I Tested Opus 5.5 on 24 Coding Tasks: Fable-Level Quality in Half the Time and Cost

It was a crowded week for frontier models. Claude Opus 5.5 arrived while OpenAI talked about GPT-6 Sol and Luna, and Mimo 2.6 was mentioned but set aside for being too slow to test fully. The core question in the video was whether 5.5 is a small increment like 5.1 or a real jump. Internal chatter had floated 5.2, and Fable had shipped as 5.1 without much impact. On the author's private coding leaderboard Opus 5 had trailed Astro medium and even Solve high by fractions — not far, but not first.

Method: One Bash Loop, Four Projects

The method was intentionally plain. A Bash script called the Claude Code CLI for 24 coding prompts. Twenty tasks were scored up to 20 points each for edge cases and imperfect paths across different projects; two were PHP, one Dart Flutter, one Go. Separately, a React and TypeScript frontend and a PHP Laravel backend were judged by GPT 5.6 up to 20 points for code quality. The first run finished all checks in about two minutes, the second even faster. Single-run variance is high, so the author may move to a three-run average in the future, but for now cost kept it to one reported run.

Speed and cost were the immediate surprise. With Opus 5, both medium and high averaged over a dollar per prompt even at the 20-dollar Anthropic plan; with Opus 5.5 the entire 24 prompts stayed under a dollar. The author usually spent three to four five-hour windows to test both Opus 5 variants. This time a single window was enough. On social media the model was framed as Fable-level but faster and cheaper than Opus 5; the first impression and the detailed numbers both pointed that way.

Scoreboard: 20 Out of 20 on Front and Back End

The score detail revealed a curious twist. In the spreadsheet both Opus 5.5 variants sat side by side. Medium scored a touch higher than high on code quality — plausible with a single try. Yet both were well above the 100-point ceiling others hit; three perfect 100s out of six runs were seen, averaged and then divided by five for the 20-point scale. Over a three-run average medium still edged high by a small margin, but high recovered on other axes.

The overall leaderboard update produced two new leaders above GPT-6 Astra. High took a clean 20 out of 20 across the four projects. Medium tripped once and logged 18 out of 19 elsewhere, still slightly ahead on code quality but behind high overall. Compared with Opus 5 the jump is stark, much closer to the 60 out of 60 ceiling. It is the second model to clear 20 out of 20 after Astra medium, but price and time are the bigger story. One viral post had argued we need not better models than Fable or Astra but normal models that are cheaper — Opus 5.5 matches that brief. Medium averaged 56 cents per API call versus over a dollar for either Opus 5 variant, and even Opus 5.5 high was cheaper than Opus 5 medium.

Why So Much Faster and Cheaper?

The official API price is only a little lower; the main saving comes from needing less compute, so the model is tuned for speed and cost overall. This was echoed by heavy users who said pushing hard barely dented their limits — Theo was cited. On timing, medium landed near two minutes per request, high near three, about twice as fast as Opus 5 and faster than Astra medium; the whole table has no other entry under two minutes except Luna low, which is off the leaderboard. Usage on the 20-dollar plan also eased: one spike over 100% was absorbed by a new cached reset, the test week stayed at 91% and then used 58% of a second window to finish, versus three to four windows for Opus 5. Anthropic also lifted five-hour window limits by 20% officially.

Community signal lined up. The author said he now trusts community feedback more than generic suites because fixing real repositories mirrors daily work. Opus 5.5 scored near the top there, cheaper than Fable, and social feeds filled with Halleluja-style praise about finally ignoring the usage cap on Fable. What is Fable's role at its premium price, or Sonnet 5 at a supposedly lower price? Sonnet 5 sits lower on the table and is neither much cheaper nor much better. Anthropic's note is clear: Opus 5.5 is the first in the Claude 5.5 family, so Sonnet and perhaps Haiku or even Fable 5.5 may follow soon. The author plans to keep adding new models to his benchmark and will cover GPT-6 Sol and Luna next, with deeper dives on his AI Coding Dealer site.