The Vibe Father
Claim vs board
When OpenAI shipped GPT-5.6 Sol on July 9, 2026, it did the thing labs do on launch day: it named a rival and drew a line above it. OpenAI reports Sol scoring 80 on Terminal-Bench 2.1 — which it says is 2.8 points above Claude Fable 5 on that eval — while using less than half the output tokens, taking less than half the time, and costing about a third less per answer. That is a specific, gutsy claim aimed at the model most people consider the coding king.
We run a live coding leaderboard, so our job today is not to referee the vibes. It is to put OpenAI's claim next to what our board actually shows, be honest about which comparisons are legitimate, and tell you what to wait for before you crown anyone.
What OpenAI is claiming
To be scrupulous: every number above is OpenAI's, on OpenAI's harness, on the Terminal-Bench 2.1 configuration. OpenAI claims Sol is its "best coding model yet" and a new state of the art on that specific eval. It has not published SWE-bench Verified or LiveCodeBench figures for Sol. So the entirety of the "beats Fable" case, as it stands on launch day, rests on one internal Terminal-Bench 2.1 run and a set of efficiency figures that no independent party has reproduced.
None of that makes the claim false. OpenAI already owns the top Terminal-Bench score on our board with GPT-5.5 (83.4), so a strong terminal result from its new flagship is entirely plausible. But "plausible" and "verified" are different words, and the gap between them is exactly where launch-week hype lives.
What our board actually shows
Here is Claude Fable 5 as it sits on our live board today, next to the OpenAI models it competes with. These are independent, verified numbers as of July 2026:
| Model | SWE-bench Verified | Terminal-Bench | LiveCodeBench | Price (in/out per M) |
|---|---|---|---|---|
 Claude Fable 5 | 95.0 | 83.1 | 89.8 | $10 / $50 |
 GPT-5.5 | 80.6 | 83.4 | 85.3 | varies |
| Claude Opus 4.8 | 88.6 | 78.9 | 87.8 | $5 / $25 |
| GPT-5.6 Sol | — | — | — | $5 / $30 |
Fable 5 sits at 95.0 on SWE-bench Verified — the real-repo-work crown, and the widest lead on the board — with 83.1 on our Terminal-Bench config and 89.8 on LiveCodeBench. Sol's row is dashes on purpose. It has no independent score yet, so we show it none. That is the whole discipline of the piece.
The cross-eval caveat that breaks the comparison
Now the part everyone skips. OpenAI's claim is on Terminal-Bench 2.1. Our board's Terminal-Bench column, where Fable reads 83.1, is a different eval version — a different configuration, a different run. You cannot subtract "Fable 83.1 on ours" from "Sol 80 on 2.1" and conclude anything, and you equally cannot line up "Sol 80" against "Fable's 2.1 number" using our 83.1 as a stand-in. They are not the same test.
This is not pedantry. Benchmark versions differ in task sets, harness rules, and scoring, and a 2-to-3-point gap is well inside the range that version differences alone can produce. When a lab reports beating a rival "on this eval," the four words "on this eval" are load-bearing. Sol may well be a superb terminal model — GPT-5.5's board-topping 83.4 says OpenAI can build one — but "2.8 points over Fable on 2.1" is a claim about one eval, not a coronation.
Where Fable's real lead lives
Even taking OpenAI's Terminal-Bench claim at face value, it addresses one of three legs. Terminal-Bench measures agentic shell work — driving a terminal, chaining commands, recovering from errors. Fable 5's signature strength is the other end: SWE-bench Verified at 95.0, the deep, cross-file, real-repository reasoning that decides whether an agent resolves a gnarly issue without breaking two other things. OpenAI has said nothing about Sol on SWE-bench. Until it does — and until someone independent confirms it — Fable's most important lead is entirely unchallenged.
Same story on LiveCodeBench, where Fable posts 89.8. No Sol number exists. So of the three benchmarks that make up the Vibe Coding Index — SWE-bench Verified 40%, Terminal-Bench 30%, LiveCodeBench 30% — OpenAI has made a contested claim about one and gone silent on the two that carry 70% of the weight. That is not a knock on Sol; it is a statement about how little we actually know.
What to wait for before you crown anyone
Three things, in order. First, an independent SWE-bench Verified run for Sol — the one that would tell us whether it can touch Fable's 95.0 on real repo work, or whether it inherits GPT-5.5's profile of terminal-strong, reasoning-merely-good. Second, an independent LiveCodeBench score. Third, a third-party Terminal-Bench run on a common config, so the flagship comparison is finally apples to apples.
When those land, they go on the leaderboard and we will say plainly what they mean, including if the answer is "Sol trades blows on terminal work and Fable keeps the reasoning crown," or "Sol is the new number one," or anything between. If you want the current champion's full profile in the meantime, our Claude Fable 5 review has it, and our cross-lab shootout walks the matchups that Sol will have to win to change the standings.
The practical move is to not pick a side from a launch post at all. Because the Vibe Father runs Fable and every OpenAI model side by side on one macOS command deck, you can drop Sol into a real task the day its API opens, run it against Fable on your repo behind a verification gate, and let your own tests break the tie the benchmarks haven't settled. That beats waiting for consensus — and it beats trusting a chart that quietly measured two different tests.
