Gemini Pro

Concept of the skill

What it is: A routing skill that decides when a task should go to Google's Gemini Pro frontier model — the current Gemini 3.1 Pro generation (moving to Gemini 3.5 Pro) — rather than Claude Opus or GPT-5. It is model-selection judgment, not a Gemini tutorial.

Mental model: Three signals push work toward Gemini Pro and one pushes it away. PUSH TOWARD: (1) the context is genuinely large — a corpus, codebase, or long document near or beyond 200K tokens, where Gemini Pro's 1M-token input avoids chunking-and-retrieval entirely; (2) the input is mixed-multimodal — audio + video + PDF + image in one call, which Gemini accepts natively; (3) the budget wants frontier-grade reasoning at the best capability-per-dollar. PUSH AWAY: the task is in a lane another frontier model owns (Claude Opus for the hardest agentic code modification; GPT-5 for best-in-class long-context retrieval quality and hard math).

Why it exists: Agents default to one model out of habit and pay for it — either too much (sending a 30K classification to the frontier) or in capability (chunking a 500K-token corpus through a 200K-window model when a 1M-window model would read it whole). This skill replaces "use my usual model" with an evidence-based per-lane decision, and corrects the most expensive misconception: that "has a 1M context window" means "best at using a 1M context window."

What it is NOT: It is not the cheap/fast tier decision (that's gemini-flash), not agent-system architecture (agent-engineering), not request-time dispatch among your local skills (skill-router), and not a Gemini API integration guide.

Adjacent concepts: gemini-flash (the cheap/fast Gemini tier and the escalation boundary up to Pro); agent-engineering (how to compose models into a system, one layer up from picking one model); skill-router (dispatches among skills, not among external models); autonomous-loop-patterns (the loop that decides which model each step in a pipeline uses).

One-line analogy: Gemini Pro is the wide-load freight truck of frontier models — when the cargo (context) is too big to fit any other vehicle in one trip, it wins by capacity, even though a sports car (Claude Opus / GPT-5) is faster on the specific stretch of road each owns.

Common misconception: That a 1M-token context window means Gemini Pro is the best at finding the needle in 1M tokens. Context capacity and long-context retrieval quality are different properties — GPT-5 leads on retrieval quality even where Gemini holds more tokens. Route to Gemini Pro to FIT the context cheaply; verify the retrieval-quality need separately.

Coverage

• The three PUSH-TOWARD signals for the Gemini Pro frontier tier: very-large single-call context, mixed-multimodal input, and frontier reasoning at the best capability-per-dollar.
• The PUSH-AWAY (escalation) boundary: which tasks another frontier model owns (Claude Opus for the hardest agentic code modification; GPT-5 for long-context retrieval quality and hard math).
• The 1M-token context window and what it buys (whole-corpus / whole-codebase / long-document reasoning without a chunk-and-retrieve pipeline).
• The 200K-token pricing cliff: the whole request reprices at ~2× once the prompt crosses 200K tokens, across both input and output.
• Native multimodal input (text · image · audio · video · PDF in one call) and what it removes (media preprocessing).
• The per-lane benchmark comparison against Claude Opus and GPT-5 — capacity vs code-modification vs retrieval-quality — and how to read it at routing time.
• That the concrete model id and prices move fast; the tier is stable, the facts must be re-checked at routing time.

Philosophy of the skill

Model choice is a routing decision driven by the task's dominant constraint, not by which model an agent reached for last time. The defaulting habit costs in two directions at once: paying frontier rates for work that never uses the capability, or crippling a task by forcing a giant corpus through a small-context model when a single-call large-context model would read it whole. The discipline is to name the dominant constraint first — is the binding limit context size, multimodality, capability-per-dollar, or a lane another model owns? — and route on that, not on familiarity. The skill also insists on a distinction that is easy to collapse and expensive to get wrong: capacity is not quality. Holding the most tokens is not the same as retrieving best from them. So the rule is to route to this tier to FIT the context cheaply, and to verify any needle-precise retrieval or hard-reasoning need against the model that actually leads that lane — never to assume the biggest window also wins every property measured inside it.

When to Route to Gemini Pro

Route here when ONE OR MORE of these is the dominant constraint:

Capability Sheet — Gemini 3.1 Pro (current Pro frontier, June 2026)

The 200K pricing cliff is the trap: a 201K-token prompt costs 2× per token versus a 199K-token prompt — across input AND output, the whole request. If a task can be kept under 200K, do it; if it genuinely needs the 1M window, budget for the 2× tier.

Strengths vs Weaknesses

Strengths

• Largest practical single-call context among the frontier trio — the only one that fits a ~1M-token corpus without retrieval engineering.
• Native, broad multimodal input in a single prompt.
• Best capability-per-dollar at the frontier (the "volume tier" default in the April-2026 cross-model snapshot).
• Strong general reasoning (GPQA Diamond ~94%) and competitive SWE-Bench Verified (~80%).

Weaknesses (the escalation boundary)

• Not the code leader. Claude Opus leads SWE-Bench Pro (the hardest agentic code-modification benchmark) — route the hardest code-mod work to Opus.
• Context size ≠ retrieval quality. GPT-5 leads long-context retrieval quality and hard math — if the task is needle-in-haystack precision or heavy math, verify against GPT-5.
• Fast-moving generation. The concrete model id (gemini-3.1-pro-preview → gemini-3.5-pro) and prices change; treat the tier as stable, re-check the concrete facts at routing time.

Verification

Confirm a Gemini-Pro routing decision before committing to it — and confirm the model facts it rests on are still current.

• [ ] The task's dominant constraint was named explicitly (context size / multimodality / capability-per-dollar / a lane another model owns) before a model was chosen — not picked by habit.
• [ ] If routed here for context size: the input genuinely exceeds a 200K window AND needs cross-document reasoning (not a lookup a retrieval pipeline would serve more cheaply).
• [ ] If the prompt crosses 200K tokens, the ~2× pricing-cliff cost (input AND output, whole request) was accounted for — or the prompt was kept under 200K where feasible.
• [ ] The decision did not assume "biggest context window" equals "best long-context retrieval" — a needle-precise retrieval or hard-math need was verified against the model that leads that lane (GPT-5), and the hardest agentic code-mod against Claude Opus.
• [ ] The concrete model id and per-token prices were re-checked against references/model-facts.md (or a live source) at routing time, since the Gemini generation moves fast.
• [ ] Pro is paid-only was accounted for (the free AI Studio tier covers Flash / Flash-Lite, not Pro).

Do NOT Use When

References

• references/model-facts.md — current context/pricing/benchmark facts with sources (Last updated 2026-06-08)
• gemini-flash skill — the cheap/fast Gemini tier and the boundary where work escalates up to this skill