Claude Opus 5.5 beats Fable 5.1 on every row of Anthropic’s launch table and costs 40% as much per input and output token. I replayed the Opus 5 production day I measured on September 4 at Opus 5.5’s list prices: the day comes to $220 instead of $359, and Fable 5.1 goes from 1.4 times the cost per turn to 2.3 times. The benchmark numbers below are Anthropic’s, not mine. The cost numbers are real transcripts priced at new rates, not a new measurement, and I mark where that difference matters.
If you want the background, the Fable 5.1 vs Opus 5 post has the day this replay is built on, and the Claude Code complete guide covers the setup it ran in.
TL;DR
Opus 5.5 shipped on 09/22/2026 as claude-opus-5-5 and became the default Opus model in Claude Code the same day (version 2.1.280). It lists at $4 input, $20 output and $0.20 cache read per million tokens. Fable 5.1 lists at $10, $50 and $0.25. Fable 5.1 is now 2.5x Opus 5.5 on every line except cache reads, where it is 1.25x. Opus 5 paid twice Fable’s cache-read rate, and Fable’s cheaper cache reads are what kept the old gap small. My measured day, replayed: $0.074 per turn on Opus 5.5 against $0.168 on Fable 5.1.
What Anthropic published on September 22
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 66.4% | 55.8% | 52.3% |
| FrontierCode v1.1, main (agentic coding) | 54.4% | 50.3% | 48.0% |
| CursorBench 4.0 (agentic coding) | 57.8% | 51.8% | 46.6% |
| GDPval-AA v2.1 (knowledge work, Elo) | 1,846 | 1,735 | 1,708 |
| AutomationBench (business workflows) | 40.0% | 31.4% | 26.9% |
| Humanity’s Last Exam, with tools | 67.7% | 65.6% | 63.6% |
| Terminal-Bench-Science 0.1 (research) | 58.7% | 52.6% | 29.0% |
| OSWorld 2.0, partial (computer use) | 81.8% | 80.7% | 74.0% |
| Chartography, with tools (charts) | 89.0% | 88.4% | 83.4% |
Source: Anthropic’s Opus 5.5 launch page, read 09/26/2026. Opus 5.5 ran at max effort except on Terminal-Bench 4.0 (xhigh); the page does not state the effort used for Fable 5.1 or Opus 5.
Opus 5.5 leads on all nine. The gaps are not even. Agentic coding, business workflows and science tasks carry the separation: 10.6 points on Terminal-Bench 4.0, 8.6 on AutomationBench, 6.1 on Terminal-Bench-Science, 6.0 on CursorBench 4.0. Computer use and chart reading are close to ties, 1.1 and 0.6 points.
Three numbers do not match between Anthropic’s two launch pages. On OSWorld 2.0 (partial), the Fable 5.1 page lists Fable 5.1 at 77.9% and Opus 5 at 75.4%, while the Opus 5.5 page lists them at 80.7% and 74.0%. On Humanity’s Last Exam with tools, Fable 5.1 is 65.0% on its own page and 65.6% on the Opus 5.5 page. The Fable 5.1 page says its OSWorld scores use the benchmark authors’ August 2026 task release and are not directly comparable to earlier published results; the Opus 5.5 page does not say which release it used. I report the Opus 5.5 page because it puts all three models in one table, and I lean on neither OSWorld figure.
These are vendor results. They say which model Anthropic’s test setup gets more out of on Anthropic’s tasks. They do not say which one finishes your tasks in fewer turns.
The price list: 2.5x on input and output, 1.25x on cache reads
| Per million tokens | Opus 5.5 | Fable 5.1 | Opus 5 |
|---|---|---|---|
| Input | $4 | $10 | $5 |
| Output | $20 | $50 | $25 |
| Cache write, 5 minutes | $5 | $12.50 | $6.25 |
| Cache write, 1 hour | $8 | $20 | $10 |
| Cache read | $0.20 | $0.25 | $0.50 |
| Batch input / output | $2 / $10 | $5 / $25 | $2.50 / $12.50 |
Source: Claude API pricing page, read 09/26/2026. Opus 5.5 fast mode lists at $8 input and $40 output.
The multiplier to notice is the cache read. Opus 5.5 reads its cache at 5% of its input price, Fable 5.1 at 2.5%, Opus 5 at the usual 10%. Opus 5 paid twice Fable 5.1’s rate to read cache, and cache reads were the largest line of its day, so Fable’s 2x sticker shrank to 1.4x per turn on agent work. Opus 5.5 cut its cache-read price by 60%, and most of that offset is gone. The general caching math for every current model, including when a cache write pays for itself, is in Claude prompt caching cost.
