Anthropic’s current Claude lineup makes a simple price comparison misleading. This review puts Claude Opus 5.5, Claude Opus 5, and Claude Fable 5.1 through the same five prompts and reports the task result, local elapsed time, route-reported token usage, and the limits of each observation.
The practical difference is visible in the task cards. Opus 5.5 was the quickest and cheapest on the measured pack, Fable 5.1 produced the fewest output tokens and the cleanest Chinese-format response, and Opus 5 took the most time, while Opus 5.5 used the most output tokens. The coding card adds an important qualification: shorter output was not automatically better output.
We copied the five prompts from the earlier Claude Opus 5.5 review and ran every prompt once for each model. The request settings were identical: an 8,192-token ceiling, high effort, one user message, and no tools. The route returned HTTP 200 and end_turn for all 15 requests.
간단한 답변
- Fastest in this five-task pack: Claude Opus 5.5 used 47.725 seconds of local end-to-end time across five sequential requests. Claude Fable 5.1 used 53.290 seconds, and Claude Opus 5 used 61.470 seconds.
- Lowest list-price estimate: Opus 5.5 came to about $0.0818 for the five requests at official token rates, followed by Opus 5 at $0.0962 and Fable 5.1 at $0.1415. These are arithmetic estimates, not platform charges.
- Quality result: all three returned grounded five-bullet extraction, parseable JSON, and the correct math answer. The coding outputs differed in edge-case handling, while Fable 5.1 followed the two-paragraph Chinese-format instruction most cleanly.
- API caveat: Opus 5.5 is not a drop-in rename for every Opus 5 integration. Its migration guide says thinking cannot be disabled and some older tool/thinking patterns need changes.
- Do not overread the ranking: this is one matched run per task on one route. It measures concrete behavior, elapsed time, and reported usage, not a universal model score.
What the three models cost and expose
The three APIs share a 1M-token context window and a 128K maximum output in Anthropic’s current model documentation. Their billing units and status differ: Opus 5.5 is the latest Opus route, Opus 5 is marked legacy, and Fable 5.1 carries the highest standard input/output rate.
| 모델 | 공식 지위 | API 모델 ID | 입력 / 출력 | 컨텍스트 / 최대 출력 | Default effort | 캐시 읽기 |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 최신 | 클로드-작품 5-5 | $4 / $20 per MTok | 1M / 128K | Medium | $0.20 / MTok |
| Claude Opus 5 | Active, legacy | 클로드-오푸스-5 | MTok당 $5 / $25 | 1M / 128K | 높음 | $0.50 / MTok |
| Claude Fable 5.1 | 최신 | 클로드-파블-5-1 | MTok당 $10 / $50 | 1M / 128K | 높음 | $0.25 / MTok |
For a simple non-cached example, 100,000 input tokens plus 20,000 output tokens costs $0.80 on Opus 5.5, $1.00 on Opus 5, and $2.00 on Fable 5.1 at the listed rates. Fable 5.1 has a lower cache-read rate than Opus 5, so cache-heavy agent workloads need a separate calculation.
What the official benchmark table shows
Anthropic의 Claude Opus 5.5 announcement places all three models in the same provider comparison table. Opus 5.5 leads the listed rows, but these are provider-reported results with benchmark-specific harnesses and effort settings. The announcement itself warns that narrow benchmark margins are a less reliable guide to real-world differences at this capability level.
| Provider benchmark | Opus 5.5 | Fable 5.1 | 오푸스 5 | What it measures |
|---|---|---|---|---|
| 터미널-벤치 4.0 | 66.4% | 55.8% | 52.3% | Agentic command-line coding |
| FrontierCode v1.1 | 54.4% | 50.3% | 48.0% | Whether agent code changes would be merged |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | Ambiguous multi-file coding tasks |
| GDPval-AA v2.1 | 1846 | 1735 | 1708 | Professional knowledge work across 44 occupations |
| AutomationBench | 40.0% | 31.4% | 26.9% | End-to-end business workflows |
| 인류의 마지막 시험 | 67.7% with tools | 65.6% with tools | 63.6% with tools | Multidisciplinary reasoning |
| OSWorld 2.0 | 81.8% partial | 80.7% partial | 74.0% partial | Computer-use tasks |
The table is useful for locating the provider’s claimed performance frontier. It is not a replacement for the matched cards below: our test uses short single-turn prompts, high effort, no tools, and a third-party controlled route.
