Google's Argon: strong for business agents, weaker on terminal tasks, price doubles Dec 1

Google's Argon model is worth a test for business agents, such as finance research and workflow automation, and is a weaker bet for terminal tasks. Outside testers put it first on Vals's Finance Agent v2 and on Artificial Analysis's AutomationBench-AA, behind Claude Opus 5.5 and Sonnet 5.5 on Terminal-Bench 4.0, and a little behind GPT-6 Astra. Its token price doubles on December 1, when Google's introductory price ends, so the number to budget on is what a task costs you, not the token price, and no source we read has measured that at Google's default thinking level. Google announced Gemini 4 Argon on September 30, 2026 and opened it on October 13 to the paid tier of the Gemini API and to Google AI Ultra and Pro subscribers in Google's own products.

This is research from public sources, not a test of ours. EP has run nothing on Argon. We re-read every source below on October 13, 2026, Central time. Leaderboards and documentation pages that show no date of their own are cited as of that day. A number Google ran itself is labelled Google's own, because it shows what the vendor claims, not what an outside tester found.

#What Argon is, and when it arrived

Argon (model name gemini-4-argon) is the first model in Google's Gemini 4 family, per Google's own guide, and Google says "Gemini 4 Argon is just the beginning" in its October 13 launch post. Its model card, published October 13, says it "is not a modification or a fine-tune of a prior model" and is a "large scale sparse mixture-of-experts."

The dates matter, because the first outside numbers came before the public release.

• September 30, 2026. Google's announcement says Argon is "rolling out to a set of trusted cyber defenders." The same day, Artificial Analysis published its first results and wrote that Argon "is currently being rolled out to selected users and is not publicly available". Neither Artificial Analysis's nor Vals's pages that we read say how they got access.
• October 13, 2026. Google's launch post says Argon "is now globally available today for developers and enterprises via the API." In practice that is the paid tier of the Gemini API. Google's pricing page lists the free tier as "Not available" for this model, and its Gemini API page says Argon "is not available on the Free tier or on the" Google AI Pro plan. The same launch post gives Google AI Ultra subscribers Google Antigravity and AI Studio, and says the Gemini app and AI Mode in Search are open to "Google AI Ultra or Pro" subscribers.

Two limits are worth knowing early. Google's September 30 announcement describes an "industry-leading 1M tokens" output limit. The October 13 launch post calls it a "1M unified input-output token limit," and Google's Vertex guide says that 1,048,576-token limit "includes both input and output," with up to 65,536 output tokens per request by default. And the Gemini API model page lists the Live API as not supported, so Argon cannot be used through the Live API. We did not test whether a separate speech setup can call it as a reasoning step.

#Where it leads and where it trails

The answer depends on the board and on who ran it. Every row below names the tester. Rows marked as printed by Google come from Google's model card, and we did not open the tester's own page for them. Google's table sets Argon against only GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Vals's boards also include Claude Sonnet 5.5 and GPT-6.1 Sol, so the comparison set in Google's table is Google's choice.

Business, knowledge-work and general boards

Four notes on that table. First, the Vals Index is not a pure business measure. Vals's formula weights finance 8.0, coding 5.6, legal 1.2 and tax 0.5, so coding results count toward Argon's first place, and the gap to Sonnet 5.5 is 1.9 points with margins of about one point (68.90% plus or minus 0.97 against 67.04% plus or minus 0.92). Vals adds that the Claude scores include some tasks served by a fallback model after a refusal. Counting those as failures lowers Sonnet 5.5 to 65.85% and Opus 5.5 to 65.05%, "with the order unchanged," which would make Argon's lead 3.05 points. Vals says Argon had no fallbacks. Artificial Analysis labels its Claude rows "with fallback" on its leaderboard too, and we did not find its figures without fallback.

Second, the Finance Agent row is Vals's primary Partial Credit score, and the Vals page says the order "flips under stricter All-Pass scoring," where Muse Spark 1.2 leads at 50.88%. We could not read Argon's All-Pass figure from the page text, and that takeaway looks out of date, since it says only two models clear 60% on Partial Credit while the same page's board shows three. So the All-Pass order for Argon is unconfirmed. Third, the legal row needs its context. The five models above Argon on that board are all Meta's, and the best Claude or OpenAI model on it scores 11.25%, so among the labs a buyer is likely to be choosing between, Argon is ahead. Fourth, Artificial Analysis's leaderboard page counts 127 models while its model page counts 216 entries, so the two lists count differently. We use the model page's rank only.

Argon is not first everywhere on Vals's business boards. Vals's Argon model page ranks it fourth on EMB (75.24%), Legal Research Bench (54.81%) and Tax Agent Bench (44.56%). On the Vals Index page, the leaders score 76.67% on EMB (Claude Fable 5.1), 55.29% on Legal Research Bench (a three-way tie) and 77.64% on Tax Agent Bench (Fable 5.1).

