Mercury 3 vs Gemini 4 Argon: Both Are on Sale, Only One Has Said When the Sale Ends

G​oogle has told the market exactly when G​emini 4 Argon gets more expensive: on December 1, 2026, input goes from $2.00 to $4.00 per million tokens and output from $10.00 to $20.00. Inception Labs has told the market that Mercury 3 is 50% off "at launch" and has published no end date at all. Two introductory prices, one deadline — and that asymmetry is more consequential for a budget than the current numbers, because the current numbers are the ones that will be wrong soonest. On October 20 the pair reads $0.35 against $10.00 per million output tokens. On December 2, assuming Inception's promotion has also been wound down, it reads $0.70 against $20.00. The ratio barely moves; what moves is what you will actually be paying. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

The models underneath the discounts are also doing different things. Mercury 3, released October 19, 2026, is Inception's diffusion language model: text in and out, 1M context, a 131,072-token output ceiling, 1,068 tokens per second on NVIDIA hardware by Inception's own count, and no numeric benchmark scores in its announcement. G​emini 4 Argon reached general availability on G​oogle's paid tier on October 13, 2026, after being announced on September 30; it is G​oogle's flagship reasoning model, with a unified 1M-token context and long decodes of up to 1M tokens, and it carries an independent Artificial Analysis Intelligence Index of 52.6. Put simply: one is the fastest cheap text model you can call this week, the other is a flagship with a deadline on its price.

Sixty days of overlap, then a different conversation

Through November 30, both models are in a promotional window that closes at different times and for different reasons. G​oogle's is a dated introductory rate with a published step-up; Inception's is a launch discount with no stated horizon. That gives a buyer an unusual opportunity — a sixty-day period in which the two cheapest credible options in their respective tiers are simultaneously discounted — and an unusual trap, which is modelling a 2027 budget on a 2026 promotion.

The Argon increase is a doubling on both sides of the rate card, and it applies to cached input too: $0.10 per million through November, $0.20 after. Context caching on Argon is implicit only — explicit caching is not supported — so the cached rate is what you get when G​oogle detects a repeated prefix, not something you configure. For a long system prompt that is useful; for a workload built around explicit cache control it is a limitation rather than a discount.

Mercury 3's cached input follows the standard 90%-off-input rule: $0.02 per million at list, $0.01 during the launch window. No expiry is published. A prudent reading is that the list price is the rate to plan on and anything below it is upside.

The spec sheet, with the caveats in place

• Output price — Mercury 3 $0.35 at launch, $0.70 at list. G​emini 4 Argon $10.00 through November 30, $20.00 from December 1.

• Input price — Mercury 3 $0.10 at launch, $0.20 at list. G​emini 4 Argon $2.00 through November 30, $4.00 from December 1.

• Output ceiling — G​emini 4 Argon allows 65,536 tokens by default on synchronous calls and up to 1,048,576 with its long-decode mode. Mercury 3 allows 131,072. The larger Argon number comes with a catch: its 1M is a unified input-plus-output budget, so a 900K-token prompt leaves only about 100K for the answer.

• Context — both 1M tokens.

• Inputs — G​emini 4 Argon is multimodal; G​oogle's documentation lists image, video, audio and document input alongside text. Mercury 3 is text in, text out.

• Throughput — Mercury 3 1,068 tok/s, Inception's figure, unreproduced. G​emini 4 Argon's output speed is not something G​oogle markets; Artificial Analysis's measurement of the model is the number to use, and it is an order of magnitude below Mercury's claim.

• Independent score — G​emini 4 Argon 52.6 on Artificial Analysis's Intelligence Index v4.3.2 at high effort, at $1.99 of compute per index task. Mercury 3 has no Artificial Analysis page at all as of today.

• Vendor task scores — G​oogle publishes a benchmark table for Argon, including Terminal-Bench 4.0 at 57.1%, SciCode at 61.8% and Humanity's Last Exam at 57.1%; these are G​oogle-reported and G​oogle's competitor columns are the competitors' own self-reported numbers. Inception publishes no numeric scores for Mercury 3.

• Availability — Argon is GA on G​oogle's paid G​emini API tier under the model string gemini-4-argon, and explicitly not available on the free tier. Mercury 3 is on Inception's platform, with 100 million free tokens for new accounts, plus enterprise deployment options.

• Access friction — Argon requires a G​oogle Cloud billing relationship and, for anything sensitive, a data-residency review. Mercury 3 requires an Inception account and an OpenAI-compatible client.

