Claude Opus 5.5: Fable-Level Performance at 40% of the Price. What It Means for AI Agents

The number worth reacting to in the Opus 5.5 launch is not a benchmark, it is a ratio. Anthropic is claiming its previous flagship's performance for roughly 40% of the price, and on the numbers so far the claim mostly holds. The frontier did not get smarter this week. It got cheaper, again, which for anyone running agents at volume is the more consequential kind of news.

Anthropic released Claude Opus 5.5 on 22 September as the first model in a new Claude 5.5 family and its new default flagship. It performs at the level of Claude Fable 5.1 on most work, costs about 40% less than Opus 5 on typical workloads, and, in a detail we will come back to, it finally fixed the way Opus talks.

What Opus 5.5 is

Opus 5.5 is priced at $4 per million input tokens and $20 per million output, a 20% cut from Opus 5's $5 and $25, with cache reads down sharply to $0.20 per million and cache writes to $5. Because it also uses fewer tokens to finish a task, Anthropic puts the real-workload saving closer to 40%, and it generates output more than 30% faster than Opus 5. It is available everywhere today, through the API and on AWS, Google Cloud, and Azure, as claude-opus-5-5.

The comparison that matters is not against Opus 5, though. It is against Fable 5.1, the model people actually reach for, which costs $10 and $50 per million. Opus 5.5 delivers similar quality at 40% of that price.

The benchmarks, and where they stop

On Anthropic's own table, Opus 5.5 beats Fable 5.1 on all nine benchmarks while costing 60% less, which is the headline. It posts 89.9% on SWE-bench Pro and 66.4% on Terminal-Bench 4.0 against Fable's 55.8%. It is not a clean sweep of the whole field: against GPT-6 Astra it trails on business workflows by about a point and a half and on agentic scientific research by around six points. But for most agent work, it is at or near the top at a fraction of the cost.

The efficiency story hides a sharper lesson about effort

The cost-per-task numbers are where it gets interesting, and where there is a real lesson for how you run it. On Terminal-Bench 4.0, Opus 5.5 at medium effort scores 57.6% for about $2.94 a task, matching Astra's high-effort 57.9% at $7.21, roughly 40% of the cost. At its default setting it beats Opus 5 running at maximum for about a fifth of the price.

Look closer and the pattern is one we keep seeing: the returns flatten fast as you push effort up. On several of Anthropic's own charts, medium effort lands within a fraction of a point of max, at a small fraction of the cost, and on one coding set medium actually edges out max. Artificial Analysis found Opus 5.5 at max effort burning around 119,000 output tokens per task, well above Fable and far above Astra's roughly 27,000. So the model is remarkably efficient at its default setting and genuinely token-hungry at max, which is the clearest argument yet that reasoning effort is a dial to tune per step, not a default to leave on maximum. With Opus 5.5, cranking effort to max is usually paying four times more for the same answer.

It also fixed the way Opus talks

One of the loudest complaints about Opus 5 was not capability, it was communication. It buried the point under jargon and idiosyncratic phrasing, to the extent that the "Opus slop" description became a running joke. Anthropic says Opus 5.5 is about 40% less verbose without losing accuracy, puts the most important information first, and actually follows the writing rules you give it. Among the specific habits it dials back, and this is a genuinely satisfying one to report, is the model's overuse of em-dashes. For agent work this matters more than it sounds, because an agent's output is often read by another step or another agent, and a model that follows format instructions and gets to the point is easier to build a reliable pipeline around.

The catch worth knowing

There is a tighter safety posture, and it is worth going in with eyes open. Opus 5.5 reroutes most cybersecurity questions to the older Opus 4.8, and gates biology-related work behind a Life Sciences Verification Program, so if a request would be refused on Fable it will be rerouted here too. It also carries Fable's anti-distillation protections, blocking attempts to edit Claude's prior context or extract its reasoning. That is sensible for enterprise deployment, but if your workload legitimately touches security or life sciences, test the rerouting behaviour before you commit, because the model you think you are calling may not be the one that answers.

What it means for AI agents

For an agent stack, Opus 5.5 is the strongest new default it has had in a while: frontier-adjacent quality, meaningfully cheaper, and faster. The move is not to switch everything to it blindly, it is the same route-per-step discipline as always, just with a better default in the mix. Point Opus 5.5 at the broad middle of demanding work where it now leads on cost per completed task, keep GPT-6 Astra for the specific things it still wins, run it at default rather than max effort unless a task proves it needs more, and confirm the safety rerouting does not sit in the path of anything you depend on. The through-line of this whole year holds: the model layer keeps getting cheaper and faster, and the advantage is the orchestration that lets you adopt a better default the day it ships.

What people are saying about Opus 5.5

In the first hours the reaction was warmer than Opus 5's, and the reasons are telling. The most common relief was about communication, with developers glad to see the end of the jargon-heavy "Claudish" style. The most common strategic question was sharper: if Opus 5.5 matches Fable 5.1 at 40% of the price, what is Fable for now, which is a fair thing to ask of your own model lineup.

Not everyone is sold. Some users burned by earlier releases and the tightening usage limits said they would stick with other providers regardless, and others welcomed the raised five-hour limits and the rate-limit reset as overdue. There was also the recurring frustration with over-refusal, the sense that the safety posture sometimes gets in the way of legitimate work. The most encouraging signal was practical: people who had been silently served Opus 5.5 in the week before launch reported noticeably better results without knowing why. As always, treat the launch numbers as a reason to pilot, and confirm the gains on your own workflows over the next few weeks.

Common questions about Claude Opus 5.5

How much cheaper is Claude Opus 5.5?

It costs $4 per million input tokens and $20 per million output, about 20% less than Opus 5 per token and roughly 40% less on typical workloads because it uses fewer tokens. Against Fable 5.1, which it matches on most work, it is about 40% of the price, since Fable runs at $10 and $50.

Is Opus 5.5 better than Fable 5.1 or GPT-6 Astra?

It beats Fable 5.1 across Anthropic's benchmark table at a large discount, so for most work it is the better value. It trails GPT-6 Astra on a couple of things, business workflows and agentic scientific research, so Astra still wins those specific jobs. For most agent work Opus 5.5 is now the stronger default.

Should I run Opus 5.5 at maximum effort?

Usually not. Its own benchmarks show medium effort landing within a fraction of a point of max on many tasks at a fraction of the cost, and at max it uses far more tokens per task. Run it at default or medium, and only raise the effort where a specific task demonstrably needs it.

What are the downsides of Opus 5.5?

A tighter safety posture. It reroutes most cybersecurity questions to Opus 4.8 and gates biology work behind a verification program, so security and life-sciences workloads should test the rerouting first. It is also token-hungry at maximum effort, which is a cost trap if you leave effort maxed by default.