Gemini 4 Argon vs Claude Opus 5.5 vs GPT-6.1 Sol: Same-Task Test

For API work, and coding agents in particular, Claude Opus 5.5 (claude-opus-5-5) is the better default over Gemini 4 Argon (gemini-4-argon) once you count cost per finished task instead of price per token. The same eight coding and reasoning tasks went to both models through the LaoZhang gateway, Argon on November 4 and Opus 5.5 on November 11, 2026. At the highest setting each was given (high thinking level for Argon, high effort for Opus), both got all 11 graded answers right, and Opus cost $0.31 against Argon's $0.40, because Argon wrote 3.3 times as many output tokens. At their defaults Argon was the cheapest run but missed one answer, and on exact counting over a 146 KB log it never got all seven answers right until it ran at high with streaming.

Argon's lower price is per token, and it ends on November 30, 2026. From December 1 Argon costs $4 input and $20 output per million tokens, exactly what Opus 5.5 charges, and the same high run becomes $0.81. On Artificial Analysis' index Argon has not been cheaper per task than Opus at high effort even at the introductory rate. As of November 11 it shows Argon High at $9.95 per task at the December rate, up from $7.68 a week earlier with identical scores, against $3.46 for Opus 5.5 High, which scores two points more.

So run Opus 5.5 at medium or high for coding agents and for exact work over long input. Keep Argon for the knowledge work, legal, finance and long-context retrieval where it leads Google's own table, call it at high with streaming, and judge it at $4 / $20. GPT-6.1 Sol (gpt-6.1-sol) remains the budget reference at $0.59 per index task. Before December 1, send a sample of your own tasks to both models and record tokens and pass rate per task, because that, not the rate card, decides the bill.

Same tasks, one gateway: Opus 5.5 high costs less than Argon high

On the same prompts, Opus 5.5 at high effort finished eight small tasks for 24% less than Argon at high thinking through November 30, and 62% less from December 1, with the same 11 of 11 answers correct. Argon ran on November 4, 2026 through the Gemini-native generateContent endpoint without streaming, once at its default (no thinking setting, which Google documents as medium) and once at thinkingLevel high. Opus 5.5 ran on November 11, 2026 through the Anthropic Messages endpoint with streaming, adaptive thinking and effort set to medium (the Claude API default) or high. Both went through the LaoZhang gateway (api.laozhang.ai), and every cost below is the returned token count priced at Google's and Anthropic's list rates.

Disclosure: laozhang.ai is our own service. Its prices are platform pricing, not an independent comparison.

The tasks were a word problem; three small Python functions graded by hidden tests (merging intervals with 8 tests, parsing durations with 18, counting top words with 10); a summary of a 6,723-word article in exactly five bullets; and three harder tasks run twice each: an expression evaluator without eval (25 hidden tests), a CSV line parser without the csv module (16) and a digit-sum count with one exact answer (55,252). That makes 11 graded answers per setting.

At the defaults the cost runs the other way. Argon was the cheapest setting, and Opus at medium cost 3.1 times as much now and 1.55 times as much from December 1, mostly because it wrote 1.63 times the output tokens. Argon's one miss was the expression evaluator: its first answer raised a ValueError on " 1 + 2 ", an input with extra spaces, and failed 1 of 25 hidden tests. A second Argon default run the same day passed all 11, so 21 of 22 Argon default answers were right. Moving Argon to high removed the miss but raised its output almost sixfold, 33,105 of the 39,514 tokens being thinking, which is why its cost passes Opus's even at today's price.

Opus 5.5 had one formatting quirk. At high it returned the merge-intervals function and its own tests as two files in one reply; with the files separated, the function passed all 8 hidden tests.

These are small, well-specified tasks, and most of them do not separate the two models. Each cell rests on one or two samples, the two models were measured a week apart on one gateway, Argon ran without streaming and Opus with it (which shifts wall time, not tokens or cost), and Argon's thinking levels and Opus's effort levels are not the same scale. Read the table as what each setting spends on the same work, not as a quality ranking; the benchmark sections below are that reference.

