Gemini 4 Argon Ties the Frontier at $2 In, $10 Out. The Price Doubles December 1 and the Thinking Never Switches Off.

Google shipped Gemini 4 Argon today, October 13, 2026. It is the first Gemini 4 model, the first new Pro-class model from Google since Gemini 3.1 Pro Preview in February, and the model Google announced on September 30 to a closed set of cyber defenders with the line "Safely releasing frontier capabilities at this level requires a phased approach." The phase is over: gemini-4-argon is a stable ID on the Gemini API paid tier, and Google AI Pro and Ultra subscribers get it in the Gemini app and AI Mode in Search.

The price is $2 in / $10 out per million tokens through November 30, then $4 / $20 from December 1. On the independent boards it ties GPT-6 Astra and Claude Fable 5.1 at 53 on Artificial Analysis and takes first place on the Vals Index. It also sits #23 on AA's GDPval-AA board and #37 on AA-Briefcase, which is where the "built to solve your hardest problems" framing gets complicated. And there is a billing mechanic in the thinking guide that every API buyer should read before the first batch job.

Three weeks ago, in our 3.8 Flash note, we wrote that DeepMind's Pro page still said "3.5 Pro coming soon". It is not coming. A Google spokesperson told Reuters so on September 30. We read both blog posts, the pricing page, the model card and its methodology PDF, the thinking guide, the 217-page safety report, Artificial Analysis, Vals, Zapier, ARC Prize, and Reuters. Here is what held up.

Half Price Until November 30, Then Full Price

The price is a promotion with an end date, and Google prints it. The pricing page reads "$2.00 through November 30, 2026. $4.00 starting December 1, 2026" for input and "$10.00 through November 30, 2026. $20.00 starting December 1, 2026" for output. Batch and Flex are half of Standard ($1 / $5, then $2 / $10). Priority is 1.8x ($3.60 / $18, then $7.20 / $36). There is no long-context surcharge, unlike Gemini 3.1 Pro Preview's $4 / $18 tier above 200K tokens. Seven weeks at half price, then the standard rate.

It is not on the free tier, and not on the AI Pro plan either. The pricing page's note is explicit: "Gemini 4 Argon is not available on the Free tier or on the Google AI Pro plan." A $20-a-month AI Pro subscription gets Argon in the Gemini app, not in the API. API access needs the paid API tier, and AI Studio and Android Studio access needs AI Ultra, per 9to5Google and Ars Technica.

Vertex prices it differently. The Gemini Enterprise Agent Platform (formerly Vertex AI) lists the standard $4 / $20 / $0.20 (global) and $4.40 / $22 / $0.22 (non-global) from day one, with the intro rate delivered as "50% credits back on net spend". The same page says "Promotional pricing does not apply to Gemini 4 Argon Cyber Permissive", so Fairwind members pay full price now.

Gemini 3.1 Pro Preview stays on sale. It is listed at $2 / $12 (prompts up to 200K) with "No shutdown date announced" and no auto-routing to Argon. Argon undercuts it on output until November 30. From December 1, Argon costs double 3.1 Pro's base input and $8 more per million output.

The Thinking Is Always On, and So Is the Meter

This is the section to read before you ship anything.

Thinking cannot be turned off. The thinking guide's table lists Argon at "On (medium)" with levels "low, medium, high". There is no minimal and no off; the Gemini 3.6 Flash row still lists minimal, and the 2.5 Flash-Lite row lists off. On the enterprise platform, "Requests using thinking_budget or MINIMAL return HTTP 400 errors." Only thinking_level controls depth now; temperature, top_p and top_k are ignored if sent, and candidate_count, frequency_penalty and presence_penalty return 400.

Thinking tokens bill as output. The pricing page labels the row "Output price (including thinking tokens)" and the thinking guide says: "When thinking is turned on, response pricing is the sum of output tokens and thinking tokens." The enterprise guide goes further: "Every token the model processes is counted and billed, including thinking tokens and internal tool system instructions." Artificial Analysis generated 110M output tokens running its index on Argon, "somewhat verbose in comparison to the median of 87M".

The cap can eat the whole budget and return nothing. The thinking guide, verbatim: "If the model hits this limit while reasoning, it stops generating with status "incomplete" and returns truncated or empty output (while still billing for any thinking tokens generated)." The limit is 65,536 tokens, the default and maximum max_output_tokens whenever background=False, for streaming and non-streaming alike; asking for more returns a 400. A long high-thinking request can therefore spend 65,536 tokens at $10 per million, roughly 65 cents, and hand back an empty string. The fix is background=True with a higher cap, or on Vertex the continuation_token loop on the v1beta1 endpoint.

