{"content_id":"7eujffjzrh","slug":"claude-opus-5-5-pricing-performance-comparison","locale":"en","schema_type":"Article","category":"ai_data","category_name":"AI Data","title":"Claude Opus 5.5 Pricing and Performance Comparison","summary":"Claude Opus 5.5 has cut its standard input and output token rates by 20% compared with Opus 5. The 40% reduction in task costs is based on Anthropic's evaluation; actual costs depend on token usage and cache use.","sponsorship_disclosure":null,"affiliate_disclosure":null,"commerce_disclosure":null,"author":{"name":"Injoys Editorial Team","url":"https://injoys.com/ko/about"},"key_points":["Claude Opus 5.5 standard rates are $4 per million input tokens and $20 per million output tokens.","The cache read rate is $0.20 per million tokens, 60% lower than Opus 5.","According to Anthropic, its Terminal-Bench 4.0 score rose from 52.3% to 66.4%.","The roughly 85% reduction in attempts to bypass boundaries comes from a specific safety evaluation.","To migrate an existing API, check the model name as well as inference and tool-calling settings."],"content_markdown":"Claude Opus 5.5 lowers standard token prices by 20%. Its coding and computer use evaluation scores have also improved. The 40% reduction in task costs is a result from Anthropic's testing.\n\nPrices and evaluation figures are based on Anthropic's September 22, 2026 announcement.\n\n## Claude Opus 5.5 Price Comparison\n\nStandard API prices are $4 per 1 million input tokens and $20 per 1 million output tokens. Tokens are the units a model uses to process text and other content. Input means content sent to the model, while output means content the model generates.\n\n| Billing item | Opus 5 | Opus 5.5 | Price reduction |\n| --- | ---: | ---: | ---: |\n| 1 million standard input tokens | $5 | $4 | 20% |\n| 1 million output tokens | $25 | $20 | 20% |\n| 1 million cache read tokens | $0.50 | $0.20 | 60% |\n\nAmounts are in US dollars. Caching stores inputs used repeatedly so they can be reused. The cost of creating a cache is calculated separately from the cost of reading it.\n\n## Conditions for 40% Savings on Task Costs\n\nThe 40% figure is not a discount that applies to every request. It is Anthropic's estimate of savings on typical tasks with default settings. It reflects both lower token prices and fewer tokens used per task.\n\n| Usage pattern | What affects cost | What to check |\n| --- | --- | --- |\n| Entering new content each time | Standard input and output usage | Actual token counts for both |\n| Reusing the same document or instructions | Cache creation and read usage | Whether the cache was actually reused |\n| A coding agent performing multiple steps | Repeated calls and retries | Total cost to complete the task |\n| Monthly subscription | Subscription price and usage limits | The distinction between token prices and subscription fees |\n\nLower API prices alone do not mean lower monthly subscription fees. To determine the savings for your service, compare usage for the same tasks. [Anthropic's Opus overview](https://www.anthropic.com/claude/opus) also describes the 40% figure as an estimate for token-billed tasks.\n\n## Calculation Example\n\nUsing 1 million input tokens and 1 million output tokens costs $24 at standard prices. For this comparison, assume usage is the same for both models. This calculation excludes caching and separate tool costs.\n\n| Calculation item | Opus 5 | Opus 5.5 |\n| --- | ---: | ---: |\n| 1 million standard input tokens | $5 | $4 |\n| 1 million output tokens | $25 | $20 |\n| Total | $30 | $24 |\n\nThe difference is $6, a 20% saving. This calculation does not include any reduction in tokens used per task. The result based on prices alone therefore differs from the announced 40%.\n\nThe price difference for reading 1 million cached tokens is $0.30. That amount compares only the read charge, however. The total bill must also account for cache creation and other token costs.\n\n## Agentic Coding and Computer Use Performance\n\nIn Anthropic's announcement, both coding and computer use scores improved. Agentic coding means using tools to carry out multistep development tasks. Computer use is the ability to view a screen and operate an interface.