The Vibe Father
Roundup
Ask "what is the best coding model?" in mid-2026 and the literal answer is easy: Claude Fable 5, by every composite measure we track. It tops SWE-bench Verified at 95.0 — the first model ever past 95 — posts 89.8 on LiveCodeBench, and sits second on Terminal-Bench at 83.1. On our Vibe Coding Index, which weights SWE-bench Verified at 40% and Terminal-Bench and LiveCodeBench at 30% each, nothing touches it. You can see the live rankings, with every score linked to its public source, at /benchmarks.
But "best" is the wrong question if you stop there. The smarter question is "best for what" — because the model that wins the index is also the most expensive model on the board, and there are at least four other crowns worth handing out. This roundup names all five.
The board, July 2026
Everything below comes from public benchmark sources; where a lab has not published a score, we say so rather than guess. Missing scores never count as zero on our index — we renormalize instead. Full methodology in how we rank coding models.
| Model | SWE-bench Verified | Terminal-Bench | LiveCodeBench | Price (in/out per M) | Context |
|---|---|---|---|---|---|
 Claude Fable 5 | 95.0 | 83.1 | 89.8 | $10 / $50 | 1M |
| Claude Opus 4.8 | 88.6 | 78.9 | 87.8 | $5 / $25 | 1M |
| Claude Sonnet 5 | 85.2 | not published | 82.4 | $3 / $15 | 1M |
 GPT-5.5 | 80.6 | 83.4 | 85.3 | varies by tier | — |
 Gemini 3.5 Flash | 79.3 | 76.2 | 87.6 | $1.50 / $9 | 1M |
 GLM 5.2 | 78.7 | not published | not published | open-weight | — |
 DeepSeek V4 Pro | 77.6 | not published | 87.5 | $0.435 / $0.87 | 1M |
 Qwen3.7 Max | 77.3 | not published | 87.1 | $2.50 / $7.50 | 1M |
 Kimi K2.6 | 76.7 | 66.7 | 86.8 | open-weight | — |
| Gemini 3.1 Pro | 75.6 | 70.7 | 88.5 | $2 / $12 | 1M |
| GPT-5.3 Codex | 74.8 | 78.4 | 87.3 | $1.75 / $14 | 400k |
 Grok 4.5 | not published | not published | 87.4 | $2 / $6 | 500k |
The absolute best: Claude Fable 5
Released June 9, 2026, Fable 5 is the model to hand your hardest problems. Its 95.0 on SWE-bench Verified is a 6.4-point gap over its own stablemate Opus 4.8, and on real multi-file repository work the gap is visible, not statistical. It carries a 1M-token context window and it plans like a principal engineer. The catch is $50 per million output tokens — the priciest output on our board. We cover exactly who should pay that in our full Fable 5 review. One honest asterisk: it does not win everything. GPT-5.5's 83.4 on Terminal-Bench edges Fable's 83.1.
The daily driver: Claude Opus 4.8
Opus 4.8 is what we would call the best default. At 88.6 SWE-bench Verified, 78.9 Terminal-Bench, and 87.8 LiveCodeBench, it is elite on every axis while costing half of Fable — $5 in, $25 out. If you run one model for everything, run this one. The delta to Fable matters on the top few percent of problems; on the other 95% of your work, Opus is indistinguishable and your invoice is half the size.
The value builders: Sonnet 5 and GPT-5.3 Codex
If your workload is high-volume implementation — the builder seat in an agent team — two models dominate the price-performance frontier. Claude Sonnet 5 posts 85.2 on SWE-bench Verified at $3/$15, streaming at 89 tokens per second with 1M context. GPT-5.3 Codex trades some SWE score (74.8) for a strong 78.4 Terminal-Bench and 87.3 LiveCodeBench at $1.75/$14. Both will happily churn through a hundred routine tasks for the cost of a dozen Fable runs. We map which model belongs in which seat in the best model for each agent role.
The terminal king: GPT-5.5
Here is the result that surprises people: GPT-5.5 holds the top Terminal-Bench score on our entire board at 83.4, ahead of even Fable 5. If your agents live in a shell — installing dependencies, debugging builds, running multi-step ops — that number is the one to watch, and OpenAI owns it. Its 80.6 SWE-bench and 85.3 LiveCodeBench are strong but not chart-topping; the terminal is where it earns its crown. We explain why shell competence diverges from pure coding skill in our Terminal-Bench explainer.
The budget pick: DeepSeek V4 Pro
At $0.435 per million input and $0.87 per million output, DeepSeek V4 Pro is absurdly cheap for what it delivers: 77.6 on SWE-bench Verified, 87.5 on LiveCodeBench, 1M context, open weights. Its Terminal-Bench score is not yet published, so we renormalize its index around the two scores that exist rather than penalizing it — but on the numbers we have, nothing else comes close per dollar. Full breakdown in the cheapest coding models worth using.
How to read a board like this
A few habits keep the hype out. Read the three benchmarks as three different questions, not one: SWE-bench Verified asks "can it fix a real bug in a real repo," Terminal-Bench asks "can it drive a shell unsupervised," LiveCodeBench asks "can it solve problems it has never seen." A model can excel at one and lag another — Gemini 3.1 Pro's 88.5 LiveCodeBench next to its 75.6 SWE-bench is the textbook case of a brilliant puzzle-solver that is merely decent at repository surgery. We unpack each suite in the SWE-bench explainer, the Terminal-Bench explainer, and the LiveCodeBench explainer.
Then put price next to the scores, because the deltas rarely cost what they charge. The gap from Sonnet 5's 85.2 to Opus 4.8's 88.6 costs you a 67% higher input price; the gap from Opus to Fable costs 100% more. On a heavy month of 50M input and 10M output tokens, that is $300 versus $500 versus $1,000 for three models from the same lab. Whether the top step is worth it depends entirely on how often you hit problems the cheaper step fails — for most people, that is less often than they think, and for a few it is daily.
What we would actually run
If we had to pick one model, it is Opus 4.8. If we had to pick a team, it is Fable 5 planning, Sonnet 5 or Codex building, and GPT-5.5 or a cross-lab reviewer checking the work — because a reviewer from a different lab has no loyalty to the builder's mistakes. And two scores to watch: Grok 4.5 launched July 9 with an 87.4 LiveCodeBench and no SWE or Terminal-Bench numbers yet, and GLM 5.2's 78.7 SWE-bench makes it the strongest open-weight repo model pending its other scores. The board moves monthly. The live version, refreshed nightly, is always at /benchmarks.