All articles

Claude Opus 5.5 beats Fable 5.1 and costs less

Opus caching at Sonnet prices: a flagship cheaper than the one it replaces

Claude Opus 5.5 shipped today, September 22, and I have been working with it in Claude Code since it showed up in the model picker. It costs $4 per million input tokens and $20 per million output, down from $5 and $25 on Opus 5. Anthropic says it performs at the level of Claude Fable 5.1, which costs $10 and $50, and on the benchmarks it published it comes out ahead. A new model that does more and costs less than the one it replaces. That has not happened in a while.

A few hours in, one thing stands out: the answers are more precise. The numbers Anthropic published point the same way, and they say more than the headline does.

Opus pricing had not moved since November 2025

The last time Opus changed price was 4.5, when it went from $15 and $75 to $5 and $25. Then came 4.6, 4.7, 4.8 and 5, all at the same rate. Ten months, four versions, not a cent of difference.

It is a flat 20% cut on every row but one.

Opus caching now costs what Sonnet's does

The odd row is the cache read. Until now a cache hit cost a tenth of the input price on every model. On Opus 5.5 the multiplier drops to 0.05x, which on a $4 input is $0.20.

Put that next to the rest of the lineup and it looks strange:

• Sonnet 5 has the same cache read, $0.20;
• Fable 5.1 is at $0.25, above Opus;
• Haiku 4.5 is at $0.10, half of it.

That column matters more than it looks, because in an agent it is the line that grows with the number of steps. The system prompt, the tool definitions and the history get re-read on every tool call. Forty calls at an average context of 150k tokens is six million cached tokens. That was $3 on Opus 5. It is $1.20 on Opus 5.5. Output is still the biggest line, but it is the one that grows least as a session gets longer.

The 40% that is not in the price list

The launch page has one sentence worth more than the table. At default settings, Opus 5.5 costs 40% less than Opus 5 on typical workloads. The price list alone accounts for 20%. The rest has to come from somewhere.

Rough math: 0.8 times 0.75 is 0.6. Even after you count the cheaper cache, part of that gap only makes sense if the model uses fewer tokens for the same task. The customers Anthropic quotes say so outright:

• Optiver: Opus 5 quality in about half the turns, time and output tokens, for 40 to 50% less;
• Kiro: more tasks solved with about 40% fewer calls and half the tokens;
• GitHub: in VS Code, more terminal tasks solved in less than half the steps.

A technical detail says the same thing from another angle. On several benchmarks, Opus 5.5 at default effort beats Opus 5 at max effort. And it generates output over 30% faster.

My read, and it is an inference, not something Anthropic explains: they did not just shave the margin. The model gets to the answer in fewer steps, and fewer tokens per answer means less compute to serve it. That is what lets you cut the price without losing money. It would also explain why it did not feel crowded today. Launch day is usually when a new model is at its slowest.

Above Fable on benchmarks, at under half the price

The widest gap is on Terminal-Bench, over ten points. OSWorld is close to a tie. The internal test that caught my eye is a different one: porting HAProxy from C to Rust. Both models passed nearly all of its regression tests, but Opus 5.5 finished in 9.5 hours against 12 for Fable 5.1, at 51% lower cost.

I don't run benchmarks. I use it to get work done, and that is where I notice the difference. It finds the problem sooner, and once it does, it looks at what surrounds it instead of stopping at the broken line. Its writing is clearer, so it is easier to follow what it is doing. That is a few hours of use, but it lines up with the table.

Fable 5.1 is still $10 and $50. Fable is not part of the 5.5 announcement, and until it moves too, I struggle to find a reason to pay two and a half times as much for day-to-day work.

Sonnet 5.5 and Haiku 5.5: the whole lineup shifts

Opus 5.5 is the first model in the 5.5 family. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks, with many of the same gains in performance, efficiency and safety. No dates, no prices.

Haiku is the one I care about. Haiku 4.5 came out in October 2025 at $1 and $5. In almost a year Opus got four versions and Haiku got none. Yet Haiku is what ends up in the nodes that run most often in an agent system: classification, routing, field extraction. Those are the nodes I wrote about in You are paying an LLM to do an if, the ones that do need a model, just a small one. If Haiku 5.5 makes the same efficiency jump as Opus, that is where the bill moves most, because that is where calls are counted in thousands.

Sonnet 5 has been at $2 and $10 since its launch price became permanent on September 1. Whether Sonnet 5.5 keeps that price, Anthropic has not said.

Where to redo the math

The model ID is claude-opus-5-5. Next time you pick a model for a new project, or open the API spend dashboard at the end of the month, redo the table with these numbers. If you put something on Fable just to be safe, that is the first node to try again on Opus 5.5.

I would not lock in the budget for the small nodes yet, though. In a few weeks that changes too.