Same prompts, same request settings
We copied the five prompts from the earlier Opus 5.5 test pack and sent each one once to each model. The client rotated model order by task, used UTF-8 JSON, and measured from request start until the complete response body was read. All 15 requests returned HTTP 200 and end_turn. For the broader rationale and repeatable workflow, see How GlobalGPT tests AI models.
Time and token comparison
| 경로 | 5-run elapsed | Mean / task | 입력 토큰 | 출력 토큰 | Route-reported thinking | List-price estimate* |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 47.725s | 9.545s | 700 | 3,952 | 2,214 | $0.0818 |
| Claude Opus 5 | 61.470s | 12.294s | 690 | 3,709 | 1,859 | $0.0962 |
| Claude Fable 5.1 | 53.290s | 10.658s | 700 | 2,690 | 1,122 | $0.1415 |
*Estimate from route-reported input/output tokens and Anthropic’s listed standard API rates. It excludes cache charges, platform credits, taxes, and markup. Elapsed time is local end-to-end wall time after the request was sent, not provider latency or tokens per second. “Thinking” is a route-reported usage field, not independently measured hidden reasoning.
The aggregate points in different directions depending on the metric. Opus 5.5 returned the fastest total and the lowest estimated cost, Fable 5.1 returned the fewest output tokens, and Opus 5 had the slowest total in this pack. A shorter answer is not automatically a better answer: the coding card shows why output content and edge-case handling still need to be read.
Five matched task cards
Can the models repair a mixed-type deduplication function?
작업 설정: The prompt required a corrected function plus two concise tests, case-insensitive string comparison, preserved first spelling, and no crash on non-string values.
| 모델 | Time / usage | 관측된 결과 |
|---|---|---|
| Claude Opus 5.5 | 15.844s local · 145 in / 1478 out · 700 thinking · end_turn | Uses casefold plus a list fallback; its docstring says identity while the code uses equality. |
| Claude Opus 5 | 16.845s local · 143 in / 1109 out · 312 thinking · end_turn | Adds type-tagged keys and an equality fallback, and calls out the invalid one-line Python syntax. |
| Claude Fable 5.1 | 12.778s local · 145 in / 783 out · 0 thinking · end_turn | Uses repr(item) for unhashable values; that passes the shown examples but can collide across types. |
Comparison result: The edge case is a task-level observation, not a claim about all coding workloads.
Can the models preserve five supplied facts without adding claims?
작업 설정: The input was a short note about a two-week test of three writing assistants. It is a short extraction task, not a large-context benchmark.
| 모델 | Time / usage | 관측된 결과 |
|---|---|---|
| Claude Opus 5.5 | 4.374s local · 210 in / 312 out · 101 thinking · end_turn | Returned five numbered bullets and stayed inside the supplied facts; labels and order differed slightly. |
| Claude Opus 5 | 4.661s local · 208 in / 265 out · 27 thinking · end_turn | Returned five numbered bullets and stayed inside the supplied facts; labels and order differed slightly. |
| Claude Fable 5.1 | 6.305s local · 210 in / 223 out · 0 thinking · end_turn | Returned five numbered bullets and stayed inside the supplied facts; labels and order differed slightly. |
Comparison result: The source passage was short; this checks grounded extraction and formatting, not million-token context handling.
Can the models return the exact requested structure?
작업 설정: The prompt required valid JSON only, five exact keys, two pros, two cons, and short strings. The fictional review content is not evidence about any model.
| 모델 | Time / usage | 관측된 결과 |
|---|---|---|
| Claude Opus 5.5 | 8.276s local · 128 in / 622 out · 390 thinking · end_turn | Returned parseable JSON with the requested keys and two pros/two cons; generated review claims are test text, not product facts. |
| Claude Opus 5 | 6.344s local · 126 in / 238 out · 0 thinking · end_turn | Returned parseable JSON with the requested keys and two pros/two cons; generated review claims are test text, not product facts. |
| Claude Fable 5.1 | 6.626s local · 128 in / 203 out · 0 thinking · end_turn | Returned parseable JSON with the requested keys and two pros/two cons; generated review claims are test text, not product facts. |
Comparison result: The JSON strings describe a fictional review setup and must not be published as product facts.
Can the models carry the rounded batch sequence through to the final answer?
