Claude Opus 5.5 Got Cheaper: The AgentSpace Bill

In a lead-worker architecture, the most expensive layer is the lead agent. Sprint planning, task decomposition and writing briefs for workers all go to the strongest model, because a mistake there is the most expensive one to fix. But that same layer is also the fattest line item on your bill. So when a cheaper top-tier model ships, the first question isn’t hype — it’s arithmetic: can I get the same work, at the same quality, for fewer tokens?

Anthropic announced Opus 5.5

Anthropic shipped Claude Opus 5.5 on September 22. According to the company, it sets a new bar for coding and knowledge-work performance, beating its own larger Fable model on many benchmarks. Pricing dropped too: output tokens are now $20 per million, down from $25 in the previous version — and the model runs faster. On the current price list Opus 5.5 is $4 per million input tokens and $20 per million output tokens (TechCrunch, Anthropic pricing).

Six days later, on September 28, Sonnet 5.5 shipped at $2 input and $10 output per million tokens, exactly half the price of Opus 5.5 (TechCrunch). Haiku 5.5 is not on the price list as of October 2; the fast tier is still Haiku 4.5 (Anthropic pricing).

Anthropic CEO Dario Amodei framed this release as part of a deliberately paced approach to capability increases:

“I have become convinced that fully addressing the risks requires even more prudence, not just investing in risk prevention, but pacing the rate of capabilities advancement so that risk prevention has time to keep up.” (TechCrunch)

The piece also notes that Opus 5.5’s cybersecurity and biology capabilities are treated as close to Fable-level, which is why it’s held to the same safety constraints.

What this actually means for us

In AgentSpace, we keep the lead agent role bound to Claude’s top tier — sprint planning, task decomposition, and writing briefs for workers all go through it, because a mistake at that stage sends workers hours in the wrong direction before anyone notices. That makes the lead agent the single most expensive line item in a sprint’s bill. So “$20 per million output tokens, down from $25” isn’t an abstract headline for us — it’s a number that lands directly on our lead-agent cost line.

The arithmetic is clear: a 20% drop in list price does not automatically mean a 20% drop in our actual sprint bill. What determines our real cost isn’t just the per-token rate — it’s also how many planning rounds the model runs, how much context it reads, how many back-and-forths it needs with a worker before a task is accepted. A model can be cheaper per token and still be more “talkative,” in which case the total bill doesn’t shrink the way you’d expect from the price cut alone. So on our next live sprint, I plan to run Opus 5.5 as lead agent against the same task set we’ve run before, and measure it directly against the previous version. If there’s a before/after delta worth showing, I’ll write it up separately — for now what I have is a plan, not a result.

The second point gets less attention than pricing but matters more to us in practice: Anthropic says Opus 5.5 uses less jargon and puts the important information at the top of a message. In AgentSpace, the lead agent’s output goes to two places — the task brief that reaches workers, and the sprint summary that reaches me. For both, we’ve had to write prompt-level constraints like “state the outcome first, then the detail,” because the model would sometimes bury the key decision in the third paragraph. If that behavior actually changes at the model level, some of the constraints we’ve been forcing in the prompt might become unnecessary — but I’m leaving that as “probably,” not claiming it until I’ve measured it.

Sonnet 5.5 landing on September 28 at half the price of Opus 5.5 changes the math a second time. Our worker pool runs across different Claude tiers, and the mid-tier workers can now move to a $2/$10 model. The open question: can Sonnet 5.5 take over part of the lead agent’s work at half the price too? So the test has two arms: Opus 5.5 on the lead layer, Sonnet 5.5 on the worker layer. When Haiku 5.5 reaches the price list, I’ll add the third arm.

There’s one more detail in the piece worth pulling out: the model runs faster, which reflects a broader drop in the total compute required to serve it. For us, that can be a gain separate from token price entirely. The lead agent repeats its planning step multiple times within a single sprint: split the task, evaluate what came back from a worker, revise the plan if needed. Every bit of latency in that loop adds to total sprint time. If the model responds faster, the same number of planning rounds finishes sooner — a separate efficiency gain, independent of per-token cost. I haven’t measured this in wall-clock hours yet either; I can just see architecturally where it would show up.

What isn’t changing is just as clear: the lead-worker architecture, the task board, the persistent memory structure, and the tool-based permission limits we give worker agents all stay exactly as they are. A cheaper model isn’t a reason to hand an agent broader authority — we keep that decision separate. Amodei’s framing of “pacing capability growth to match the speed safety measures can keep up with” points in the same direction as something we already do at small scale: no matter how powerful a model gets, the permissions we grant a worker agent are set by the task, not by the model’s raw power. That principle didn’t change with Opus 5.5 — it just got reconfirmed.

The added constraints around biology and cybersecurity capabilities don’t create any practical difference on our end — AgentSpace workers don’t operate on tasks like exploit development or biological design, so those constraints simply fall outside our use case.

What to do today

If you’re running a lead-worker system yourself, or running agents through a single strong model, set up this comparison before you migrate on the strength of a pricing announcement alone:

Opus 5.5 migration checklist:
1. Run the same task set across two separate sprints:
   - Sprint A: your current lead-agent model
   - Sprint B: Opus 5.5
2. Log separately for each sprint:
   - input/output token counts
   - total $ cost
   - number of planning rounds / worker round-trips
3. Compare output quality:
   - does the key information land at the top of the message
   - is there unnecessary jargon
4. Result: calculate $ per completed task for each version separately —
   don't just look at list price
5. Repeat the same test on your worker pool with Sonnet 5.5 ($2/$10, half of Opus 5.5)

The list-price drop is real and confirmed. But the actual delta that lands on your sprint bill only shows up after you run this test.

Closing

The instinct when a model gets cheaper is to migrate immediately; I measure first instead, because a lead-agent mistake burns worker hours that a cheap token doesn’t refund. I’m running this test on AgentSpace’s next live sprint — if you want to see the lead-worker architecture and this test’s results firsthand, you can join the Muratify AI community.