GPT-6.1 Sol vs Gemini 4 Pro: Google Never Shipped That Name, and the Price Match Ends December 1

There is no model called Gemini 4 Pro. If you searched for a comparison against GPT-6.1 Sol, the Google side of that matchup is Gemini 4 Argon — the frontier model Google announced on September 30, 2026 to a closed cohort of cyber defenders, and opened to paid developers on October 13. The two now carry what looks like an identical card: $2.00 per million input tokens and $10.00 per million output. That identity has an expiry date. Argon's introductory rate runs through November 30; from December 1 it becomes $4.00 and $20.00. GPT-6.1 Sol, released at DevDay on September 29, is priced at $2.00 and $10.00 as its standing rate, not a promotion. So the real question is not which model wins today. It is which card you are still holding in February.

Everything below comes from the two vendors' own model and pricing documentation, read on October 27, 2026, or from Artificial Analysis, which is the only independent board where both models appear. Where a number is a vendor's own benchmark run, the sentence carrying it says so. Google's model list is the reason the naming correction is not pedantry: the identifier it serves is gemini-4-argon, and there is no gemini-4-pro row to call.

The name in the search box, and the model on the invoice

"Gemini 4 Pro" survived as a community label from the leak cycle. In mid-September, reports described a mystery model appearing anonymously on arena-style leaderboards, producing unusually long reasoning traces, and being widely identified as the next Pro-tier Gemini. When Google actually announced the model, it called it Argon and described it as the start of the Gemini 4 family rather than a replacement for a named Pro tier. Reports at the time said Google had cancelled the Gemini 3.5 Pro line and redirected that compute, which is part of why the Pro label stuck: there was a gap where a Pro model would have been.

The label matters for a practical reason. Searching for Gemini 4 Pro mostly returns pages built on leak specs — a ten-million-token context, permanent cross-session memory, sandbox isolation — none of which Google has published. Searching for Argon returns the vendor's own documentation. If you are choosing a model rather than reading a rumour, the second is the only useful object. Argon's public documentation describes a unified one-million-token workspace shared across input, thinking and output, a default output allowance of 65,536 tokens, and long-decode runs that can extend to roughly a million output tokens.

Two rate cards, one of which is a promotion

Here is the whole commercial comparison in the dimensions that change a bill. Every line is from the vendors' own published terms on October 27, 2026.

• Standard rate today — GPT-6.1 Sol $2.00 input / $10.00 output per million; Gemini 4 Argon $2.00 / $10.00.

• Rate on December 1, 2026 — GPT-6.1 Sol unchanged at $2.00 / $10.00; Gemini 4 Argon $4.00 / $20.00.

• Cached input — Sol $0.10 per million; Argon about $0.10 today at a 95% discount, reported to rise to roughly $0.20 after the step.

• Long-context pricing — Sol reprices the entire request above 272,000 input tokens, to $4.00 input and $15.00 output with cached input at $0.20; Argon publishes no equivalent step because its workspace is a flat million tokens.

• Context — Sol 1,050,000 input tokens with a 128,000-token output cap; Argon one million tokens unified, with a 65,536-token default output.

• Reasoning control — Sol is always reasoning, with five rungs from low to max and no off switch; Argon's effort control is exposed through Google's API but not documented in the same ladder terms.

• Input — Sol accepts text, images and files; Google positions Argon as multimodal across text, image and video.

• How you buy it — Sol through OpenAI's API, Bedrock, and routers that carry it, including ours; Argon through Google's own API, AI Studio and enterprise channels.

That fourth line is the one people miss. Sol's long-context step is a cliff attached to the length of a single request: cross 272,000 input tokens and the whole prompt reprices, not just the excess. Argon's flat million-token workspace has no equivalent boundary. For an agent that assembles a large repository context on every turn, that difference can matter more than the headline rate — and it cuts in Argon's favour, at least until December.