Exact counts over a 146 KB log: Argon needs high and streaming

On a long-input task with one right answer, Opus 5.5 was right every time at its default, and Argon was right only at high with streaming. The input was a synthetic application log of 146 KB (1,500 lines, 326 of them ERROR lines) with seven questions: the ERROR count for five error codes, how many distinct users hit an E503 error, and which request had the highest retry value. The script that generated the log also wrote the answer key, and a grep over the file gives the same answers.

Opus 5.5 got all seven right in 4 of 4 runs on November 11, 2026 (two at medium, two at high), in 56–59 seconds, using 78,537 input and 7,699–8,255 output tokens, for $0.47–0.48 a run before and after December 1. Argon at its default, on November 4, answered in about seven seconds with 85 output tokens and no thinking, and both runs were wrong: the five error counts came back 20–37% too low (59 E503 errors instead of 94), and the E503 user count was 46 instead of 78. Each of those runs cost $0.16, or $0.32 from December 1.

Argon at high without streaming ran into the 32,768-token output cap while it was still thinking, returned no answer and was billed $0.49 anyway. With streaming, both high runs were right, in 271 and 402 seconds, at $0.53–0.55 a run now and $1.06–1.09 from December 1. Two more Argon high calls were cut off by a local proxy at about 300 seconds; they returned nothing, cost nothing and are left out. For exact work over long input, then, the Argon setting that works costs slightly more than Opus now, about 2.2–2.3 times as much from December 1, and takes roughly five to seven times as long. This is one task type with two to four samples per setting.

Same token price as Sol, half of Opus, and why that misleads

Argon's rate card matches GPT-6.1 Sol and is half of Opus 5.5 until November 30, but Argon spends far more tokens per task than either, so the rate card is the weaker guide to what you will pay.

List rates for identical tokens

All rates are USD per million tokens on each vendor's standard tier, as of October 14, 2026; Argon's are unchanged as of November 11, 2026, and Opus 5.5's as of November 10.

With U as fresh input, R as cache reads and O as output including reasoning, all in millions of tokens, the bill for one request is:

• Argon through November 30: 2U + 0.10R + 10O
• GPT-6.1 Sol up to 272K input: 2U + 0.10R + 10O, plus 2.5 per million tokens written to cache
• Opus 5.5: 4U + 0.20R + 20O, plus 5 per million tokens written to the 5-minute cache
• Argon from December 1: 4U + 0.20R + 20O

Before cache-write fees, a request with 100K fresh input and 20K output costs $0.40 on Argon through November 30 and on GPT-6.1 Sol, and $0.80 on Opus 5.5 and on Argon from December 1. With 50K fresh input, 150K cache reads and 20K output, the same split is $0.315 against $0.63.

So for the same tokens, Argon equals GPT-6.1 Sol and costs half of Opus 5.5 until November 30. From December 1 it equals Opus 5.5 on every token type. The one difference left is that Argon charges no cache-write fee. It also does not let you create a cache: the Gemini API pricing page lists implicit caching only, so the $0.10 or $0.20 rate applies when Google's cache matches your repeated prompt prefix, and you cannot pin one yourself.

The other tiers move on the same date. Argon's Flex tier is priced like Batch. Priority costs $3.60 input and $18 output through November 30, then $7.20 / $36. Argon has no separate long-prompt rate, while GPT-6.1 Sol bills the whole request at its higher tier once input passes 272K. Grounding with Google Search includes 5,000 free requests a month, shared across Gemini 3 and newer models, then costs $14 per 1,000.

Measured cost per task: Argon $4.98 to $9.95, Opus 5.5 $3.46, Sol $0.59

Real runs never use identical tokens. Artificial Analysis runs each model through its Intelligence Index, now v5.0.1, a set of nine evaluations, and reports what the whole run cost per task. Version 5.0 replaced v4.3.2 between October 27 and 28, 2026, and every score and cost changed, so figures quoted before then (Argon at 53 for $1.99, for example) differ from these. Version 5.0.1 (November 2026) then regraded one of the nine evaluations, CritPt, and of the cells below only Opus at max moved, from 58 to 59. Argon's later change, described under the table, is a separate change that came later, with no change in score. Argon is published at "high" effort only, so high is the one setting where all three can be compared directly.