What 52.6 buys that "outperforms Haiku 5.5" does not

The evidence asymmetry here is stark enough to be the deciding factor for many teams. Artificial Analysis ran G​emini 4 Argon on its own harness and published a number that anyone can compare against 200 other models. G​oogle, separately, published a task-level table with named benchmarks. Neither of those is a guarantee, but together they let a buyer reason about where Argon sits relative to models they already use.

Mercury 3 offers a comparative sentence — it "outperforms GPT-6 Luna and Claude Haiku 5.5" — with no margin, no benchmark names, and no scores, plus one agent result on 50 live Mind2Web tasks where it completed half and beat GPT-6.1 Sol. That is a genuine signal about web navigation and nothing at all about coding, math, long-context recall or instruction following.

There is a fair reading of this and an unfair one. The fair reading is that Inception shipped a model one day ago and independent labs have not caught up; the previous Mercury, 2.5, does have an Artificial Analysis page and reads at 705.9 tokens per second, so the vendor's speed numbers have historically been in the right neighbourhood under neutral measurement. The unfair reading is to treat "outperforms Haiku 5.5" as though it were a score. It is a positioning statement about the speed tier, and Argon is not in that tier.

The December 1 arithmetic, run properly

Suppose a product team settles on 40 million output tokens and 200 million input tokens a day — a large assistant, or a batch pipeline with generous prompts. On G​emini 4 Argon at the introductory rate that is $400 of input and $400 of output, or $800 a day. From December 1 the same traffic is $800 and $800, or $1,600 a day. The step-up is $800 a day, or roughly $24,000 a month, for identical work. Nobody sends a memo about that; it simply appears on the invoice.

On Mercury 3 at the launch rate the same traffic is $20 of input and $14 of output, or $34 a day. At list price it is $40 and $28, or $68 a day. Even after both discounts end, Mercury 3 is about 24 times cheaper on this workload — and the absolute gap has grown from $766 a day to $1,532 a day, because a ratio applied to a bigger number is a bigger number.

What that money buys on the Argon side is the ability to emit up to a million tokens in one response, to read a video, and to score 52.6 on a public index. For a workload that summarises a two-hour recording, or writes a 200-page structured document in a single pass, that is not a premium — it is the only way to do the job. For a workload that emits four hundred tokens per call, none of it is being used.

Where a router helps, and where it does not

Neither model is on our catalogue today. Our model API has no G​emini 4 entry and no Inception entry — google/gemini-4-argon returns model-not-found, and so does every Mercury model. That is worth stating without decoration, because a price comparison between two models we cannot serve is only useful if you know that going in. Argon runs on G​oogle's own paid tier; Mercury 3 runs on Inception's platform and the third-party hosts Inception lists.

What we can do is cover the rest of the stack around them. The Google models we do route — gemini-3.8-flash, gemini-3.6-flash, gemini-3.5-flash, gemini-3.1-pro-preview and their lite variants — sit on the same OpenAI-compatible endpoint as more than 200 other models, at provider list price with 0% markup, which means a G​oogle price change lands on our side the same day rather than at the next contract renewal. That property is about to be tested in public on December 1, and it is the reason to keep the routing layer between your code and the vendor rather than hard-coding either rate card.

For the two models in this piece, the useful pattern is a capability split with an explicit cost ceiling: Argon for the requests that genuinely need a long decode or multimodal input, a routed model for everything else, and the routing DSL holding the rule so the split is enforced rather than remembered. Model fusion is worth a look for the hard cases — running a panel and taking the consensus is cheaper than defaulting the whole workload to a flagship.

The decision, dated

Before December 1, G​emini 4 Argon at $2/$10 is a reasonable flagship price and the case for it rests on capabilities Mercury 3 does not have. After December 1, at $4/$20, it is a premium product at a premium rate, and the case has to be made on the same terms as Claude Opus 5.5 — which charges the same $4/$20, carries a slightly higher independent index score at 57.6, and caps output at 128K tokens where Argon offers up to 1M. The long decode is Argon's defensible premium; at the higher rate it is the only thing that is.

Mercury 3 is the opposite trade. It is extraordinarily cheap, fast in a way that matters for interactive products, and almost entirely unmeasured. Its own launch discount is the least durable thing about it, because a vendor that has not published an end date can end it whenever the growth numbers suggest it should.

Plan on list prices. Use the promotional window to run the evaluation you would otherwise put off until Q1, and decide in November what you are willing to pay in December. The deadline on Argon's rate card is the most useful piece of information in this comparison, and it is the one that expires soonest.