Replaying my measured Opus 5 day at Opus 5.5 prices
On 09/04/2026 my Claude Code setup ran 2,979 Opus 5 turns and 4,501 Fable 5.1 turns in Claude Code, and I priced every transcript at list. The Opus 5 turns averaged 113,957 cache-read tokens and 980 output tokens each, and the day’s cache writes came to $117 at $6.25 per million, about 18.7 million tokens. Same tokens, Opus 5.5 prices:
| Cost line | Opus 5, measured day at list | Same tokens at Opus 5.5 list |
|---|---|---|
| Cache reads | $170 | $68 |
| Cache writes (5-minute tier) | $117 | $94 |
| Output | $73 | $58 |
| Total | $359 | $220 |
| Per turn | $0.121 | $0.074 |
Claude Code transcripts from 09/04/2026, priced at list on 09/05/2026 (Opus 5) and 09/26/2026 (Opus 5.5). Rows are rounded to the dollar. Not a bill.
Cache reads account for $102 of the $139 difference. Fable 5.1 cost $0.168 per turn on the same day. Against Opus 5 that was 1.4x. Against Opus 5.5 it is 2.3x. Per turn, most of that gap is now cache writes ($0.082 against $0.031) and output ($0.056 against $0.020), both priced at the 2.5x sticker ratio. If Claude Code writes to the one-hour cache tier instead, the day comes to $430 on Opus 5, $276 on Opus 5.5 and $980 on Fable 5.1, and the ratio is 2.35x. The conclusion does not depend on which tier you assume.
Three boundaries on these numbers.
- Same tokens is an assumption. The replay assumes Opus 5.5 spends what Opus 5 spent per turn. Anthropic says Opus 5.5 generates output more than 30% faster than Opus 5 and costs 40% less at default settings, and it defaults to medium effort where Opus 5 ran at high. So on my workload the real figure is more likely below $220 than above it. That is an inference, not a measurement.
- It is a replay. The measurement is a transcript run on Opus 5.5 turns, and I have not run it yet.
- The replay rebuilds the published day. Recomputed from the per-turn figures, the Opus 5 lines come back at $169.74 for cache reads and $72.99 for output, and the one-hour variant at $430 against the $429 I published. That is the check that the replay starts from the right day.
Anthropic’s launch post says Opus 5.5 costs 40% less than Opus 5 at default settings on typical workloads. My replay says 39%. The two numbers measure different things: theirs includes medium effort and any change in tokens spent, mine holds my tokens fixed and changes only the prices. They do not confirm each other.
What changes when you switch in Claude Code
Claude Code 2.1.280 made Opus 5.5 the default Opus model, with a 1M-token context window. The default model alias now resolves to Opus 5.5 on most plans. A few things carry over badly from Opus 5.
- Default effort is medium. Opus 5 defaulted to high. Anthropic’s effort guide for Opus 5.5 says to run an effort sweep on your own evaluations instead of carrying settings over, and to give it a large
max_tokensat the higher levels. - Thinking is always on. A request that sets thinking to disabled returns an error at every effort level.
- Your instruction files were written for an older model. Claude Code 2.1.283 added
/doctor prompt-audit(also/checkup prompt-audit), which audits CLAUDE.md files, skills, agents and commands for prompting patterns written for older models. When I moved from Opus 4.8 to Opus 5, guardrails tuned for the old model pointed the wrong way. Run the audit before you judge the new one. - Fast mode costs double. $8 input and $40 output per million tokens, for up to 2.5x the speed.
Where Fable 5.1 still earns 2.5x
On Anthropic’s own table: nowhere. What the table does not settle is narrower.
- The benchmarks it leaves out. My routing notes kept Fable in the seat for one kind of task, a single hard, novel code diff, because Fable 5 held its edge over Opus 5 on SWE-bench Pro and DeepSWE. The Opus 5.5 table reports neither.
- Tool errors. On my September 4 day, Fable 5.1 made 8.3 tool errors per 100 tool calls against Opus 5’s 11.0. I have no Opus 5.5 figure yet.
- Plan billing. On a Max plan you do not pay per token, and my Fable-heavy September 4 day (2,443 Fable 5.1 turns and 1,120 Opus 5 turns before midnight) used 19% of the weekly window, plan-wide. If you are on a plan, the question is quality on your own tasks, not price.
So Fable 5.1 keeps a reservation for work where a wrong answer costs more than 2.3x the tokens and Anthropic’s table does not already answer it.
My routing now
My Claude Code setup has run Opus 5.5 at medium effort as its main loop since 09/23/2026. The tiebreaker I use is intelligence, then taste, then cost. Last time Opus 5 and Fable tied on the first two, so cost decided. This time there is no tie to break: Opus 5.5 leads Anthropic’s table and costs less.
The next measurement is a seven-day transcript run on Opus 5.5, priced the same way. It replaces the replay in this post when it exists.
Price your own workload
The replay is my token mix, and yours is different. An agent that re-reads a large context on every turn behaves like my day. A workload with short prompts and long answers sits closer to the 2.5x sticker gap. The Claude token counter and cost calculator prices a brief of your own on Opus 5.5, Fable 5.1 and the rest, with the cache multipliers above.
For the version history of this comparison, the Fable 5 vs Opus 4.8 post has the first generation, and how to use Claude Opus 5 has the failure-replay method I use before trusting any new model.