Each cell is index score / cost per index task on v5.0.1. Argon's row and both max rows are as of November 11, 2026; the other Opus rows are as of November 10 and the other Sol rows as of November 4. Artificial Analysis costs Argon at the $4 / $20 December rate. Every Argon token price is exactly half of that through November 30, so about $4.98 is Artificial Analysis' figure halved for the introductory period, not a number it publishes.

Artificial Analysis' published cost for Argon rose in the first week of November while its score stayed put. On November 4 it showed Argon High at $7.68 per task and 22.7 minutes per task; on November 11 the same score of 50, with all nine evaluation results identical, came with $9.95 and 28.6 minutes, and the output tokens for the whole index went from 110 million to 113 million. The page carries no note explaining the change. Any figure derived from $7.68, including the halved $3.84 introductory number, is out of date.

This chart is outdated: it predates Argon's October 13 launch and still labels Argon as having no model ID, and its cost-per-task bars and scores are Artificial Analysis' old v4.3.2 figures. For current numbers use the v5.0.1 table above (Argon as of November 11, 2026) and the same-task test at the top of this page.

At today's rate, the high row puts Argon at about 8.4 times GPT-6.1 Sol's cost per task ($4.98 ÷ $0.59) for the same score of 50, even though the two charge the same list price. The gap is tokens consumed: Argon produced 113 million output tokens across the index, about 116,000 per task. Against Opus 5.5, Argon costs about 1.44 times as much per task ($4.98 ÷ $3.46) and scores two points lower, although Opus charges twice as much per token. And GPT-6.1 Sol at max effort scores 55 for $1.27, five points above Argon for about a quarter of Argon's introductory cost.

A two-point gap on one composite index is small, and the Vals AI result further down ran the other way between Argon and Opus on its earlier v2.1 index.

What the December 1 price does: $4 / $20 puts Argon near $10 per task

From December 1, every Argon token type costs twice as much: input, output and cached input. Artificial Analysis already prices Argon this way and shows $9.95 per task, twice the roughly $4.98 the same tokens cost at today's rate. Opus 5.5 and GPT-6.1 Sol have no announced price change, so their figures stay put.

At that point Argon at high effort would cost about 2.9 times Opus 5.5 at high ($9.95 ÷ $3.46) and about 17 times GPT-6.1 Sol at high ($9.95 ÷ $0.59). Opus at xhigh would score seven points more than Argon for 39% less, and Opus at max would score nine points more for about the same money. The $9.95 is Artificial Analysis' own figure for its token mix, not a bill: your tasks may make Argon spend more or fewer tokens than the index does.

Equal token prices from December 1 also do not mean equal bills for the same prompt, because the two models count the same text differently. Measured through the same gateway (Argon on November 4, Opus 5.5 on November 11, 2026), an English passage counted 1.45 times as many input tokens on Opus 5.5 as on Argon, and the 6,723-word article from the same-task test 1.38 times as many (15,307 against 11,110). Passages in Chinese, Japanese, Korean, Russian and Spanish came out 1.72 to 2.26 times higher on Opus. The 146 KB log, mostly ASCII, numbers and IDs, counted almost the same on both (0.99 times). From December 1, then, English prose input costs about 1.4 times as much on Opus for the same text, while output depends on how much each model writes, which the same-task test shows can favor either one.

Vals AI already prices Argon at $4 / $20, the December rate, and its result points the other way on Opus. On the Vals Index, Argon at high effort cost $15.68 per test against $32.14 for Opus 5.5 on v2.1; on the current v2.2 (October 30) it is $13.37 against $28.12. According to Kingy AI's notes on the leaderboard, the Opus run used max effort with fallbacks, so the two rows are not at matched effort. Vals also reports that Argon's cost climbs on long agentic tasks, to $193.78 per test on CUA-bench and $57.82 on Code Migration.