작업 설정: The sequence was 12, then 25% more with floor rounding, across four batches.
| 모델 | Time / usage | 관측된 결과 |
|---|---|---|
| Claude Opus 5.5 | 3.830s local · 89 in / 257 out · 74 thinking · end_turn | Computed 12, 15, 18, 22 and a total of 67; all three ended with the requested standalone 67. |
| Claude Opus 5 | 5.079s local · 87 in / 277 out · 90 thinking · end_turn | Computed 12, 15, 18, 22 and a total of 67; all three ended with the requested standalone 67. |
| Claude Fable 5.1 | 5.347s local · 89 in / 127 out · 0 thinking · end_turn | Computed 12, 15, 18, 22 and a total of 67; all three ended with the requested standalone 67. |
Comparison result: All three passed this arithmetic prompt; one prompt cannot estimate general reasoning reliability.
Can the route keep the requested two-paragraph structure?
작업 설정: The prompt requested two Chinese paragraphs of 80–120 Chinese characters each: a value answer followed by the test plan.
| 모델 | Time / usage | 관측된 결과 |
|---|---|---|
| Claude Opus 5.5 | 15.401s local · 128 in / 1283 out · 949 thinking · end_turn | Returned the two requested paragraphs, then added Markdown framing and an English note. |
| Claude Opus 5 | 28.541s local · 126 in / 1820 out · 1430 thinking · end_turn | Returned the two requested paragraphs, then added an English note after a separator. |
| Claude Fable 5.1 | 22.234s local · 128 in / 1354 out · 1122 thinking · end_turn | Returned two clean Chinese paragraphs with no extra note or Markdown framing. |
Comparison result: This records instruction following through the tested route; it is not a Chinese-language quality benchmark.
API 및 마이그레이션 시 유의사항
Opus 5.5 is not a drop-in rename for every Opus 5 integration. Anthropic’s migration guide says thinking cannot be disabled on Opus 5.5; requests that try to disable or manually enable the old thinking mode return a 400 error. The migration also changes forced tool use, ties thinking blocks to the model and conversation, and replaces the older computer-use tool on the Claude API and Google Cloud path.
The shared test used output_config.effort: high for all three models so the prompt results were easier to compare. That means these runs do not reproduce each model’s default effort setting. Fable 5.1 is documented with adaptive thinking always on and high default effort; Opus 5.5 is documented with adaptive thinking always on and medium default effort; Opus 5 uses adaptive thinking with high default effort.
What this test actually supports
Task-level verdict
Opus 5.5: fastest and lowest estimated list-price cost in this five-task pack, while also producing the most output tokens. Fable 5.1: shortest output overall and cleanest compliance with the Chinese two-paragraph constraint, but its higher token price made the five-run estimate highest and its coding fallback had a concrete cross-type edge case. 오푸스 5: correct on the extraction, JSON, and math tasks, but slowest in this route sample and marked legacy in the official docs.
That is enough to justify a practical preference for Opus 5.5 on this measured route, not enough to claim that it wins every workload. Repeat the test with tools, longer inputs, and multiple runs before treating the result as a production benchmark.
If you want to reproduce the comparison without opening three separate accounts, try the model comparison workflow in GlobalGPT. The direct model route and the upstream provider mapping should still be verified separately for any production decision.
자주 묻는 질문
Which model was fastest in the matched test?
Claude Opus 5.5 had the lowest total local elapsed time: 47.725 seconds across five sequential requests. This is a route observation, not provider-side latency.
Which model used the fewest output tokens?
Claude Fable 5.1 used 2,690 route-reported output tokens across the five tasks, compared with 3,709 for Opus 5 and 3,952 for Opus 5.5.
Which model was cheapest in the test?
Using official standard API rates and the returned input/output token counts, Opus 5.5 produced the lowest five-run estimate at $0.0818. This excludes cache, credits, taxes, and platform markup.
Did all three models pass the same tasks?
All three returned grounded five-bullet extraction, parseable JSON with the requested keys, and the correct total of 67. The coding and Chinese-format cards show meaningful output differences.
Was this a long-context benchmark?
No. The reused extraction prompt contains a short passage. It tests grounding and exact formatting, so it should not be presented as evidence of million-token context performance.
Is Claude Opus 5 still current?
Anthropic’s model docs label Opus 5 as active but legacy and recommend considering Opus 5.5 for improved performance. Existing API integrations should review the migration guide before switching.
Official specifications: Opus 5.5 docs, Opus 5 docs, 및 Fable 5.1 docs. Provider benchmark values are attributed to Anthropic; hands-on values come from the matched test pack described above.