\n\n| Evaluation | Opus 5 | Opus 5.5 | Score difference |\n| --- | ---: | ---: | ---: |\n| Terminal-Bench 4.0 | 52.3% | 66.4% | 14.1 percentage points |\n| CursorBench 4.0 | 46.6% | 57.8% | 11.2 percentage points |\n| OSWorld 2.1 | 74.0% | 81.8% | 7.8 percentage points |\n\nThe evaluation name in the official announcement is OSWorld 2.1. Its score is marked `partial`. Keep both the version and the scoring label to avoid confusing it with other results.\n\nThe Opus 5.5 result for Terminal-Bench 4.0 uses the `xhigh` setting. Its evaluation conditions are not the same as those used for the announcement's cost estimate with default settings. A higher score alone cannot tell you the success rate for a real project.\n\n## Changes in Speed and Writing\n\nAnthropic announced that output generation is more than 30% faster than with Opus 5. This figure concerns the speed at which text is generated. Any reduction in total task time, including search and testing, must be measured separately.\n\nAnthropic says writing has been improved to put the main point first and explain it clearly. It did not provide a separate improvement rate for the quality of Korean documents. Check the following with documents your team uses.\n\n- Whether conclusions and supporting evidence are distinct\n- Whether the requested style and format are followed\n- Whether figures and quotations can be checked against the source\n- Whether the reasons for code changes are explained well enough to review\n\n## What the 85% Reduction in Safety Evaluations Means\n\nThe approximately 85% reduction refers to attempts to bypass isolation boundaries. The comparison is with Opus 5 or Claude Mythos 5.1. It does not mean that the rate of incidents in actual services fell by the same percentage.\n\nA sandbox is an execution environment that limits a program's access. Model safety evaluations and that environment's access restrictions are separate matters. Better evaluation scores do not, by themselves, justify expanding operational permissions.\n\nAnthropic explains that it is still difficult to detect every failure before deployment through evaluation. You can check the scope of the figure in the [safety section of the Claude Opus 5.5 announcement](https://www.anthropic.com/claude-opus-5-5). During operation, check both the scope of file access and execution logs.\n\n## Settings to Check When Switching APIs\n\nExisting API integrations require checking settings beyond the model name. The Claude API model identifier is `claude-opus-5-5`. Other clouds use their respective platform model identifiers.\n\nThe official migration guide states the change to thinking settings as follows.\n\n\u003e Thinking can't be disabled\n\nThe source is Anthropic's [Migrating to Claude Opus 5.5](https://platform.claude.com/docs/en/models/opus-5-5/migration-guide). This means a setting to disable thinking is not supported.\n\n1. Check existing requests for settings that disable thinking. Opus 5.5 rejects those settings.\n2. Specify `effort`, which controls thinking intensity. The default for Opus 5.5 is `medium`.\n3. Check settings that force a specific tool call. Unsupported settings cause request errors.\n4. Process response blocks according to their `type`. The first block is not always answer text.\n5. Measure cost and completion time again using the same tasks. Thinking tokens are also included in output charges.\n\nSupport for AWS, Google Cloud, and Microsoft Azure was also announced at launch. Model identifiers and tool support conditions may differ by platform. Check the migration guide for the model used by your existing integration.\n\n## Common Mistakes\n\nRead price and performance figures according to what they measure. Token prices, task costs, and evaluation scores are different metrics. Keeping the following distinctions in mind helps preserve the conditions for comparison.\n\n| Easy-to-confuse interpretation | Confirmed meaning |\n| --- | --- |\n| Every bill falls by 40% | An estimate for typical tasks with default settings |\n| Caching cuts total costs by 60% | Only the cache read price falls by 60% |\n| Performance rises by 14.1% | The Terminal-Bench score rises by 14.1 percentage points |\n| Every task finishes 30% faster | Output generation speed improves by more than 30% |\n| Actual incidents fall by 85% | Boundary-bypass attempts in a specific evaluation fall by approximately 85% |\n\n## Comparing Cost per Completed Task\n\nWhen choosing a model, compare the cost per task that passes review. Token prices alone do not show the cost of retries and revisions. This is a way to apply the announced figures to actual work.