Gemini 4 Argon vs Gemini 3.1 Pro: The Successor Is Cheaper, and the Old Model Still Wins Two Fights

Gemini 4 Argon costs less to run than the model it is meant to replace. That is not a typo. Go​ogle's introductory output rate for Gemini 4 Argon is $10 per million tokens and Gemini 3.1 Pro Preview — the model sitting in the OrcaRouter catalogue today, and the one Argon's whole existence is an argument against — charges $12. Same $2 input on both, unless you push Gemini 3.1 Pro Preview past 200,000 input tokens, at which point its rate doubles to $4 in and $18 out. A cheaper successor is unusual enough to be worth the lede on its own.

The rest of the picture is less tidy. On Artificial Analysis's Intelligence Index, Gemini 4 Argon scores 52.6 to Gemini 3.1 Pro Preview's 29.7 — a 22.9-point gap that would, on its own, suggest the older model is not in the same league. But the older model wins GPQA Diamond 94.1% to nothing published, wins long-context reasoning 82.0 to 79.7, and posts a 95.6 τ²-Bench agentic tool-use score that DeepSeek has been quoting as its own flagship's headline. So the honest framing is not "Argon replaces 3.1 Pro". It is "Argon replaces 3.1 Pro's job, at a lower output price, while 3.1 Pro keeps a couple of narrow advantages that will matter to a specific minority of workloads."

What Gemini 3.1 Pro Preview actually is

It has been in the catalogue since 2026-02-19 and it has been called a preview for all seven months since. Google spent the intervening period shipping Flash-tier models — the 3.6 Flash, the 3.8 Flash, the cybersecurity-tuned 3.8 Flash Cyber — while the Pro tier stayed where it was. That gap is the context for everything Argon claims, and it is why a May-era question ("when does the Pro tier get replaced?") is still the right question in October.

The card is a real frontier model, not a leftover. 1,048,576-token context, multimodal input across text, image, audio, video and files, tool use, structured output and reasoning, and five capability tags on our side. Its weaknesses are where it was trained: a 65,536-token maximum output, and a task profile that looks dated next to a 2026 model.

The table you should read twice

Prices, both as listed, per million tokens:

• Input — Gemini 4 Argon $2.00 versus Gemini 3.1 Pro Preview $2.00. Level.

• Output — Gemini 4 Argon $10.00 versus Gemini 3.1 Pro Preview $12.00. Argon is 17% cheaper.

• Cached input — Gemini 4 Argon $0.10 versus Gemini 3.1 Pro Preview $0.20.

• Long-input tier — Gemini 4 Argon states a flat introductory rate; Gemini 3.1 Pro Preview steps to $4.00 in / $18.00 out above 200,000 input tokens.

• Maximum output — Gemini 4 Argon 1,000,000 tokens versus Gemini 3.1 Pro Preview 65,536.

• Measured cost per index task — Gemini 4 Argon $1.99 versus Gemini 3.1 Pro Preview $0.67, per Artificial Analysis. Here the older model is three times cheaper, because it is a much smaller job for the harness to run.

That last line is the reconciliation. Per-token Argon is cheaper; per-task it is three times more expensive, because it answers harder questions. If your workload is fixed-input, fixed-output — classification, extraction, summarisation at a known length — the per-token numbers are the ones that hit your invoice. If your workload is "solve this", the per-task number is the honest one and Argon's premium is real.

The 22.9-point gap, and where it comes from

Artificial Analysis runs a fixed battery, so a gap that large should show up consistently. It mostly does, and the pattern is informative:

• Terminal-Bench 4.0 — Gemini 4 Argon 57.1% versus Gemini 3.1 Pro Preview 4.0%. Not a misprint. Long-horizon software engineering is the single largest capability change between the two.

• AutomationBench-AA — Gemini 4 Argon 77.5% versus Gemini 3.1 Pro Preview 35.4%.

• ApexAgents — Gemini 3.1 Pro Preview 32.0% against no published Argon figure on the same board.

• SciCode — Gemini 4 Argon 61.8% versus Gemini 3.1 Pro Preview 58.7%. Close.

• Humanity's Last Exam — Gemini 4 Argon 57.1% versus Gemini 3.1 Pro Preview 47.0%.

