Anthropic’s Claude Opus 5.5 is cheaper and faster to run than Opus 5, but a small set of reasoning tests found no clear capability improvement. In The New Stack’s comparison, both models solved two problems and failed a difficult ordering task, while Opus 5.5 sometimes used substantially fewer output tokens.
Lower API pricing
Anthropic lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens, compared with $5 and $25 for Opus 5. The source estimates that the direct price reduction accounts for a 20% saving; the larger reduction claimed by Anthropic depends on the newer model generating fewer tokens. Anthropic also claims Opus 5.5 costs 40% less on typical workloads and produces output more than 30% faster.
Mixed reasoning results
The comparison used identical prompts through Anthropic’s API, with adaptive thinking enabled at the default effort level. Each problem was run once per model. Both systems completed a logic-grid puzzle correctly and returned the same correct results for a stone-removal game. On the constrained-ordering problem, neither produced an answer, even after the output allowance was increased.
Faster, but below the claim
Opus 5.5 completed the logic-grid task in 65 seconds for $0.16, versus 108 seconds and $0.27 for Opus 5. On the stone game, it used 28,740 output tokens and cost $0.58, while Opus 5 used 74,981 tokens and cost $1.88. Across the calls, the newer model averaged 103.4 output tokens per second against 93.1 for Opus 5, an improvement of about 11%, below Anthropic’s stated 30% figure.
Long failures still matter
The ordering test exposed a practical limitation: both models spent extensive time thinking without producing a result. With a 128,000-token limit, Opus 5.5 ran for 19 minutes, consumed 112,733 tokens and ended with a refusal stop reason. The source argues that this appeared to be an erroneous safety-filter response to a harmless counting problem. The results support lower costs for some workloads, but they do not establish better reasoning from this limited, single-run comparison.
AI-generated editorial illustration; not a documentary photograph.
Source: Claude Opus 5.5 vs. Opus 5 on reasoning tasks: Cheaper, faster, but not better
SIGNAL TITANS
Leave a Reply