\n\nCalculate this by dividing the total execution cost by the number of tasks that pass review. Include the cost of failed attempts in the total execution cost. Recording human review time as a separate item also helps with comparison.\n\n- Use the same task list and completion criteria\n- Separate standard input, cache, and output costs\n- Record total costs, including failures and retries\n- Record whether each task passes final review\n- Separate execution time from human revision time","content_html":"\u003cp\u003eClaude Opus 5.5 lowers standard token prices by 20%. Its coding and computer use evaluation scores have also improved. The 40% reduction in task costs is a result from Anthropic's testing.\u003c/p\u003e\n\u003cp\u003ePrices and evaluation figures are based on Anthropic's September 22, 2026 announcement.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#claude-opus-55-price-comparison\" class=\"anchor\" id=\"claude-opus-55-price-comparison\"\u003e\u003c/a\u003eClaude Opus 5.5 Price Comparison\u003c/h2\u003e\n\u003cp\u003eStandard API prices are $4 per 1 million input tokens and $20 per 1 million output tokens. Tokens are the units a model uses to process text and other content. Input means content sent to the model, while output means content the model generates.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eBilling item\u003c/th\u003e\n\u003cth\u003eOpus 5\u003c/th\u003e\n\u003cth\u003eOpus 5.5\u003c/th\u003e\n\u003cth\u003ePrice reduction\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Billing item\"\u003e1 million standard input tokens\u003c/td\u003e\n\u003ctd data-label=\"Opus 5\"\u003e$5\u003c/td\u003e\n\u003ctd data-label=\"Opus 5.5\"\u003e$4\u003c/td\u003e\n\u003ctd data-label=\"Price reduction\"\u003e20%\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Billing item\"\u003e1 million output tokens\u003c/td\u003e\n\u003ctd data-label=\"Opus 5\"\u003e$25\u003c/td\u003e\n\u003ctd data-label=\"Opus 5.5\"\u003e$20\u003c/td\u003e\n\u003ctd data-label=\"Price reduction\"\u003e20%\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Billing item\"\u003e1 million cache read tokens\u003c/td\u003e\n\u003ctd data-label=\"Opus 5\"\u003e$0.50\u003c/td\u003e\n\u003ctd data-label=\"Opus 5.5\"\u003e$0.20\u003c/td\u003e\n\u003ctd data-label=\"Price reduction\"\u003e60%\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eAmounts are in US dollars. Caching stores inputs used repeatedly so they can be reused. The cost of creating a cache is calculated separately from the cost of reading it.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#conditions-for-40-savings-on-task-costs\" class=\"anchor\" id=\"conditions-for-40-savings-on-task-costs\"\u003e\u003c/a\u003eConditions for 40% Savings on Task Costs\u003c/h2\u003e\n\u003cp\u003eThe 40% figure is not a discount that applies to every request. It is Anthropic's estimate of savings on typical tasks with default settings. It reflects both lower token prices and fewer tokens used per task.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eUsage pattern\u003c/th\u003e\n\u003cth\u003eWhat affects cost\u003c/th\u003e\n\u003cth\u003eWhat to check\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Usage pattern\"\u003eEntering new content each time\u003c/td\u003e\n\u003ctd data-label=\"What affects cost\"\u003eStandard input and output usage\u003c/td\u003e\n\u003ctd data-label=\"What to check\"\u003eActual token counts for both\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Usage pattern\"\u003eReusing the same document or instructions\u003c/td\u003e\n\u003ctd data-label=\"What affects cost\"\u003eCache creation and read usage\u003c/td\u003e\n\u003ctd data-label=\"What to check\"\u003eWhether the cache was actually reused\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Usage pattern\"\u003eA coding agent performing multiple steps\u003c/td\u003e\n\u003ctd data-label=\"What affects cost\"\u003eRepeated calls and retries\u003c/td\u003e\n\u003ctd data-label=\"What to check\"\u003eTotal cost to complete the task\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Usage pattern\"\u003eMonthly subscription\u003c/td\u003e\n\u003ctd data-label=\"What affects cost\"\u003eSubscription price and usage limits\u003c/td\u003e\n\u003ctd data-label=\"What to check\"\u003eThe distinction between token prices and subscription fees\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eLower API prices alone do not mean lower monthly subscription fees. To determine the savings for your service, compare usage for the same tasks. \u003ca href=\"https://www.anthropic.com/claude/opus\"\u003eAnthropic's Opus overview\u003c/a\u003e also describes the 40% figure as an estimate for token-billed tasks.