token. Point your longest-lived contexts at the new model first. Second, treat effort settings as a cost knob: pin them low by default and raise them only where output quality measurably improves. Third, if you prompt for output style, re-read your style rules against the new model, since it claims to follow them more literally now, and a stale rule list will shape output more than it did before.
There is also a competitive read here. Every frontier lab now sells a top tier and a cheap efficient tier, and the gap between them keeps shrinking. When a flagship model gets 20 percent cheaper while claiming equal or better output, the cheap tier has to move too, and the models that do not follow will look overpriced within a quarter. If you are on a second tier model purely for budget reasons, it is worth re-running that comparison now rather than at your next contract renewal.
My take: the price cut is the safe claim, the efficiency claim is plausible but bill-dependent, and the quality claims need independent numbers. The pattern across the last few releases is that flagship tiers keep getting cheaper at the same quality level, which is good news for anyone running long agentic workflows. The cost per task matters more than the cost per token, and this release is one of the more explicit acknowledgments of that.