Anthropic cuts Opus 5.5 prices 20% on tokens, 60% on cache reads, citing fewer tokens burned for 40% total savings

Every line on the new price sheet moved 20% except one: input from $5 to $4 per million, output from $25 to $20, cache writes from $6.25 to $5, and cache reads from $0.50 to $0.20, which is 60% ^[2][3][4][1]. Anthropic says the saving comes from two places, cheaper tokens and the model burning fewer tokens per job ^[5].

Take the price half on its own. Half the bill in cache reads gets you to the headline: 0.5 times 0.4 plus 0.5 times 0.8 is 0.6, a 40% cut ^[2]. Below that share the saving slides toward 20%, so a team whose spend is mostly output tokens gets a fifth off ^[2].

The consumption half carries the bigger numbers. Anthropic said a tester audited and repaired a 200,000-line codebase in under three hours, work that took Opus 5 more than 20 hours and 2.5 times the tokens ^[11]. At 40% of the tokens and 80% of the price, that job runs at 32% of the old cost, and 16% if it is cache-read heavy ^[3]. Another tester migrated 680,000 lines of code in less than a day ^[12].

Against Fable 5.1 at $10 and $50 per million, Opus 5.5's base rates are 60% lower on both lines ^[6][6]. An internal test translated the HAProxy load balancer from C to Rust in 9.5 hours against Fable's 12, at 51% less cost ^[13].

The OpenAI comparison runs the other way. GPT-6 Sol launched the same Tuesday at $2 and $10 per million, half Opus 5.5's base rates ^[18]. On Zapier's AutomationBench, Cryptopolitan reported Opus 5.5 at 40.0% for about $1.28 a task and Sol at 32.0% for $0.34 ^[19], which is 31 points of score per dollar against 94 ^[4]. Anthropic said benchmark margins are a weaker guide to real differences at this capability level ^[10].

In my view the cache-read line is the one that will show up in bills, because Anthropic says cache reads drive most of the cost in coding and agentic work ^[4]. The counter-case is plain: a shop whose spend is mostly output tokens sees 20% and stops there ^[1][2]. Anthropic did not disclose what serving those tokens costs it, and a cut of this shape is consistent with cheaper inference and with buying share at a thinner margin.

Consumption is what would undo the budget case. Cryptopolitan has reported a forecast that inference cost per agentic workflow climbs more than five times by 2028 ^[20]. Cheaper tokens do not contradict that forecast; they lower the rate at which a longer loop bills. Opus 5.5 is on the API as claude-opus-5-5 and on Amazon Web Services, Google Cloud and Microsoft Azure ^[21].