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#calculation-example\" class=\"anchor\" id=\"calculation-example\"\u003e\u003c/a\u003eCalculation Example\u003c/h2\u003e\n\u003cp\u003eUsing 1 million input tokens and 1 million output tokens costs $24 at standard prices. For this comparison, assume usage is the same for both models. This calculation excludes caching and separate tool costs.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eCalculation item\u003c/th\u003e\n\u003cth\u003eOpus 5\u003c/th\u003e\n\u003cth\u003eOpus 5.5\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Calculation item\"\u003e1 million standard input tokens\u003c/td\u003e\n\u003ctd data-label=\"Opus 5\"\u003e$5\u003c/td\u003e\n\u003ctd data-label=\"Opus 5.5\"\u003e$4\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Calculation item\"\u003e1 million output tokens\u003c/td\u003e\n\u003ctd data-label=\"Opus 5\"\u003e$25\u003c/td\u003e\n\u003ctd data-label=\"Opus 5.5\"\u003e$20\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Calculation item\"\u003eTotal\u003c/td\u003e\n\u003ctd data-label=\"Opus 5\"\u003e$30\u003c/td\u003e\n\u003ctd data-label=\"Opus 5.5\"\u003e$24\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eThe difference is $6, a 20% saving. This calculation does not include any reduction in tokens used per task. The result based on prices alone therefore differs from the announced 40%.\u003c/p\u003e\n\u003cp\u003eThe price difference for reading 1 million cached tokens is $0.30. That amount compares only the read charge, however. The total bill must also account for cache creation and other token costs.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#agentic-coding-and-computer-use-performance\" class=\"anchor\" id=\"agentic-coding-and-computer-use-performance\"\u003e\u003c/a\u003eAgentic Coding and Computer Use Performance\u003c/h2\u003e\n\u003cp\u003eIn Anthropic's announcement, both coding and computer use scores improved. Agentic coding means using tools to carry out multistep development tasks. Computer use is the ability to view a screen and operate an interface.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eEvaluation\u003c/th\u003e\n\u003cth\u003eOpus 5\u003c/th\u003e\n\u003cth\u003eOpus 5.5\u003c/th\u003e\n\u003cth\u003eScore difference\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Evaluation\"\u003eTerminal-Bench 4.0\u003c/td\u003e\n\u003ctd data-label=\"Opus 5\"\u003e52.3%\u003c/td\u003e\n\u003ctd data-label=\"Opus 5.5\"\u003e66.4%\u003c/td\u003e\n\u003ctd data-label=\"Score difference\"\u003e14.1 percentage points\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Evaluation\"\u003eCursorBench 4.0\u003c/td\u003e\n\u003ctd data-label=\"Opus 5\"\u003e46.6%\u003c/td\u003e\n\u003ctd data-label=\"Opus 5.5\"\u003e57.8%\u003c/td\u003e\n\u003ctd data-label=\"Score difference\"\u003e11.2 percentage points\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Evaluation\"\u003eOSWorld 2.1\u003c/td\u003e\n\u003ctd data-label=\"Opus 5\"\u003e74.0%\u003c/td\u003e\n\u003ctd data-label=\"Opus 5.5\"\u003e81.8%\u003c/td\u003e\n\u003ctd data-label=\"Score difference\"\u003e7.8 percentage points\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003cp\u003eThe evaluation name in the official announcement is OSWorld 2.1. Its score is marked \u003ccode\u003epartial\u003c/code\u003e. Keep both the version and the scoring label to avoid confusing it with other results.\u003c/p\u003e\n\u003cp\u003eThe Opus 5.5 result for Terminal-Bench 4.0 uses the \u003ccode\u003exhigh\u003c/code\u003e setting. Its evaluation conditions are not the same as those used for the announcement's cost estimate with default settings. A higher score alone cannot tell you the success rate for a real project.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#changes-in-speed-and-writing\" class=\"anchor\" id=\"changes-in-speed-and-writing\"\u003e\u003c/a\u003eChanges in Speed and Writing\u003c/h2\u003e\n\u003cp\u003eAnthropic announced that output generation is more than 30% faster than with Opus 5. This figure concerns the speed at which text is generated. Any reduction in total task time, including search and testing, must be measured separately.