Terminal and coding boards

Argon lands at 57 or 58 on Terminal-Bench on all three runs, and each puts Opus 5.5 ahead. Three cautions on that. First, the Vals lead is smaller than it looks. Vals says 22 of Opus 5.5's 198 task attempts there were served by a fallback model, and that counting those as failures lowers its score from 65.15% to 58.08%, against Argon's 57.58%. Sonnet 5.5 (62.63%) and Astra (59.60%) stay ahead on the same count. Second, the Astra gap is inside the noise on that board: Vals prints Argon at 57.58% plus or minus 2.31 and Astra at 59.60% plus or minus 4.40. Third, Google's row sets its own run of Argon beside rival scores that, by its methodology page, come from the official public leaderboard, so it is not like for like, and we did not check whether Google's 66.4% figure for Opus 5.5 has the same fallback issue.

Terminal-Bench 4.0 is also wider than the name suggests. Vals files it under coding, but its page says roughly three quarters of the benchmark is outside traditional software engineering, and that Argon leads two of its categories, Operations (55.56%, 9 tasks) and Media (83.33%, 4 tasks). Astra leads the Software category (61.11%) and Opus 5.5 leads Science (71.43%). Google's own Terminal-Bench Science run also has Argon behind Astra and Opus 5.5 (57.6% against 68.1% and 63.3%). Vals lists the same Mini-SWE-agent harness for every model compared here. That is why the title says terminal tasks and not coding. Vals's model page ranks Argon sixth of 48 there, so "weaker" here means weaker than the leaders, not weak. Vals also calls automated code migration "far from turnkey" and says even the leader's ports "mostly work but aren't drop-in."

The coding row that makes the best headline is the least checkable one. Google's methodology page says Argon's DeepSWE score is "self computed," with Astra taken from the public leaderboard and the Claude scores taken from "their respective system cards." The benchmark owner's DeepSWE board, updated September 22, 2026, shows no Argon row in the view we read. That board predates the September 30 announcement, so its silence is expected. It does mean that board gives no outside confirmation of 77.9% yet.

Computer-use results are mixed. Vals's Argon model page puts it at 4.83% on CUA-bench, eighth of nine on Vals's ranking, at $193.78 per test. Google's model card has OSWorld-2.0 at 69.2% against Astra's 72.6%, and Agent's Last Exam at 39.5% against Opus 5.5's 38.2% and Astra's 34.2%. Both Argon scores are Google's own runs, and the OSWorld score is the best of three.

Two more readings. On Arena's text leaderboard, a ranking built from human votes, Argon (high) is first at 1525 but marked "Preliminary," with Meta's muse-wonder-max at 1513 and Claude Opus 5.5 (high) at 1508, all as of October 13, 2026. The margins shown, plus or minus 9 and plus or minus 13, overlap. And a low hallucination rate is not the same as being right. Artificial Analysis found a 15% hallucination rate on its AA-Omniscience test, against 51% for GPT-6 Astra (max), but only 50% accuracy, against Astra's 63%, and an overall AA-Omniscience score of 42 that Artificial Analysis calls "in line with GPT-6 Astra (43)" (September 30, 2026). On its AA-Briefcase knowledge-work test, Argon reached 1494 Elo on the highest rubric pass rate Artificial Analysis has recorded, 65%, but with lower Analytical Quality (1576 Elo) and Presentation Quality (1308 Elo).

#Cost per task is not the token price

Google's pricing page, read October 13, 2026, lists Argon's standard rate at $2.00 per million input tokens and $10.00 per million output tokens through November 30, then $4.00 and $20.00 from December 1. Google's Gemini API page calls $2 and $10 an "introductory price." So the December 1 price is the standard one, and the current price is the promotion. The output price includes thinking tokens. Cached input costs $0.10 per million, rising to $0.20 on December 1, and only implicit caching is supported. Vals lists Opus 5.5 at $4 and $20, Sonnet 5.5 at $2 and $10, and GPT-6 Astra at $10 and $50. From December 1, Argon's listed token price equals Opus 5.5's and is double Sonnet 5.5's.

A token price does not tell you what a task costs, because models use different numbers of tokens. Artificial Analysis measured $1.99 per task on its Intelligence Index for Argon at the introductory price, against $3.26 for GPT-6 Astra (max). It says the saving "is driven by lower token prices, rather than reduced token use": Argon averaged 62k output tokens per task against 27k for Astra. Artificial Analysis then projects $3.98 per task at the full price, about 1.2 times Astra. That figure is exactly double the measured $1.99, so it is arithmetic and not a second run.

The same leaderboard, as of October 13, 2026, shows cheaper runs at similar scores. Argon scored 53 at $1.99. Claude Opus 5.5 at high setting, labelled "with fallback," scored 54 at $1.82. GPT-6.1 Sol (max) scored 52 at $0.72. Artificial Analysis scored Argon at its high setting only.