Fifty-three against fifty-two, and the honesty of a one-point gap

Artificial Analysis puts Gemini 4 Argon at 53 on its Intelligence Index and GPT-6.1 Sol at 52, both on version 4.3.2 of the suite. Argon ranks ninth of 221 models; Sol sits just behind it. One point on a composite of roughly ten evaluations is not a ranking, it is a band. What the index is better at is telling you that neither model is buying its score with a dramatically different cost structure — at least not on the headline rates, where they are the same number.

The configurations are not identical, and that is worth stating rather than hiding. Artificial Analysis charts Sol at max effort, its most expensive setting. Argon's score is published at a single configuration Google exposes. A comparison across different effort settings is a comparison of two products as shipped, not of two capabilities at their ceilings, and no one has published a matched-effort run of both.

Vendor numbers exist on both sides and should be read as claims. Google's launch material reported first-place finishes on 13 of 19 evaluations it ran, and described Argon as its strongest model for difficult scientific and engineering work. OpenAI's launch material for GPT-6.1 Sol reported parity with its own flagship GPT-6 Astra on DeepSWE v1.1 at roughly one-fifth the cost, a seven-point gain on OSWorld 2.0 offline, and a drop in factual error rate at low reasoning from 11.4% to 7.7%. None of those four claims has a published independent reproduction. They are useful for understanding what each vendor thinks it built, and useless as a tiebreaker between the two.

The arithmetic of the thirty-five days between now and December 1

Take a mid-sized production workload: 40 million input tokens and 8 million output tokens a month, no caching, no long-context crossings. On either card today that is $160 a month — $80 of input at $2.00, $80 of output at $10.00. From December 1, the same workload on Argon is $320: $160 of input at $4.00 and $160 of output at $20.00. On GPT-6.1 Sol it stays at $160.

Over the twelve months after the change, that is $3,840 on Argon against $1,920 on Sol, a difference of $1,920 for a workload that did not change. If your traffic pattern includes prompts above 272,000 tokens, Sol's cliff complicates the picture: a single request at 300,000 input tokens bills at the long-context rate on the full 300,000, which is $1.20 rather than $0.60. Run forty such turns a month and the saving above shrinks by roughly $240 a year. It does not reverse. The cliff is a reason to measure your prompt lengths, not a reason to prefer a card that doubles.

What the comparison cannot tell you

Neither model has a practitioner corpus. Argon has been generally available for two weeks; Sol for four. There is no body of published failure modes, no long-running agent telemetry, and no independent cost study of either on a real production trace. The index measures a suite, not your workload. The vendor benchmarks measure the vendor's chosen tasks, at the vendor's chosen effort settings.

The one asymmetry that is not about capability is availability. Argon is sold by Google, on Google's terms, with a price change already scheduled. Sol is sold by OpenAI and carried by routers, which is a different kind of dependency: if a single provider has an incident, a routed caller can fail over; a direct caller cannot. That is not a reason to pick a model. It is a reason to keep the model choice separate from the transport, which is the argument for a gateway in general and not for either model in particular.

Which one to call, given the calendar

If you are choosing today and your workload is price-sensitive, GPT-6.1 Sol is the safer card because it is a commitment: $2.00 and $10.00 with a published long-context rule, no expiry, and a cached-input meter at a tenth of the input rate. It is on our catalogue at OpenAI's list price with 0% markup, so the rate you see on the vendor's page is the rate you are billed, and a change on OpenAI's side would be reflected the same day rather than after a repricing cycle.

If you need what Argon specifically does — the flat million-token workspace, Google's multimodal input, or its placement on Google's own scientific evaluations — then the honest advice is to budget the December number now. A team that plans against $2.00 and $10.00 and renews in January will experience the change as a 100% increase. A team that plans against $4.00 and $20.00 and gets thirty-five days of discount is simply early. Gemini 4 Argon is not in our catalogue, so if you want it, the vendor's own API and several third-party platforms are where it lives.

The name in the search box was never real. The model is, the discount is real for another month, and the only decision that survives December is whether the workload needs Argon's shape badly enough to pay twice for it. For most coding and document-heavy work, the answer this month and next is the model whose price is not a countdown.