\u003c/p\u003e\n\u003cp\u003eAnthropic says writing has been improved to put the main point first and explain it clearly. It did not provide a separate improvement rate for the quality of Korean documents. Check the following with documents your team uses.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eWhether conclusions and supporting evidence are distinct\u003c/li\u003e\n\u003cli\u003eWhether the requested style and format are followed\u003c/li\u003e\n\u003cli\u003eWhether figures and quotations can be checked against the source\u003c/li\u003e\n\u003cli\u003eWhether the reasons for code changes are explained well enough to review\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2\u003e\n\u003ca href=\"#what-the-85-reduction-in-safety-evaluations-means\" class=\"anchor\" id=\"what-the-85-reduction-in-safety-evaluations-means\"\u003e\u003c/a\u003eWhat the 85% Reduction in Safety Evaluations Means\u003c/h2\u003e\n\u003cp\u003eThe approximately 85% reduction refers to attempts to bypass isolation boundaries. The comparison is with Opus 5 or Claude Mythos 5.1. It does not mean that the rate of incidents in actual services fell by the same percentage.\u003c/p\u003e\n\u003cp\u003eA sandbox is an execution environment that limits a program's access. Model safety evaluations and that environment's access restrictions are separate matters. Better evaluation scores do not, by themselves, justify expanding operational permissions.\u003c/p\u003e\n\u003cp\u003eAnthropic explains that it is still difficult to detect every failure before deployment through evaluation. You can check the scope of the figure in the \u003ca href=\"https://www.anthropic.com/claude-opus-5-5\"\u003esafety section of the Claude Opus 5.5 announcement\u003c/a\u003e. During operation, check both the scope of file access and execution logs.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#settings-to-check-when-switching-apis\" class=\"anchor\" id=\"settings-to-check-when-switching-apis\"\u003e\u003c/a\u003eSettings to Check When Switching APIs\u003c/h2\u003e\n\u003cp\u003eExisting API integrations require checking settings beyond the model name. The Claude API model identifier is \u003ccode\u003eclaude-opus-5-5\u003c/code\u003e. Other clouds use their respective platform model identifiers.\u003c/p\u003e\n\u003cp\u003eThe official migration guide states the change to thinking settings as follows.\u003c/p\u003e\n\u003cblockquote\u003e\n\u003cp\u003eThinking can't be disabled\u003c/p\u003e\n\u003c/blockquote\u003e\n\u003cp\u003eThe source is Anthropic's \u003ca href=\"https://platform.claude.com/docs/en/models/opus-5-5/migration-guide\"\u003eMigrating to Claude Opus 5.5\u003c/a\u003e. This means a setting to disable thinking is not supported.\u003c/p\u003e\n\u003col\u003e\n\u003cli\u003eCheck existing requests for settings that disable thinking. Opus 5.5 rejects those settings.\u003c/li\u003e\n\u003cli\u003eSpecify \u003ccode\u003eeffort\u003c/code\u003e, which controls thinking intensity. The default for Opus 5.5 is \u003ccode\u003emedium\u003c/code\u003e.\u003c/li\u003e\n\u003cli\u003eCheck settings that force a specific tool call. Unsupported settings cause request errors.\u003c/li\u003e\n\u003cli\u003eProcess response blocks according to their \u003ccode\u003etype\u003c/code\u003e. The first block is not always answer text.\u003c/li\u003e\n\u003cli\u003eMeasure cost and completion time again using the same tasks. Thinking tokens are also included in output charges.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eSupport for AWS, Google Cloud, and Microsoft Azure was also announced at launch. Model identifiers and tool support conditions may differ by platform. Check the migration guide for the model used by your existing integration.\u003c/p\u003e\n\u003ch2\u003e\n\u003ca href=\"#common-mistakes\" class=\"anchor\" id=\"common-mistakes\"\u003e\u003c/a\u003eCommon Mistakes\u003c/h2\u003e\n\u003cp\u003eRead price and performance figures according to what they measure. Token prices, task costs, and evaluation scores are different metrics. Keeping the following distinctions in mind helps preserve the conditions for comparison.