I spent the last stretch of October watching Google’s model drop the same way I watch a new image model land on Civitai—curious, a little skeptical, and already thinking about the commercial brief sitting in my inbox. Gemini 4 (Argon) showed up with the usual fanfare. Gemini 3.1 Pro has been sitting there since February as a long-running preview. The internet immediately started ranking them. Most of those rankings are noise. Here’s the clean version that actually helps you decide what to run today.

Core Specs That Aren’t Marketing
Google never put the two models on the same scorecard. That alone should slow everyone down.
Item | Gemini 4 (Argon) | Gemini 3.1 Pro |
|---|---|---|
Status | Phased rollout | Preview (gemini-3.1-pro-preview) |
Output limit | 1M tokens | 65,536 tokens |
Input context | Not publicly stated | 1,048,576 tokens |
Multimodal | Heavy emphasis on long video, charts, documents | Text, image, video, audio, PDF |
Thinking levels | Highest level tested; others undocumented | Low / Medium / High (default High) |
Training cutoff | Not stated | January 2025 |
The one number Google made crystal clear is the output ceiling. Gemini 4 jumps from the previous 64K class to a full 1M tokens. Everything else is still thin.
Pricing Reality Check (the 200K Line)
Launch pricing for Gemini 4: $2 / $10 per million tokens (input/output). It later moves to $4 / $20.
Gemini 3.1 Pro: $2 / $12 under 200K tokens, $4 / $18 above that. Batch is half price on 3.1 Pro. Caching is cheaper on 4 during the launch window.
Quick math on two real workloads:
150K input + 20K output → Gemini 4 launch ≈ $0.50, later $1.00; 3.1 Pro ≈ $0.54
500K input + 20K output → Gemini 4 launch ≈ $1.20, later $2.40; 3.1 Pro ≈ $2.36
During the launch window Gemini 4 is competitive or cheaper. Once the higher rates kick in, short prompts favor 3.1 Pro and long prompts land roughly even. Remember that 3.1 Pro counts thinking tokens as output, so your actual bill can move.
Benchmarks You Cannot Directly Compare
Google’s own tables share almost nothing. Terminal-Bench uses different versions. Long-context tests are different tests. The scores that exist look strong on both sides in their own silos, but they do not form a head-to-head. Independent leaderboards are starting to fill in; treat them as directional, not gospel.

Practical Decision Rules
Use Gemini 3.1 Pro right now if:
You need the model in production today
Your agents rely on the custom-tool endpoint
Most of your prompts stay under 200K
You want documented thinking levels and batch pricing
Switch or test Gemini 4 when you get access if:
You regularly need outputs longer than 64K
You are running multi-document or long-video agent loops
You can afford the post-launch rate and still want the higher output ceiling
Migration checklist I actually use:
Keep the 3.1 Pro preview ID live until Gemini 4 has a public model ID and wider access.
Re-measure latency and cost under whatever thinking settings Gemini 4 eventually documents.
Cap max output tokens hard—1M is powerful and expensive when a loop goes sideways.
Keep 3.1 Pro as the fallback for anything that must not fail while access is still gated.
What I Am Doing in Practice
I run the same recurring task (contract summaries, codebase audits, long prompt-engineering loops for image models) on both whenever I can get tokens. Two weeks of side-by-side data beats any leaderboard. Quality, retry rate, and actual dollar cost per accepted answer are the only metrics that survive a real commercial brief.
Gemini 4 looks like the next flagship. On paper it has the bigger lungs. Gemini 3.1 Pro is the model you can call today, that still has a 1M input window, clear thinking controls, and lower short-prompt pricing. Until Google opens the gates wider and publishes the missing numbers, the practical choice for most production work remains 3.1 Pro, with Gemini 4 parked as the promising upgrade path you test the moment it becomes available.
Test the models on your own work. Everything else is just commentary.
For AI tool features, rules and availability, refer to their official sites.


Leave your thoughts here, too.