Vals reports a cost per test, and its Argon model page says Argon "is priced at $4.00 per million input tokens and $20.00 per million output tokens," so we read these figures as the December 1 rate, though Vals does not say so in those words. The same page says the evaluations ran at reasoning effort "high." Neither figure is the cost of a finished task on your work: Artificial Analysis reports cost per task in its own index, Vals reports cost per test, and neither says it is adjusted for how often the model succeeds. The two are not comparable with each other. The answer flips with the job:

On the broad index and on finance, Argon costs less per test than Opus 5.5, Sonnet 5.5 and Astra, though Sol costs far less and scores lower (61.15% on the index, 52.03% on finance). On Terminal-Bench 4.0 Argon costs more than those three and scores lower than all three. So whether it is cheaper has no single answer. It changes with the job, the thinking setting and the date.

#Five checks before switching

1. Price your own task at the December 1 rate. Take one real job, count the billed tokens including thinking, and multiply at $4 and $20, not at the $2 and $10 that ends November 30. Put your current model's cost for the same job beside it, and count how often each one finishes the job.
2. Test at the thinking level you will run. Google says its results use "the highest thinking settings", Artificial Analysis scored Argon at high, "the highest available," and Vals says it ran Argon at reasoning effort "high." Google's Gemini API page lists medium as the default thinking level. We found no page that reports a score or a cost at medium.
3. Match the board to your job. Argon's best evidence is on business-agent boards, and its weakest is on terminal tasks. Pick the board nearest your work, then run a small version of that work.
4. Run your tool loop unchanged, in a test copy with no live write access, and count the failures. Google's Gemini API page says sampling settings such as temperature are ignored, and the model page says explicit caching is not supported. Google's Vertex guide adds that thinking_budget and MINIMAL return an HTTP 400 error and that prefilled model turns must be removed. We did not confirm that those last two apply to the Gemini API. The Vertex guide also lists "Agents that call tools in many short turns" as a poor fit for long decode, and the Gemini API page says to set HTTP client timeouts to "at least 15 minutes" because hard prompts can take several minutes.
5. Read the access and data terms. Argon is on the paid tier only, with no free tier. Google's launch post names Ultra subscribers for Antigravity. There is no Live API. Google's pricing page says content on the paid tier is not used to improve its products.

#What we have not tested

EP has run nothing on Argon. Everything above is reading, not testing. We have no cost per task of our own, at either price, and no view of how Argon behaves on EP's work.

Some limits on the sources. We did not open Zapier's board, so the AutomationBench row rests on Google's table. We did not review Google's safety report. We read each page once on October 13, and the boards can move. Vals and Artificial Analysis scored Argon from September 30, before the public release, and we found no page that says whether the public model is the same build.

What we would run first, as a plan and not a finding:

• One real long coding job, same repo and prompt, against our current default, logged as billed tokens and dollars at $4 and $20, with reviewer-accepted and reviewer-rejected changes counted.
• A check for stubs and false completion on that run: empty functions, TODO markers, and claimed test passes against actual ones.
• Our existing tool-calling loop, unchanged, counting API errors.
• One long-context retrieval check at 256k to 1M tokens on a real document set. Google's lead there, 84.2% against Astra's 71.8% on GraphWalks, is self-computed for all models.

Related field notes: Claude Opus 5.5 sells the same work for less. Its benchmark table runs at a setting you may not use. covers the same effort-setting question for a rival, and What OpenAI's own system card says about GPT-6 Astra, without the word neuralese covers the model Argon is most often set against.

The $2 and $10 rate is Google's introductory price. The task price is the one your budget will carry.

#Sources

• Google, Gemini 4 Argon: our next era of frontier intelligence (September 30, 2026)
• Google, Gemini 4 Argon: Built to solve your hardest problems (October 13, 2026)
• Google DeepMind, Gemini 4 Argon model card (published October 13, 2026)
• Google DeepMind, Gemini 4 Argon evaluation methodology (read October 13, 2026)
• Google, Gemini API pricing (read October 13, 2026)
• Google, Gemini API latest model page (read October 13, 2026)
• Google, Gemini API model page for gemini-4-argon (read October 13, 2026)
• Google Cloud, Gemini 4 Argon guide (updated October 13, 2026)
• Artificial Analysis, Gemini 4 Argon: Google is back as one of the top three labs in intelligence achieved (September 30, 2026)
• Artificial Analysis, Gemini 4 Argon model page (read October 13, 2026)
• Artificial Analysis, model leaderboard (read October 13, 2026)
• Vals AI, Vals Index (read October 13, 2026)
• Vals AI, Gemini 4 Argon model page (read October 13, 2026)
• Vals AI, Terminal-Bench 4.0 (read October 13, 2026)
• Vals AI, Harvey's Legal Agent Benchmark (read October 13, 2026)
• Vals AI, Code Migration (read October 13, 2026)
• Arena, text leaderboard (read October 13, 2026)
• Datacurve, DeepSWE board (updated September 22, 2026)