\u003c/p\u003e\n\u003cdiv class=\"overflow-x-auto\"\u003e\u003ctable\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth\u003eEasy-to-confuse interpretation\u003c/th\u003e\n\u003cth\u003eConfirmed meaning\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Easy-to-confuse interpretation\"\u003eEvery bill falls by 40%\u003c/td\u003e\n\u003ctd data-label=\"Confirmed meaning\"\u003eAn estimate for typical tasks with default settings\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Easy-to-confuse interpretation\"\u003eCaching cuts total costs by 60%\u003c/td\u003e\n\u003ctd data-label=\"Confirmed meaning\"\u003eOnly the cache read price falls by 60%\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Easy-to-confuse interpretation\"\u003ePerformance rises by 14.1%\u003c/td\u003e\n\u003ctd data-label=\"Confirmed meaning\"\u003eThe Terminal-Bench score rises by 14.1 percentage points\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Easy-to-confuse interpretation\"\u003eEvery task finishes 30% faster\u003c/td\u003e\n\u003ctd data-label=\"Confirmed meaning\"\u003eOutput generation speed improves by more than 30%\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd data-label=\"Easy-to-confuse interpretation\"\u003eActual incidents fall by 85%\u003c/td\u003e\n\u003ctd data-label=\"Confirmed meaning\"\u003eBoundary-bypass attempts in a specific evaluation fall by approximately 85%\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\u003c/div\u003e\n\u003ch2\u003e\n\u003ca href=\"#comparing-cost-per-completed-task\" class=\"anchor\" id=\"comparing-cost-per-completed-task\"\u003e\u003c/a\u003eComparing Cost per Completed Task\u003c/h2\u003e\n\u003cp\u003eWhen choosing a model, compare the cost per task that passes review. Token prices alone do not show the cost of retries and revisions. This is a way to apply the announced figures to actual work.\u003c/p\u003e\n\u003cp\u003eCalculate this by dividing the total execution cost by the number of tasks that pass review. Include the cost of failed attempts in the total execution cost. Recording human review time as a separate item also helps with comparison.\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eUse the same task list and completion criteria\u003c/li\u003e\n\u003cli\u003eSeparate standard input, cache, and output costs\u003c/li\u003e\n\u003cli\u003eRecord total costs, including failures and retries\u003c/li\u003e\n\u003cli\u003eRecord whether each task passes final review\u003c/li\u003e\n\u003cli\u003eSeparate execution time from human revision time\u003c/li\u003e\n\u003c/ul\u003e\n","tags":["AI Agents","Anthropic","AI Development","Claude","Coding Agent"],"faqs":[{"question":"When was Claude Opus 5.5 released?","answer":"Anthropic released Claude Opus 5.5 on September 22, 2026."},{"question":"What is the standard API price for Claude Opus 5.5?","answer":"At release, the price was $4 per million input tokens and $20 per million output tokens. Each rate is 20% lower than for Opus 5."},{"question":"Are costs always reduced by 40%?","answer":"The 40% figure is Anthropic's estimate for typical tasks. Actual savings vary depending on token usage and cache use."},{"question":"How much did the cache read price drop?","answer":"It dropped from $0.50 to $0.20 per million tokens. That is a 60% reduction in the read rate. Cache creation costs need to be checked separately."},{"question":"What is the exact name of the computer use evaluation?","answer":"The evaluation is called OSWorld 2.1 in the official announcement. Opus 5.5 scored 81.8%, with a partial label."},{"question":"Does the 85% safety figure mean an 85% reduction in incidents?","answer":"It means that attempts to bypass isolation boundaries decreased by about 85% in a specific evaluation. It does not mean a reduction in incidents in actual services."},{"question":"Can I just change the model name in my existing API requests?","answer":"Additional changes may be needed depending on your existing request settings. Check the migration documentation for reasoning settings, forced tool calls, and response block handling."},{"question":"Is it available on AWS and Google Cloud too?","answer":"The launch announcement included support for AWS, Google Cloud, and Microsoft Azure. Check the model identifiers and conditions for feature support on each platform."},{"question":"Do faster output generation and shorter task completion times mean the same thing?","answer":"They are different metrics. If a task includes tool execution and testing in addition to output generation, the total completion time may differ."