The Vibe Father
Model review
Gemini 3.1 Pro landed on February 19, 2026, from Google DeepMind, and five months later we can say exactly what it is: the most specialized frontier model on our leaderboard. Its 88.5 on LiveCodeBench is the second-highest score we track — behind only Claude Fable 5's 89.8, ahead of every other flagship from every other lab. Its SWE-bench Verified score is 75.6. That thirteen-point spread between contest coding and real-repo work is the entire review, and understanding it will save you money in one direction or frustration in the other.
It ships with a 1M-token context window, streams at 147 tokens per second, and costs $2 per million input, $12 per million output. Solid numbers everywhere. But the benchmark shape is what you're actually buying.
The numbers
Our Vibe Coding Index weights SWE-bench Verified at 40% because resolving real GitHub issues is the job most of us are paying for, with Terminal-Bench and LiveCodeBench at 30% each. Here's 3.1 Pro against its nearest rivals:
| Model | SWE-bench Verified | Terminal-Bench | LiveCodeBench | $ in / $ out per M |
|---|---|---|---|---|
 Claude Fable 5 | 95.0 | 83.1 | 89.8 | 10.00 / 50.00 |
 Gemini 3.1 Pro | 75.6 | 70.7 | 88.5 | 2.00 / 12.00 |
| Gemini 3.5 Flash | 79.3 | 76.2 | 87.6 | 1.50 / 9.00 |
 Grok 4.5 | — | — | 87.4 | 2.00 / 6.00 |
The awkward row in that table is DeepMind's own. Gemini 3.5 Flash — released three months later, cheaper on both ends, 20 tok/s faster — beats 3.1 Pro on SWE-bench Verified by 3.7 points and on Terminal-Bench by 5.5. The only column where Pro wins is LiveCodeBench, by 0.9. In our weighted index, Flash comes out ahead of the model that carries the "Pro" badge. That's not a knock on DeepMind; it's what happens when a lab ships fast. But it means the case for 3.1 Pro has to rest entirely on that contest-coding score.
The price math
A heavy month — 50 million input tokens, 10 million output — runs 50 × $2 + 10 × $12 = $220. The same month on Gemini 3.5 Flash is $165. On Grok 4.5, which sits 1.1 points behind on LiveCodeBench, it's $160. So you're paying a $55-to-$60 monthly premium over the nearest alternatives for nine-tenths of a point on one benchmark, while giving up ground on the other two. If you spend most of your tokens on ordinary repo work, that math doesn't close — our cheapest-models breakdown shows how far $220 goes elsewhere. If you spend them on algorithmic problems, keep reading.
Its best seat on the team
3.1 Pro's seat is the algorithm specialist. LiveCodeBench measures self-contained, contest-style problem solving — clean spec, checkable answer, no legacy code in the way (we break down what the benchmark does and doesn't measure in our LiveCodeBench explainer). At 88.5, this model out-solves Claude Opus 4.8, GPT-5.3 Codex, and everything else short of Fable 5 on exactly that kind of work. Hand it the hard function: the scheduling algorithm, the parser, the query optimizer, the data-structure problem your builder model has been circling for three attempts. It's genuinely elite there, and at $12 per million output it's a far cheaper elite than Fable 5's $50.
What the SWE-bench gap means in practice — and our SWE-bench Verified explainer covers why this benchmark is the one we weight heaviest — is that 3.1 Pro is weaker the messier the repository gets. Real issues demand navigating unfamiliar code, respecting conventions nobody documented, and changing the right file without breaking three neighbors. At 75.6, it does that noticeably less reliably than its own cheaper sibling. The 70.7 Terminal-Bench tells the same story about multi-step shell sessions. So: don't make it your default builder. Make it the specialist your builder escalates to when the problem is algorithmic rather than archaeological, and keep Flash or another all-rounder from our rankings in the daily-driver seat.
In practice, the escalation pattern is simple to wire up: your default builder flags a task as algorithm-shaped — usually because its second attempt still fails the tests — and the ticket routes to 3.1 Pro with the function signature, the failing cases, and nothing else. Contest-style problems don't need the whole repo in context, which is exactly why this model thrives on them, and stripping the context down also keeps the $2-per-million input bill honest. We've watched it one-shot problems that a cheaper generalist had circled for half an hour. That's the trade you're buying: not a better colleague, a better closer.
Verdict
Gemini 3.1 Pro is a legitimately elite contest coder wearing a general-purpose price tag. Used as a specialist — algorithmic escalations, competitive-programming-shaped problems, isolated hard functions — it earns its $220 month easily. Used as your everything model, it's beaten by its own younger sibling at 75% of the price. Who should skip it: anyone whose work is mostly real-repo maintenance rather than algorithm design; anyone already paying for Gemini 3.5 Flash, which is the better generalist from the same lab; and budget-tier teams, for whom the open-weight models in our open-weight guide deliver 86-to-87 LiveCodeBench scores at a fraction of the cost. Specialists are worth having. Just don't schedule one for every shift.