The harnesses and the same-task test agree on one point. Argon's bill depends on how many tokens your task makes it spend, and the token price is a poor predictor of it.

Is Argon actually better than Opus 5.5?

Not clearly. At matched high effort the two land within about two points on independent indexes, Argon leads in knowledge work and long-context retrieval, and Opus leads in terminal-heavy coding work.

Google's table: Argon ahead on 14 of 18 rows, Opus on 4

Google's September 30 launch table compares Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Opus has a score on 18 of its rows, and Argon is ahead on 14 of them. Opus leads on four.

The pattern is consistent. Opus leads where an agent works in a terminal for a long time: terminal coding, FrontierSWE, the science terminal suite and ML post-training. Argon's largest margins are in knowledge work, long-context retrieval and video understanding. Some of its other leads are thin, such as CWE-bench v1 at 68.0 against 67.0.

These are vendor-reported numbers, and the table as a whole has not been independently reproduced. Kingy AI's reading of Google's methodology document reports harness differences that affect how far individual rows can be compared: highest thinking settings, a mini-swe harness for DeepSWE, six times the verifier timeout on the science run, and unequal video frame budgets on LVBench. Counting wins also depends on which columns you include. VentureBeat counts 12 outright leads for Argon across all four models, and the count against Opus alone is 14. Google's October 13 post adds new figures of its own, such as 87% on ERQA, but names no competitor scores.

Independent measurement: Argon and Opus 5.5 within about two points

Two third parties have run Argon themselves.

Artificial Analysis puts Opus two points ahead at high effort on its v5.0.1 index, 52 against 50 for Argon; on the older v4.3.2 it was 54 against 53. Its own comparison page for the two models sets Argon at High against Opus 5.5 at max (labeled "Max, Default Fallback"), and as of November 11, 2026 that pairing reads 59 against 50, with the two costing about the same per task at the December price ($10.04 against $9.95) and Opus finishing sooner (24.7 against 28.6 minutes). That comparison is fair if you would pay about $10 per task; at the same high setting the gap is two points.

Vals AI ranked Argon first of 41 models on its Vals Index as of October 1, at 68.90%, ahead of Claude Sonnet 5.5 at 67.04% and Opus 5.5 at 66.97%. Google's October 13 post repeats the first-place claim. The margin over Opus was about two points. That ranking is from Vals Index v2.1. Vals released v2.2 on October 19 (it updated Finance Agent), and as of October 30 its leaderboard puts Argon fourth of 32 at 69.24%, behind Sonnet 5.5 (70.85%), Gemini 4 Carbon (70.45%) and Opus 5.5 (70.18%), with GPT-6.1 Sol eighth at 63.94%. Read November 4, 2026; do not mix v2.1 and v2.2 numbers. Vals also flags a weak spot in computer use, where Argon scored 4.83% on CUA-bench, seventh of eight.

Kingy AI reports that Arena's text leaderboard on September 30 had Argon at high effort first with 1,525 against 1,504 for Opus 5.5 at high. Arena measures which answer people prefer, which is a different thing from whether a task was completed correctly.

Taken together: Argon and Opus 5.5 are within about two points of each other at the same effort. Opus is ahead on Artificial Analysis; on Vals, Argon led on v2.1 and Opus leads by about one point on the current v2.2. Argon is a little ahead on finance, legal and long-context work, and behind on terminal-heavy coding agents by Google's own numbers. Nothing here supports calling either one clearly better across the board.

Argon vs Opus 5.5 for coding: Opus 5.5 is the safer default

For coding agents, Opus 5.5 is the safer default, though the three kinds of evidence agree only in part. Google's launch table, which is vendor-reported, gives Opus the four terminal-heavy rows, including Terminal-bench 4.0 at 66.4 against 57.4 and FrontierSWE v2 at 62.3 against 55.0, and gives Argon two coding rows by narrow margins: DeepSWE v1.1 at 77.9 against 74.2 and Vibe Code Bench at 91.9 against 90.3.

