We tested Fable 5.2, Opus 5.5 and Codex on our own code. Extra high effort takes almost twice as long and isn't any smarter
From November 10 to 14 we put three models through tasks taken from our working code. The short version: for everyday coding we use Opus 5.5 on high effort, it gives the best quality for the money. Fable 5.2 is a little smarter on long tasks and a stronger reviewer, but it costs about a third more. Codex running GPT-6 Astra keeps up with Opus and uses 4 to 5 times fewer tokens.
How we tested
Every task came from our own git history: a bash script, a Swift app, research and a plan for moving a server, and fixes that came with a test that fails before the fix. A blind judge scored the work without knowing which model did it. In one of the four long tasks the blinding broke: the author's name was left in a file. Claude Code and Codex ran in their usual apps, the same way we use them every day.
Fable 5.2 vs Opus 5.5: which is better?
| What we compared | What others report | What we got | Match |
|---|---|---|---|
| Long tasks | Anthropic: Fable is almost twice as strong at remote work (Remote Labor Index 41.3% vs 21.3%); Cursor on long tasks: 56.3% vs 56.0% | judges picked Fable in 3 tasks out of 4, each time by a small margin | partly: our gap is smaller |
| Short code | no separate benchmarks found | both passed 27 of 27 tests | our data |
| Speed | an enthusiast on X: Fable finished a big task an hour and a quarter faster | Fable was 30% slower for us: 107 minutes vs 83 across four tasks | no |
| Reviewing other work | paddo.dev: Fable found the flaw in 4 runs out of 4, Opus in 1 | Fable's findings in Opus's work held up 13 of 13, Opus's findings in Fable's work 5 of 7 | yes |
Fable is smarter on long tasks and at reviewing, but it takes a third longer and burns a third more tokens.
Which Claude Code effort level should you use?
Same two tasks from our code, same blind judge. Opus 5.5 in Claude Code and GPT-6 Astra in Codex at different effort levels:
| effort level | Opus medium | Opus highour pick | Opus extra high | Astra medium | Astra high | Astra extra high | Astra ultra |
|---|---|---|---|---|---|---|---|
| Minutes for both tasks | 4.5 | 4.7 | 8.5 | 3.4 | 4.9 | 9.8 | 10.2 |
| Tokens, M | 4.1 | 5.7 | 10.3 | 1.0 | 1.2 | 1.9 | 2.0 |
| Judge scores (medium / hard task) | 7 / 7 | 8 / 9 | 8 / 8 | 8 / 7 | 8 / 7 | 8 / 8 | 9 / 8 |
| Errors on 40 tricky bash commands | 6 | 4 | 0 | 2 | 2 | 1 | 0 |
| Kept the neighboring test intact | no | yes | no | no | no | no | no |
| Left junk files behind | no | no | no | no | no | yes | yes |
All seven passed the tests we set. Astra writes our code about as well as Opus: a bit better in places, a bit worse in others. On edge cases Astra is more reliable: 2 errors on medium and high, against 6 and 4 for Opus. But only Opus on high kept intact the neighboring test it wasn't asked to touch.
Extra high runs almost twice as long and eats almost twice the tokens, and the scores don't go up. It has one upside: zero errors on the 40 bash commands, against four. Cursor saw the same thing: Opus scored an identical 56.0% on high and extra high, and high cost almost half as much. That's why we run high everywhere.
We don't use max at all. Two nights in a row, Opus on max kept getting cut off by a safety filter, and instead of 29 work items we got 4.
Claude vs Codex: which is better for coding?
| What we compared | What others report | What we got | Match |
|---|---|---|---|
| Quality | Vals, Terminal-Bench: Opus 65.2%, Astra 59.6% | judge scores of 7 to 9 for both | closer for us |
| Tokens | Vals, Terminal-Bench: Astra uses 9.4 times fewer per task; testers on Habr measured 3.2 to 4.2 times | 4 to 5 times fewer | yes, fewer |
| Edge cases | no data | on 40 bash commands, Astra made 0 to 2 errors, Opus on medium and high made 4 to 6 | our data |
| Cleaning up | no data | on extra high, Astra leaves helper files behind | our data |
Our judge was a Claude model, so it had a horse in the race. And there were only two tasks. Claude and Codex also count tokens differently: almost all of Codex's tokens are cache it reads again.
From our community: Astra is better at reviewing code, Opus is better at writing it.
How fast does Fable eat your subscription limit?
In 18 minutes on Fable, on the Ultra plan with ten subagents running at once, the five-hour window went from 10% to 35%. A whole night and morning of subagents on Opus, around 9.8M tokens, used 3% of the weekly limit. Fable is fine for the odd hard task. Run it all day and the limit is gone.
Models catch mistakes in each other's work, not their own
Both models left costly mistakes in their server migration plans: three from Fable, one from Opus. Most of them were caught by the rival model, not the author. Our rule now: before any step you can't undo, a different model reviews the work.
- 1database copy
- 2keys to the new server
- 3service checks
- 4switch the address
- 5shut down the old one
- 1database copy
- 2keys to the new server
- 3service checks
- 4switch the address
- 5shut down the old one
What to do if Claude bans your account
On November 13 at around 21:30, the Claude account on one of our Macs stopped responding and sent us to a review. Anthropic gave no reason. Our overnight memory processing stopped, and so did our server, because its key was tied to the same account. By morning we were working from another account and had lost no data: our chats are stored as files on our side and our memory lives on our own server.
- 21:30account sent to review, no reason given
- nightmemory processing and the server stopped: the key was on the same account
- morningworking from another account, no data lost
We're not the only ones. According to Anthropic's transparency report, the company banned 11.4 million accounts in the first half of 2026, almost 8 times more than in the previous half. You can only appeal from inside the banned account. According to press reports, Anthropic's shares will be priced on November 18 and start trading on the 19th. The idea that the bans are linked to the IPO is only a rumor for now, with no primary source.
What to do before it happens:
- keep your chats, memory and working files on your side, not only in the model's cloud
- have a second account or a second provider ready
- don't put your server key on the same account your team works from
Which model for which job
| Job | What to use |
|---|---|
| Everyday coding | Opus 5.5 on high |
| Long hard tasks, review before an irreversible step | Fable 5.2 |
| Second developer, saving tokens | Codex on GPT-6 Astra, reviewed by Opus |
| Extra high and max effort | we don't use them |
AI-powered development for your project
We write code with AI agents and proper checks, just like in this article: websites, bots, mini apps and internal tools.
From $1,500
FAQ
- Is Fable 5.2 smarter than Opus 5.5?
- A little, on long tasks, and it's better at reviewing other models' work. On short code the result is the same, and Fable is slower.
- Which effort level should I use in Claude Code?
- High. For us, extra high took twice as long with no gain in quality.
- Claude or Codex: which is better for coding?
- On quality we found them close. Codex uses 4 to 5 times fewer tokens, while Opus on high pays more attention to the code around the change. We keep Opus as our main model and Codex as the second.
- What's the best AI for coding?
- For everyday code, Opus 5.5 on high came out on top for us. For long tasks and reviews we use Fable 5.2, with Codex as a second developer.
- Why does Claude ban accounts?
- Anthropic usually doesn't say. You appeal from the banned account itself. Keep your work on your side and have a backup account, and a ban won't stop your business.
If you've tested these models on your own tasks, tell us what you found and we'll compare notes.
