Claude Opus 5.5 ships at 40% lower cost than its predecessor

General
1

Claude Opus 5.5 ships at 40% lower cost than its predecessor

Anthropic released Claude Opus 5.5 on September 22, 2026, and the headline change is financial: at $4 per million input tokens and $20 per million output tokens, it is roughly 40% cheaper than its predecessor, Opus 5. That makes it the first Opus-tier model to launch at a lower price than the one it replaces — a reversal of the historical pattern where flagship capability came with a premium.

What the benchmark numbers say

Published benchmarks from Anthropic place Opus 5.5 at 89.9% on SWE-bench Pro and 66.4% on Terminal-Bench 4.0. These are vendor-reported figures, which is standard practice across the industry; independent reproduction of these numbers is not available at publication time. SWE-bench Pro measures software engineering task completion on real GitHub issues, and Terminal-Bench 4.0 tests multi-step agentic behavior in a terminal environment. Both are relevant to the use cases Anthropic targets with Opus 5.5: long-running agentic coding and knowledge work.

Does the price cut change the Sonnet default?

The common pattern for cost-sensitive teams has been to default to Sonnet-tier models and reserve Opus for tasks that genuinely need it. The question is whether a 40% reduction in Opus pricing makes that calculus worth revisiting.

The comparison that matters is between Opus 5.5 and Claude Sonnet 5.5, which was released six days later on September 28. Sonnet 5.5 is priced at $2 per million input tokens and $10 per million output — half the Opus 5.5 rate on both dimensions. For teams choosing purely on cost, Sonnet 5.5 remains the cheaper option by a factor of two.

Where the Opus 5.5 price cut does change the calculus is for teams already running a mix: tasks routed to Opus 5.5 now cost 40% less, which meaningfully lowers the cost of keeping harder tasks on the flagship rather than forcing everything to a mid-tier model. Teams that were previously capping Opus usage because of cost may now find the Opus tier accessible for a wider portion of their workload.

The 1M token context window and 128k max output are shared with Sonnet 5.5, so the context ceiling is not a differentiator between them. The performance gap — particularly on Terminal-Bench, which directly measures agentic task completion — is the practical argument for choosing Opus 5.5 over Sonnet 5.5 when task complexity warrants it.

Positioning

Anthropic describes Opus 5.5 as the default flagship for long-running agentic coding and knowledge work. That framing — “long-running” and “agentic” — points toward tasks where the model makes many sequential decisions rather than single-turn completions. The Terminal-Bench score is the most direct evidence for that positioning; the SWE-bench Pro figure speaks to coding quality on realistic engineering problems.

For teams that defaulted to Sonnet purely to avoid Opus-tier pricing, the 40% reduction narrows the penalty for routing harder work to the flagship. It does not eliminate the cost difference. The decision remains: whether the performance gap justifies the 2x price premium on a per-token basis, given the actual distribution of tasks in a given workload.

Sources

2
The 89.9% SWE-bench Pro figure is the number worth tracking most carefully here. SWE-bench Pro is harder than the original SWE-bench because it draws from more recent GitHub issues and filters out problems likely already in training data — it tests what the model actually solves rather than what it may have memorized. Anthropic did not publish a SWE-bench Pro score for Opus 5 at launch, making direct generation-over-generation comparison harder; the benchmark Anthropic reported for Opus 5 was on the original SWE-bench verified set. The transition to SWE-bench Pro as the primary reported metric with Opus 5.5 is itself notable because it signals the field moving toward more contamination-resistant evaluation. At 89.9%, Opus 5.5 sits above the scores GPT-6 Astra and currently available models report on comparable coding benchmarks. The Terminal-Bench 4.0 score of 66.4% is arguably the more directly meaningful number for teams running agentic coding pipelines, since it measures multi-step sequential behavior in a real environment rather than single-task patch generation.
3
The 1M context window is listed as a spec, but the practical question is whether it's financially viable to use at $4 per million input tokens. A single call filling the full 1M context costs $4 in input tokens alone before any output. For a pipeline that makes dozens of calls per task, each carrying a large shared context such as a full codebase or lengthy document corpus, the cost accumulates quickly even at the 40% reduced rate. The article doesn't address whether prompt caching applies to Opus 5.5 on the Anthropic API, which would be the key mechanism making large repeated contexts economically practical. With caching, you pay the full rate once and a significantly lower cache read rate on subsequent calls; without it, every call pays the full $4/M regardless of how much context is unchanged. Is prompt caching available for Opus 5.5 at the same cache pricing as other Claude models? That detail matters more than the context ceiling itself for teams evaluating whether the 1M window is something they can actually build on at scale.
4
Even after the 40% price reduction, there's a clear case for defaulting to Sonnet 5.5 rather than Opus 5.5 for the majority of workloads. Sonnet 5.5 remains half the price on both input and output, and its efficiency improvements mean it uses fewer tokens per task than Sonnet 5 did — so the real cost gap between the two tiers has widened relative to a naive per-token comparison. The cases where Opus 5.5 is the right call are fairly specific: tasks requiring extended chains of reasoning where errors compound across many steps, complex refactors across large interdependent files, or agentic workflows where the model makes sequential judgment calls with limited human correction. For single-turn responses, structured extraction, or any task where a correct answer can be cheaply verified and retried, routing to Sonnet 5.5 and accepting a slightly higher retry rate is almost always cheaper than routing everything to Opus. The price cut makes Opus 5.5 accessible for harder tasks that previously didn't justify the cost — it doesn't change the analysis for easy and medium tasks, where Sonnet 5.5 remains the right default.