Artificial Analysis' own runs (v5.0.1, as of November 11, 2026) shrink one gap and widen another. On Terminal-Bench 4.0, Argon at High scores 57.1% against 56.6% for Opus at high and 59.6% at max, so at matched effort they are even, not nine points apart. On FrontierCode 2.0, Argon scores 25.7% against 35.9% for Opus at high and 46.4% at max, the widest coding gap in either direction.

The same-task test adds the cost and reliability side. Both models solved every small coding task once Argon ran at high, but Argon at high took about three times as long and cost more even at today's price, and on exact work over long input it also needed streaming. If your agent edits code in a terminal for long stretches, start on Opus 5.5 at medium and raise effort before you switch models. If it mostly generates self-contained code that a test suite checks, Argon at its default is the cheapest per call, and how often your tests catch a miss decides whether that saving survives.

Where GPT-6.1 Sol stands: cheapest per task, with a real quality gap

Google's table has no GPT-6.1 Sol column, so every figure for Sol comes from third parties.

On Artificial Analysis' v5.0.1 index, Sol ties Argon at high effort (50 each) and trails Opus by two points (52), for about one-sixth of Opus's cost per task and about one-eighth of Argon's at the introductory rate. On the Vals Index, Kingy AI's retrieval of the v2.1 leaderboard shows GPT-6.1 Sol at 61.15% for $3.24 per test, 7.75 points below Argon at about one-fifth of Argon's $15.68. On v2.2 (October 30) Sol is at 63.94% for $3.61 per test against Argon's 69.24% for $13.37.

Sol is the budget choice with a real quality gap, and the size of the gap depends on the suite. It is small on Artificial Analysis' mixed index, where Sol at max (55) even passes Argon at high, and larger on Vals' finance, legal and tax work. Its price position also improves on December 1, when Sol stays at $2 / $10 and Argon doubles. If you are weighing Sol against Opus alone, Claude Opus 5.5 vs GPT-6.1 Sol works through when Opus pays off. Within OpenAI's lineup, GPT-6.1 Sol vs GPT-6 Astra covers when Astra is worth the step up.

One safety figure circulates with the wrong model attached. Google's prompt-injection chart, as reported by VentureBeat, shows attack success rates of 0.7% for Argon, 1.0% for Opus 5.5 and 27.0% for GPT-6 Sol. That last number is the older Sol. No value for GPT-6.1 Sol has been published. Google's October 13 post restates Argon's 0.7% as measured across 15 adaptive attempts.

Which of the three you can call today

All three can be called from code: Argon since October 13, 2026 on the Gemini API paid tier, Opus 5.5 since September 22 and GPT-6.1 Sol since September 29. "GPT 6.1" in this matchup means GPT-6.1 Sol. Which chat apps you also get depends on the subscription you hold.

Google's Gemini 4 Argon launch post says the model is "globally available today for developers and enterprises via the API in AI Studio and Gemini platform." The Gemini API models page lists gemini-4-argon under Gemini 4 as Stable, and the Google Cloud post on Gemini 4 Argon adds Gemini for business on the web, desktop and mobile.

On the Claude side, the Free plan includes neither Opus 5.5 nor Claude Code, and on October 30, 2026 Max 5x and Max 20x were renamed Max ($100) and Ultra ($200). How to Access Claude Opus 5.5: 6 Ways In, Which Free Routes Work lists each route.

Where you personally get Argon depends on which Google product you pay for:

The Google AI Pro line needs a careful reading. The Gemini API pricing page says "Gemini 4 Argon is not available on the Free tier," and API calls are billed separately from any Google subscription. The launch post separately gives AI Pro subscribers Argon in the Gemini app and in AI Mode. An AI Pro subscription therefore gets you Argon in chat, and calling gemini-4-argon from code takes a paid API account. How to Access Gemini 4 Argon for Free: Who Actually Qualifies covers the free and trial routes.

Rate limits for Argon do not appear on the model page or the pricing page. Check the limits on your own project before you route production traffic to it.

