Don't switch models yet: Opus 5.5 and Fable 5.2 are only 1.9 points apart, but the bill differs by 2.8x

Garry Tan released a set of benchmarks measuring coding tasks in Claude Code 2.1.322. There were 52 Terminal-Bench 2.1 tasks; each model ran each task twice, for 312 total attempts across three models, all scored, with none dropped midway.

The results are straightforward: Opus 5.5 had an 84.6% pass rate, Fable 5.2 had 82.7%, a difference of 1.9 percentage points. The 95% interval is -9.6 to +5.8, crossing zero, so this run did not detect a difference in pass rate between the two. Cost is unambiguous: Opus 5.5 cost $0.65 per solved task, Fable 5.2 cost $1.83, or 2.80x, with a 95% interval of 2.05x to 3.99x.

The report includes a key caveat: Opus already solves most tasks, and distinguishing the two models with this set would require a harder test suite. Opus's 84.6% in this run means it is only about 15 points from a perfect score, and the ceiling itself compresses the gap. Fable 5.2 is not noticeably worse on this task set; the issue is that the tasks are not hard enough, and the extra cost does not reliably buy a higher pass rate.

One commenter replied that 84.6 vs. 82.7 is noise, sitting next to a 2.8x bill. Blunt, but consistent with the report's conclusion about measurable difference.

Addy Osmani's usage advice is closer to everyday practice: Opus 5.5 at medium is a very good daily workhorse. When it gets stuck, raise the effort level first, then switch models. Save Fable 5.2 for genuinely hard work, such as large migrations or long unattended runs. Effort is the setting for how much thinking, tool use, and self-checking the model is willing to spend on a task. In his claude.dev blog post, he lays out this path in more detail: first confirm whether the model can self-check, e.g., run tests, builds, or call APIs; if not, raise effort one level; if that is still not enough, switch to a larger model, and start from a lower effort.

This sequence is backed by cost data. The blog mentions that on a subset of SWE-bench Pro, Opus 5.5 run entirely at low and rerunning only failed tasks at high achieved about 97% pass rate at an average of $0.17 per task; running all tasks directly at high achieved 95.3% at an average of $0.29. Running at the lower setting first and retrying failures is cheaper than blindly maxing out. Another data point is that medium is about 2.5 percentage points lower than high and costs about 70% of high. Effort is fine-tuning; switching models is coarse-tuning. Many people switch to the most expensive model as soon as they hit a wall, skipping the two cheaper knobs in front of them. Switching models is usually faster than adjusting effort, but the bill does not slow down for it.

The blog also clearly lays out pricing for four models. Opus 5.5 is $4 per million input tokens and $20 per million output tokens; Fable 5.2 is $8 and $40. Both have cache reads at $0.20 per million tokens. In long sessions where cache reads dominate, the gap narrows; but for uncached input and output, Fable is still twice Opus. Haiku 5.5 is cheaper and suited to high-volume work such as classification, routing, and extraction, as well as sub-agents doing odd jobs for the main model. Having an expensive model read logs and find files is another kind of waste.

So when is Fable 5.2 worth using? One commenter asked whether complex planning and changes across million-line codebases should use it. Another approach is to have Fable 5.2 orchestrate, breaking tasks down for Opus and Haiku to execute. Someone else said they call in Fable only for bugs where Opus has failed twice. These are usage experiences, not benchmark conclusions, but the direction is consistent: put the expensive model on judgment and decomposition, and don't make it do everything.

What is truly counterintuitive about this data is that the gap between strong models may be so small that it takes harder problems to measure, while the bill does not lie. Fable 5.2 is certainly strong, but strong does not mean it is worth paying 2.8x more on this task set. Default to Opus 5.5 at medium, raise effort first, and only then consider switching models. If you want to switch to Fable 5.2, test it on your own tasks; don't make decisions off a leaderboard. Model choice ultimately comes down to the ledger.

Sources: Garry Tan's tweet, benchmark report, Addy Osmani's tweet, claude.dev blog.