}],"sources":[{"url":"https://www.anthropic.com/claude-opus-5-5","title":"Anthropic, Introducing Claude Opus 5.5, September 22, 2026","type":"source"},{"url":"https://www.anthropic.com/claude/opus","title":"Anthropic, Claude Opus model and pricing information","type":"source"},{"url":"https://platform.claude.com/docs/en/models/opus-5-5/migration-guide","title":"Claude Platform Docs, Migrating to Claude Opus 5.5","type":"source"},{"url":"https://platform.claude.com/docs/en/build-with-claude/prompt-caching","title":"Claude Platform Docs, Prompt caching","type":"source"}],"images":[{"id":1595,"url":"https://injoys.com/rails/active_storage/blobs/proxy/eyJfcmFpbHMiOnsiZGF0YSI6MjM5MjIsInB1ciI6ImJsb2JfaWQifX0=--9ab6a9e985c76b07acd43a43425994585ce87958/ai-15b093c3.webp","is_representative":true,"generation_method":"ai_photo","license":"ai_generated","mime_type":"image/webp","width":1536,"height":1024,"translations":{"ko":{"alt":"서버실에서 태블릿을 들고 열린 서버 랙을 살펴보는 엔지니어","caption":"실제 운영 비용은 토큰 사용량과 캐시 활용에 따라 달라집니다.","description":null},"en":{"alt":"Engineer holding a tablet and inspecting an open server rack in a server hall","caption":"Actual operating costs depend on token usage and cache use.","description":null},"ja":{"alt":"サーバールームでタブレットを持ち、開いたサーバーラックを確認するエンジニア","caption":"実際の運用コストはトークン使用量とキャッシュの活用次第で変わります。","description":null},"es":{"alt":"Ingeniero con una tableta inspecciona un rack de servidores abierto en una sala de servidores","caption":"El coste operativo real depende del uso de tokens y de la caché.","description":null},"id":{"alt":"Teknisi memegang tablet sambil memeriksa rak server terbuka di ruang server","caption":"Biaya operasional sebenarnya bergantung pada penggunaan token dan cache.","description":null},"pt":{"alt":"Engenheiro com um tablet inspeciona um rack de servidores aberto numa sala de servidores","caption":"O custo operacional real depende do uso de tokens e do aproveitamento do cache.","description":null},"zh-hant":{"alt":"工程師在伺服器機房手持平板，檢查開啟的伺服器機櫃","caption":"實際營運成本會隨權杖用量與快取使用情況而變動。","description":null},"de":{"alt":"Ingenieur hält ein Tablet und prüft ein offenes Serverrack in einem Serverraum","caption":"Die tatsächlichen Betriebskosten hängen von der Token-Nutzung und dem Cache-Einsatz ab.","description":null}}},{"id":1596,"url":"https://injoys.com/rails/active_storage/blobs/proxy/eyJfcmFpbHMiOnsiZGF0YSI6MjM5MjgsInB1ciI6ImJsb2JfaWQifX0=--677963acf29b560200099cc3de29dd3eebb74a30/ai-8bc52f98.webp","is_representative":false,"generation_method":"ai_semi","license":"ai_generated","mime_type":"image/webp","width":1536,"height":1024,"translations":{"ko":{"alt":"개발자가 테스트실에서 연결된 스마트폰 여러 대 앞에 앉아 태블릿을 확인한다.","caption":"기존 API를 전환할 때는 모델명과 함께 추론·도구 호출 설정을 점검해야 합니다.","description":null},"en":{"alt":"A developer checks a tablet at a test bench lined with connected phones.","caption":"When switching an existing API, check reasoning and tool-call settings as well as the model name.","description":null},"ja":{"alt":"開発者がテスト室で、接続された複数のスマートフォンを前にタブレットを確認している。","caption":"既存のAPIを切り替える際は、モデル名に加えて推論とツール呼び出しの設定も確認しましょう。","description":null},"es":{"alt":"Un desarrollador revisa una tableta ante una fila de teléfonos conectados en un laboratorio de pruebas.","caption":"Al cambiar una API existente, hay que revisar el nombre del modelo y los ajustes de razonamiento y llamadas a herramientas.","description":null},"id":{"alt":"Seorang pengembang memeriksa tablet di depan deretan ponsel yang terhubung di ruang pengujian.","caption":"Saat beralih dari API yang ada, periksa nama model serta pengaturan penalaran dan pemanggilan alat.","description":null},"pt":{"alt":"Um desenvolvedor verifica um tablet diante de uma fileira de celulares conectados em um laboratório de testes.","caption":"Ao migrar uma API existente, verifique o nome do modelo e as configurações de raciocínio e de chamadas de ferramentas.","description":null},"zh-hant":{"alt":"開發者在測試室裡，坐在一排已連接的手機前查看平板電腦。","caption":"轉換現有 API 時，除了模型名稱，也要檢查推理與工具呼叫設定。","description":null},"de":{"alt":"Ein Entwickler prüft ein Tablet vor einer Reihe angeschlossener Smartphones in einem Testlabor.","caption":"Bei der Umstellung einer bestehenden API sollten neben dem Modellnamen auch die Einstellungen für Schlussfolgerungen und Werkzeugaufrufe geprüft werden.","description":null}}}],"published_at":"2026-10-02T10:07:05+09:00","updated_at":"2026-10-02T10:07:05+09:00","license":"cc_by","translation_status":"reviewed","available_locales":["ko","en","ja","es"],"data_locales":["ko","en","ja","es","id","pt","zh-hant","de"],"url":"https://injoys.com/en/articles/claude-opus-5-5-pricing-performance-comparison"}