Claude Opus 5.5 shipped yesterday. Fable 5.1 level on most work, 40% cheaper to run than Opus 5, bigger five-hour limits, a banked reset you can spend when you like. Sonnet 5.5 and Haiku 5.5 to follow. Every model lab dropped something this week, so I am not going to add to the benchmark screenshots.
What I did instead: updated Claude Code, noticed that --model opus now resolves to claude-opus-5-5 with no announcement, and ran my two boring test tasks on four models. Twice each. Sixteen runs, all claude -p, project settings only so none of my CLAUDE.md or plugins leak in.
The two tasks
Same tiny repo as my do-not-overengineer experiment: a constants.ts with four exports and a six-line fetchUser. Task one: “Add retry with exponential backoff to fetchUser. Keep it minimal.” Task two: “Add unit tests for constants.ts using bun test.” I measure wall time, cost from the JSON output, output tokens, lines added, and for the tests task, how many assertions are nothing but the constant restated (expect(MAX_RETRIES).toBe(3)).
One accident made it better. I ran with --permission-mode acceptEdits, which allows file writes but not Bash, so no model could actually run bun test. Every one of them shipped unverified tests. Every one of them also said so, unprompted. Opus 5.5 opened with “I haven’t run it.” Good.
Retry task, averaged over two runs
| Model | Time | Cost | Output tokens | Lines added |
|---|---|---|---|---|
| Opus 5 | 29s | $0.19 | 2129 | 17.5 |
| Opus 5.5 | 13s | $0.09 | 904 | 12 |
| Fable 5.1 | 21s | $0.26 | 1211 | 10 |
| Sonnet 5 | 10s | $0.05 | 610 | 8.5 |
Opus 5.5 is the 12-line version. It retries on network errors, 429 and 5xx, throws straight away on other 4xx, reuses the MAX_RETRIES constant it found in the file next door, and does it in four turns. Opus 5 wrote the same logic plus a sleep helper, a comment explaining 4xx, and twice the output tokens. Fable was tighter than both and cost the most. Sonnet was smallest and fastest, and skipped the 429 case.
for (let attempt = 0; ; attempt++) {
let res: Response | undefined;
try {
res = await fetch(`${API_BASE}/users/${id}`);
} catch (err) {
if (attempt >= MAX_RETRIES) throw err;
}
if (res) {
if (res.ok) return res.json();
const retryable = res.status === 429 || res.status >= 500;
if (!retryable || attempt >= MAX_RETRIES) throw new Error(`fetchUser ${id}: ${res.status}`);
}
await new Promise((r) => setTimeout(r, 2 ** attempt * 100));
}That is Opus 5.5’s whole diff. I would merge it.
Tests task, averaged over two runs
| Model | Time | Cost | Tests written | Tautological |
|---|---|---|---|---|
| Opus 5 | 49s | $0.24 | 11 | 3.5 |
| Opus 5.5 | 17s | $0.11 | 9 | 3.5 |
| Fable 5.1 | 27s | $0.28 | 10.5 | 4 |
| Sonnet 5 | 29s | $0.07 | 7.5 | 4 |
Two things. First, Opus 5.5 is again a third of Opus 5’s time and less than half its cost, with fewer, tighter tests. Second, every model, including Fable, wrote three or four tests that restate the source. API_BASE equals the URL. DEFAULT_TIMEOUT_MS is 30000. This is not a model quality problem, it is a prompt problem. I ran Opus 5.5 once more with three words added, “No tautological tests”, and got eight tests and zero restatements, plus a sentence explaining that asserting the literal would just copy the source. Three words. Put them in your CLAUDE.md.
Totals
| Model | 4 runs | Wall time | Output tokens |
|---|---|---|---|
| Opus 5 | $0.86 | 154s | 10743 |
| Opus 5.5 | $0.39 | 60s | 4870 |
| Fable 5.1 | $1.08 | 95s | 6459 |
| Sonnet 5 | $0.24 | 79s | 4023 |
The list price moved from $5/$25 to $4/$20 per million tokens, which is 20% not 40%. The rest of the claim is token efficiency, and on these two tasks it was real. Half the output tokens of Opus 5 for a diff I preferred. Total came out 55% cheaper and 2.5x faster.
What changes for me
Not much, which is the point. opus already means 5.5 in Claude Code, so my routing setup got cheaper without an edit. Fable stays for the hard, long jobs where it earned the extra 20 seconds here. Sonnet stays for one-liners. And I am now watching the five-hour meter with a banked reset in my pocket, which feels a lot like having a free pass on a bad afternoon.
Same tasks, same repo, same prompts. Just cheaper.
