Claude Opus 5.5 does Fable-level work for 40% less than Opus 5

Anthropic released a new model today, and the announcement led with price, not power. The @claudeai account introduced Claude Opus 5.5 as the first model in a new 5.5 family, saying it "performs at the level of Claude Fable 5.1 for most tasks" and costs 40% less to run than Opus 5.

Soon after, @lydiahallie put it more casually: it "feels like Fable, but ~30% faster and ~40% cheaper per task." Launch-day posts are a pitch, and I'd read both that way. Still, the numbers Anthropic published alongside them are specific enough to check.

What changed

The launch page cuts list prices by 20%: $4 per million input tokens, down from $5, and $20 per million output tokens, down from $25. Cached input drops from $0.50 to $0.20 per million, a 60% cut. Anthropic's 40% figure is the combined effect on what it calls typical workloads, where a lot of the input is reused context read from cache.

Anthropic also says Opus 5.5 generates output "more than 30% faster" than Opus 5. That's easy to skim past, but a faster model keeps an agent loop moving, and anyone sitting there waiting on a draft will notice.

The benchmark gains are largest on long agentic work. Anthropic reports Terminal-Bench 4.0 at 66.4%, up from 52.3% for Opus 5, CursorBench 4.0 at 57.8% versus 46.6%, and computer use on OSWorld 2.0 at 81.8% versus 74.0%. On GDPval-AA, a test built from office knowledge work, Opus 5.5 posts a higher Elo than Fable 5.1 itself. That's the basis for the "Fable-level" line.

The quieter change: shorter answers

The change I care most about isn't on a benchmark chart. Anthropic says Opus 5.5 puts the most important information first and writes shorter, better-structured replies with less jargon. One early customer quoted on the launch page called it 40% less verbose without losing accuracy.

For a business, this cuts the bill twice. You pay for every output token, so a model that says the same thing in fewer words costs less per task, on top of the lower rate. And a summary that leads with the answer is one your staff will actually read.

It's also the opposite of the trend we wrote about yesterday with Grok 4.7, where the gains came from a model that thinks longer and spends more than twice the tokens to get there. Same week, two labs, opposite bets on how much a model should say.

Where to stay skeptical

Anthropic picked these benchmarks and ran most of them. "Most tasks" also leaves room: Fable 5.1 is still the bigger model, and there will be work where the gap shows. TechCrunch notes Opus 5.5 ships with the same cyber and biology safeguards as Fable, so a few specialized uses stay restricted.

The 40% figure also depends on your workload looking like Anthropic's. If your prompts are short and you rarely reuse context, the cache discount does little for you, and your saving is closer to the 20% list price cut plus whatever the shorter answers save.

Anthropic also framed the release around safety. It calls Opus 5.5 its first model since the company called for pacing the frontier, and reports better scores than any recent Claude model on its automated behavioral audit. Sonnet 5.5 and Haiku 5.5 are due in the coming weeks.

What it means for a business using AI

Price cuts at the top end change which jobs are worth handing to the best model. A month ago, running an Opus-class model on every inbound email or every quote draft was hard to justify. At 40% less, with shorter output, some of those workflows pencil out.

The habit worth keeping: measure cost per finished task, not price per token. Run the same ten real jobs through your current model and Opus 5.5, count what each one costs and how often a person had to fix the result, then decide.

That's the kind of comparison we do before recommending a model for a client workflow at New Face Design. If you want to know which of your processes are worth automating at today's prices, our free process audit is a good place to start.