Claude Opus 5.5 tops the Artificial Analysis index at $4/$20, using 1.6x Opus 5's output tokens per task (2026)

Anthropic released Claude Opus 5.5 on September 22, the first model in a new Claude 5.5 family, priced below the model it replaces: $4 per million input tokens and $20 per million output, against Opus 5's $5 and $25. 1 It is on AWS, Google Cloud, Microsoft Azure and the Claude Platform as claude-opus-5-5. 1

Anthropic says it matches Claude Fable 5.1, its larger sibling, on most work while costing 40% less to run on typical workloads than Opus 5. 1 Artificial Analysis measures it at 58 on its Intelligence Index, the highest that firm has recorded. 2

Signal	Confirmed detail	What it means for you
What shipped	Adaptive thinking is always on; `effort` sets thinking depth and defaults to `medium`, where Opus 5 defaulted to `high`. 3	One model ID to change; a reasoning default that moved.
What it costs	$4 and $20 per million input and output tokens; cache reads fall to $0.20 from $0.50. 1	Cache-heavy agent loops gain most.
The catch	About 119,000 output tokens per index task at max effort, against Opus 5's 73,000. 2	Cost per task is what reaches your bill.
What gets rerouted	Most cybersecurity tasks fall back to Opus 4.8; biology and frontier-model work goes to Opus 5. 1	Check where your workload lands first.

Where Anthropic's own numbers stop

The vendor table puts Opus 5.5 ahead of Fable 5.1 on every row it published, and warns that at this capability level benchmark margins are a less reliable guide, with the gap to Fable 5.1 narrower than the scores suggest. 1 Its safeguards intervened during evaluation too, handing cybersecurity tasks to Opus 4.8 and biology and frontier-model tasks to Opus 5, which it says likely reduced its own scores. 1 Two rows still run the other way: Zapier's AutomationBench gives GPT-6 Astra 41.4% to Opus 5.5's 40.0%, and Terminal-Bench-Science 0.1 gives Astra 64.6% to 58.7%. 1

Where it stands on someone else's scale

Artificial Analysis ran it with Anthropic's default fallback enabled. It leads six of the index's ten evaluations, taking Humanity's Last Exam to 61.4% against a previous best of 59.1%, and trails on CritPt, AA-LCR and GDP.pdf. 2 That firm also supplies the number complicating the price story: Opus 5.5 spends about 1.6 times Opus 5's output tokens per index task, leaving its cost per task level with Opus 5's at roughly $5.98. 2 4

Two stacked Artificial Analysis panels: a bar chart of Intelligence Index scores where Claude Opus 5.5 is the tallest bar at 58, and a scatter plot of index score against cost per index task with a dotted Pareto line — Artificial Analysis's Intelligence Index and its cost-per-task frontier, in the copy The Decoder published; the top panel places Claude Opus 5.5 at 58 with Fable 5.1 and GPT-6 Astra at 53, and the lower panel plots tracked models by score against cost per index task. 2 4 5

Before you move an Opus 5 integration

Four changes break code already running on Opus 5. Thinking can no longer be switched off: a request carrying thinking: {"type": "disabled"} or a manual token budget returns an error. Forced tool use is rejected, so tool_choice set to any or to a named tool becomes auto. Thinking blocks are bound to the model that produced them, so a conversation moved from Opus 5.5 outside the Opus 5 lineage loses the earlier reasoning. The older computer_20251124 computer-use tool is refused on the Claude API and Google Cloud. 3

Why it matters

Anthropic is cutting price into a field where OpenAI's GPT-6 Astra and cheaper Chinese models hold the lower-cost position; The Decoder reads the cut as a response to that pressure. 5 This is Anthropic's first release since chief executive Dario Amodei called for pacing the frontier, and the company places it inside that argument. 1 6 Sonnet 5.5 and Haiku 5.5 follow in the coming weeks, without dates. 1