• Long-context reasoning — Gemini 3.1 Pro Preview 82.0% versus Gemini 4 Argon 79.7%. The old model wins, narrowly.

• GPQA Diamond — Gemini 3.1 Pro Preview 94.1%. Argon publishes nothing on this board.

• Omniscience — Gemini 4 Argon 42.4 versus Gemini 3.1 Pro Preview 31.9.

• GDPval — Gemini 4 Argon 1,611 versus Gemini 3.1 Pro Preview 776.

• ITBench-SRE — Gemini 4 Argon does not appear; Gemini 3.1 Pro Preview 30.3%.

So: the gap is almost entirely about agentic endurance and long-horizon task execution, and it is enormous there. On scientific multiple-choice reasoning, on a 1M-token reasoning task and on GPQA, the two models are within noise or the older one leads. Terminal-Bench 4.0's 4.0% against 57.1% is the number that should decide most migration decisions, and it is the number the older model's defenders will skip past.

Two specifications that reshape your architecture

Argon's output ceiling is 1,000,000 tokens, up from the 64K that the previous generation capped at — and 64K is exactly what Gemini 3.1 Pro Preview offers, 65,536 to be precise. Sixteen times the headroom in a single trajectory changes what a single call can be asked to do.

Concretely: anything you currently do by chaining calls to stay under a 65K output cap — multi-file patch generation, long audit reports, full document drafting in one pass — becomes a single call. Fewer calls means fewer places for an agent loop to lose state, and it means your retry logic gets simpler rather than more elaborate. If you have built elaborate scaffolding around the 64K ceiling, that scaffolding becomes the migration cost.

Gemini 3.1 Pro Preview keeps one specification advantage that Argon's announcement does not address: audio and video in. Argon's published modality list covers text, images, video and files — audio is not named in the launch material, and Google's video example is LVBench, a long-video-understanding benchmark. If your pipeline ingests audio, verify the modality before assuming the upgrade carries it. That is a checkable fact about a model that is not yet callable, which is exactly the category of claim that ages badly.

The Preview stuck in the way

There is a genuine, boring, expensive problem here: the older model is called a preview, it has been called that for seven months, and nobody has said what happens to it. Two scenarios are plausible and they have opposite implications for your codebase. Either Gemini 3.1 Pro Preview is retired when Argon reaches general availability and you must migrate, or it is superseded and slowly deprecates while still serving, and you have time.

Google's own materials do not settle it. The Argon announcement says broader access comes "as soon as possible", starting with paid API customers and Google AI Ultra subscribers, and says nothing about the fate of the Pro tier it replaces. The conservative engineering position is the one that costs nothing: keep the model id in configuration rather than in code, keep an evaluation harness that can run your real prompts against a candidate model in an afternoon, and treat the eventual swap as an operational task rather than a rewrite.

That position is cheap on OrcaRouter specifically, because the swap is a string. google/gemini-3.1-pro-preview is live in the catalogue at the rates above, behind the same OpenAI-compatible endpoint as the other 200-plus models we route, at provider list price with zero markup — so when Google moves 3.1 Pro's price or Argon gets published, the rate on our side moves the same day. Running your prompts against both is a config change and a second call, not a second integration, and automatic failover means an import that goes stale does not take the endpoint down with it.

What to do this month

If you are on Gemini 3.1 Pro Preview and your workload is retrieval over long context, GPQA-shaped scientific reasoning, or anything where the scientific-reasoning accuracy at 94.1% is load-bearing, there is no case for moving yet — Argon is not callable, and the specific task families where the old model leads have not been demonstrated to have closed. Do not migrate on the strength of an index gap alone.

If your workload is agentic coding, multi-step business task execution, or anything that generates large outputs, the case is already strong on paper: sixteen times the output ceiling, a 53-point Terminal-Bench spread, and a lower output price. The blocker is purely that Argon is a phased rollout to trusted cyber defenders through Google's Fairwind Program with no public endpoint. Put the migration on your roadmap, build the evaluation harness now while there is no pressure, and run it the week Argon becomes callable. The successor is cheaper; being ready for it costs almost nothing.