Fairwind now matters only for one case. Google is extending Argon to "hundreds of additional" partners in its program for vetted security defenders, and those teams get the model "without cyber guardrails on a rolling basis." The Fairwind eligibility guide explains what the program is and who qualifies. It was written for Gemini 3.8 Flash Cyber, so treat its steps as background for Argon rather than an application path.

Two limits apply on the OpenAI side. GPT-6.1 Sol is not in ordinary ChatGPT chat or on the Free and Go plans, and tool calls require the Responses API. The GPT-6.1 Sol pricing and access guide covers what breaks when you switch from the older Sol.

Argon's 1M token limit: shared by input and output, 65,536 out by default

Argon's million tokens are one budget for input and output together, not a million tokens of output on top of a full prompt. The Gemini 4 Argon model page lists a context window of 1,048,576 as a "unified input and output token limit." Its maximum output reads "65,536 (default synchronous) / up to 1,048,576 (with long decode)."

In practice, a normal API call returns at most 65,536 output tokens, which is about half the 128K that Opus 5.5 and GPT-6.1 Sol allow per response, and Google's API docs (as of November 4, 2026) cap a call without streaming at 32,768. Thinking counts against that cap, so a long high run without streaming can stop before it writes any answer and still be billed, which is what happened on the 146 KB log above. Stream any Argon call that may think for a long time. Going past 65,536 takes long decode, a separate mode. Read how the Gemini API docs set it up before you design around it. VentureBeat notes that Google has not clarified how the unified limit relates to the longest single uninterrupted response.

This replaces the picture from launch week. On September 30 Google described a 1M output limit, up from 64K, and Vals listed the endpoint it evaluated with 262K maximum output. The published model page is the contract to plan against.

A larger output budget also cuts both ways on cost. At $10 per million, a single response that uses 500K output tokens costs $5, and $10 from December 1.

What to run now, and what to test before December 1

All three can be called today, so the decision is which one wins your own tasks at the price it will have from December 1, and the weeks until November 30 are the cheapest time to find out.

1. Start on GPT-6.1 Sol if cost per task drives the decision. Raise effort before you change models. Sol at max scored 55 for $1.27 per index task, which is still below the cost of Opus at medium ($2.45) and above Argon's score at high (50).
2. Use Opus 5.5 for coding agents, exact work over long input, and anywhere one failed run costs more than the price difference. Its default medium effort was enough for every task in the same-task test, including the log count, and it leads the terminal rows in Google's own table. The Opus 5.5 pricing guide explains what its rates and limit reset change. If Opus at max still fails your hardest tasks, Anthropic's newer Claude Fable 5.2 (released November 9 at $8 / $40) is the next step; Claude Fable 5.2 vs Opus 5.5: When Fable High Beats Opus Max covers when it pays.
3. Test Argon now on the work where it leads, at high with streaming. Finance, legal, long-context retrieval and video understanding are its strongest rows in Google's table, and AutomationBench is its clearest independent lead (77.5% against 63.2% for Opus at high on Artificial Analysis). Run your evaluation set before November 30, while each run costs half what it will later.
4. Price every Argon decision at $4 / $20. If Argon wins your tasks only at $2 / $10, the case for it ends on December 1. Keep the model name in configuration, so moving traffic after the switch is a config change.

On model spend alone, cost per accepted task is cost per attempt divided by pass rate. At the Artificial Analysis high-effort figures, Argon is cheaper than Opus 5.5 per accepted task only if its pass rate on your work is more than 1.44 times Opus's now ($4.98 ÷ $3.46) and 2.88 times from December 1 ($9.95 ÷ $3.46). Since no pass rate exceeds 100%, from December 1 that means Opus would have to fail more than 65% of your tasks (pass fewer than 34.7%, which is 1 ÷ 2.88) even if Argon passed every one. GPT-6.1 Sol stays cheaper than Argon as long as its pass rate is more than 11.9% of Argon's now ($0.59 ÷ $4.98), or 5.9% from December 1 ($0.59 ÷ $9.95). Review time and retries move these lines, so measure them on your own work.

