Benefits of Moving Production Agents to Claude Opus 5.5

If your production agents have been running on Claude Opus 5, the math changed on September 22, 2026. Anthropic cut the price, raised the agentic coding scores, and kept the 1M-token context window in the same release. Moving production agents to Claude Opus 5.5 is a one-line model ID swap, but it touches your bill, your latency, and a few things that will return a 400 error if you skip them. Below is what you actually gain, and what to check before you flip it in prod.

The bill drops before you change a line of prompt

Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5 and $25. Cache reads fall further, to $0.20 per million tokens (5% of the new base price, versus $0.50 on Opus 5), and batch requests stay at 50% off both directions. Anthropic puts the combined effect at roughly 40% lower cost for typical token-billed workloads, since most of a long-running agent's tokens are repeated system prompt and tool schema served from cache, not fresh input.

Opus 5:    10M in x $5  + 2M out x $25 = $100/day
Opus 5.5:  10M in x $4  + 2M out x $20 = $80/day  (20% off list price)

That's the list-price case. Add cache reads and the gap widens toward Anthropic's 40% figure.

Effort replaces "thinking on or off" as your cost dial

Opus 5.5 can't turn thinking off: a request that sends thinking: {"type": "disabled"} now returns a 400 error. What replaces it is output_config.effort, one parameter with five levels (low, medium, high, xhigh, max) that controls how much the model reasons, calls tools, and writes, in one knob instead of a thinking budget plus a model choice.

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=4096,
    output_config={"effort": "high"},
    messages=[
        {"role": "user", "content": "Summarize the failed deploy and propose a rollback plan."}
    ],
)

Set effort explicitly. Opus 5.5 defaults to medium, one step down from the high default on Opus 5, so a request that omits the field now runs cheaper and thinner than it did before the swap. That's a benefit if you tune it on purpose, and a silent regression if you don't.

It wins the benchmarks closest to the job

The benchmarks that matter for a production agent look like the job: long tool-calling sessions in a terminal, not a single-turn quiz. On Anthropic's published numbers, Opus 5.5 scores 66.4% on Terminal-Bench 4.0, beating GPT-6 Astra at roughly 40% of the cost; 54.4% on FrontierCode v1.1 against GPT-6 Astra's 53.3% for about a fifth of the price; and 57.8% on CursorBench 4.0 versus GPT-5.6 Sol's 41.7% for about a third of the cost. Those three track what a coding or ops agent actually does all day: read a repo, run commands, fix what breaks.

The same 1M-token context, now cheaper to fill

Context and output limits didn't move. Opus 5.5 still gives you a 1M-token context window and a 128K-token output ceiling on the synchronous API (300K on the Batch API with the output-300k-2026-03-24 beta header). What changed is the price of filling it: an agent that keeps a full repo, a ticket history, or a multi-hour transcript in context now pays the lower $4 input rate for every one of those tokens on every turn.

Fast mode for latency-bound agent steps

Opus 5.5 also ships a research-preview fast mode: $8 per million input tokens and $40 per million output tokens for up to 2.5x the speed, on the Claude API only (not yet on Bedrock, Google Cloud, or Microsoft Foundry).

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: fast-mode-2026-02-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5-5",
    "max_tokens": 4096,
    "speed": "fast",
    "messages": [{"role": "user", "content": "Classify this support ticket and draft a reply."}]
  }'

It costs double the standard rate, so save it for the interactive step a human is waiting on, not the background batch job.

When it's not worth switching yet

Four things break on the swap immediately, not eventually. If your Opus 5 integration sets thinking: {"type": "disabled"}, forces a specific tool with tool_choice: {"type": "tool", ...} or "any", relies on the older computer_20251124 tool for computer use on the Claude API or Google Cloud (Bedrock is unaffected), or swaps models mid-conversation expecting old thinking blocks to carry over, fix that code before you change the model ID. None of it is hard, replace disabled thinking with a lower effort, swap forced tool choice for auto plus strict tool use, but it's real migration work, not a drop-in.

And if your agents already run on Sonnet 5 for anything short of long-horizon coding, this release doesn't change that trade. Sonnet 5 is still $2/$10 per million tokens and faster. Opus 5.5 is the upgrade for agents that were already justifying Opus 5's price on capability, not the ones you put there by default.

Cheaper, ahead on the benchmarks that resemble the job, and priced to make the context window you already pay for less expensive to actually use: for a team already running agents on Opus 5 in production, that's a rare upgrade where the cost case and the migration checklist both fit on one page.