Before December 1, check these on Argon:

• Tokens per accepted task on your own workload. This decides the bill. Record input, cache reads, output and reasoning tokens for the same tasks on all three models, then reprice Argon's at $4 / $20. Count input in your own language, since Opus counts prose at 1.4–2.3 times Argon's tokens.
• Whether long calls are streamed. Without streaming, Argon stops at 32,768 output tokens including thinking, and a stopped call is still billed.
• How often the implicit cache hits. Argon has no explicit caching, so long reused prompts only get the discounted rate when Google's cache matches them.
• Whether any response needs more than 65,536 output tokens. If so, test long decode before you depend on it.
• Your project's rate limits. They are not published on the model or pricing page.
• Your terminal and long-running agent tasks. Compare against Opus 5.5 at the same effort, and watch cost on long trajectories, where Vals saw it rise sharply.
• Computer use, if you depend on it. It is a preview feature on Argon, its CUA-bench result on Vals is the weakest number in its profile, and Google's computer-use guide names Gemini 3.8 Flash as the recommended model for it.
• Batch or Flex for jobs that can wait. Both halve Argon's rates, to $2 / $10 from December 1.

Gemini 4 Argon vs Opus 5.5 and Sol: common questions

Can I use Gemini 4 Argon now?

Yes, since October 13, 2026. Developers can call gemini-4-argon on the Gemini API paid tier, in Google AI Studio and through Gemini platform on Google Cloud. Google AI Pro and Ultra subscribers get it in the Gemini app and AI Mode in Search. Ultra subscribers and, since October 15, 2026, paid Google AI Pro subscribers (and, as of October 23, 2026, Pro trials that came with eligible premium devices) also get it in Google Antigravity, in the Antigravity app, CLI and IDE extensions but not the Antigravity IDE, according to the Antigravity models page; Pro access is still rolling out. It is also in Gemini for business. It is not on the Gemini API free tier.

When does Gemini 4 Argon's introductory price end?

The introductory rate of $2 input, $10 output and $0.10 cached input per million tokens runs through November 30, 2026. From December 1, the standard rate is $4 / $20 with cached input at $0.20. Batch, Flex and Priority double on the same date.

Is Gemini 4 Argon the same price as GPT-6.1 Sol?

Per token, yes until November 30: $2 input, $10 output and $0.10 cached input per million, the same as GPT-6.1 Sol for requests up to 272K input tokens. From December 1, Argon costs twice as much per token as Sol. Per task, no, even now: on Artificial Analysis' index (v5.0.1, as of November 11, 2026) at high effort, Argon costs about 8.4 times what Sol does at the introductory rate, because it uses far more tokens. For other models, the AI API price comparison ranks options by input and output mix.

Is Gemini 4 Argon half the price of Opus 5.5?

Only per token, only for identical token counts and only until November 30, 2026. Per task it is not cheaper even now: on Artificial Analysis' index (v5.0.1) at high effort, Argon costs about 1.44 times Opus 5.5 at the introductory rate and 2.88 times from December 1, as of November 11. In a same-task run through one gateway, Argon at high cost $0.40 against $0.31 for Opus at high. From December 1, Argon's $4 / $20 rate equals Opus 5.5 per token.

Is Opus 5.5 faster than Gemini 4 Argon?

Yes, whenever both reach a correct answer. In the same-task run (Argon on November 4, Opus 5.5 on November 11, 2026), Opus at high finished eight small tasks in 147 seconds against 463 for Argon at high, Opus at medium took 126 seconds against 163 for Argon's default, and on the 146 KB log Opus answered correctly in under a minute while Argon needed 271–402 seconds; Argon's default replied in seven seconds there, but with wrong counts. On Artificial Analysis' index as of November 11, Argon High averages 28.6 minutes per task against 9.5 for Opus 5.5 at high.

Should I wait for Gemini 4 Argon or build now?

There is nothing left to wait for, since Argon is open on the paid API. Build on whichever of the three wins your evaluation tasks. Independent results put Argon within about two points of Opus 5.5 at matched effort, so switching does not buy a clear step up by default. Use the introductory window to measure Argon's tokens per task, and judge it at the December price.