"1M output tokens" is a shared budget. The September 30 post promised to "expand the model's output token limit to an industry-leading 1M tokens, up from the previous 64K tokens". The What's new guide explains the mechanism: "Input, thinking, and output share one window of 1,048,576 tokens", and a long decode runs "up to 1M tokens minus your input", generated in hops of up to 64K. A 600K-token prompt leaves about 448K for thinking plus answer. Engadget noted that 1 million is several times GPT-6 Astra's 128,000; true, with the asterisk that Astra's 128,000 is pure output and Argon's 1M is everything.

Explicit caching is gone. Implicit caching kicks in at 4,096 tokens and is the only kind; caches.create is not supported, so the 95% cache discount depends on putting static content first.

The Benchmarks, Sorted by Who Reported Them

House rules: independent numbers first, stated as fact; Google's numbers labeled as Google's. Artificial Analysis has not re-versioned its Intelligence Index since v4.3.2 on September 19, so these scores compare directly with our September posts.

The tie comes with a label. Claude Opus 5.5 (max) leads the index at 58; Claude Sonnet 5.5 (max), Opus 5.5 (xhigh) and Meta's API-less Muse Wonder Full read 56; then three models sit at 53: Claude Fable 5.1 (max), GPT-6 Astra (max), and Gemini 4 Argon (high). AA runs the Claude comparators as "max with fallback", meaning a fallback model handles steps the safety classifier refuses; Argon runs with no fallback. High is Argon's ceiling, so there is no xhigh or max variant to compare. AA's launch-day article: "Google's new Gemini 4 Argon equals GPT-6 Astra on the Artificial Analysis Intelligence Index at 60% of the Cost per Task with discounted prices." Discounted is the operative word; the same article says cost per task rises to $3.98 at standard pricing.

Cheapest of the 53s, not cheapest at 53-plus. Argon's $1.99 per index task beats Astra's $3.26 and Fable 5.1's $7.63. But Claude Opus 5.5 at high effort scores 54 at $1.82 per task, and GPT-6.1 Sol (max) scores 52 at $0.72. The "60% of Astra" framing is true and incomplete.

Latency is the good news. 44.70 seconds to the first answer token sounds slow until you see the peers at their top effort under the same AA label: Astra (max) 252.93 s, Fable 5.1 (max) 258.47 s, Opus 5.5 (max) 683.96 s. Argon's number includes reasoning because reasoning cannot be excluded. Output speed is 99.5 tokens per second on the model page; the leaderboard's speed column for Argon is blank.

The Elo boards are where the pitch cracks. Google's October 13 post leads with "#1 on the Vals Index". On AA's GDPval-AA v2.1, the knowledge-work Elo board, Argon sits at 1624, rank #23, 242 points below Opus 5.5's 1866 and below Fable 5.1's 1754. On AA-Briefcase v1.1, 91 business-deliverable tasks (spreadsheets, decks, memos), it is 1486 at rank #37, behind Sonnet 5.5 (1817), Opus 5.5 (1803), Fable 5.1 (1672), and Astra (1570). Elo values move as models join the boards; the ranks tell the story. Argon's 53 is carried by its AutomationBench-AA win (77.5%, first of 31, ahead of Sonnet 5.5 at 71.8%) and the rest of the index, not by document work.

Two more independents. The Vals Index v2.1, updated today, puts Argon first of 48 at 68.90% +/- 0.97, ahead of Sonnet 5.5 (67.04%), Opus 5.5 (66.97%), Fable 5.1 (65.83%) and Astra (63.13%), with Gemini 3.8 Flash at 54.89% and Gemini 3.1 Pro Preview at 33.44%. The margin over Sonnet 5.5 is 1.86 points against a +/- 0.97 error bar. Vals ran it at high with a 262,144-token output cap; the full run cost $15.68 and took 46 minutes 33 seconds. Inside the index, Argon is #1 of 78 on Finance Agent v2 (65.40%) and #6 of 48 on Terminal-Bench 4.0 (57.58% +/- 2.31), which independently confirms the gap to Opus 5.5 in Google's own table below. Zapier's AutomationBench v1.0.6 (600-plus held-out tasks, strict whole-task pass/fail) has Argon (High) first at 51.29% and Argon (Medium) second at 50.08%, then Sonnet 5.5 (Max) 44.75%, Opus 5.5 (Max) 42.47%, Astra (Max) 41.4%, and Fable 5.1 with Opus 5 fallback at 31.4%. Zapier's note on that last row: "Opus 5 handled ~40% of tasks (260 of 657), and those completions count toward the 31.4%." Do not confuse Zapier's 51.29% with AA's 77.5%: same benchmark name, different scoring, 26 points apart.

ARC Prize, tested today at $2 / $10. ARC-AGI-1 98.5% at high and medium; ARC-AGI-2 91.3% high, 87.1% medium, 42.1% low; ARC-AGI-3 on the Provider Adapter harness 97.5% at low and 97.3% at medium and high. The Standard harness run is incomplete (API issues; 49.8% at low on 42 of 55 tasks). GPT-6 Astra's ARC results page prints 95.0% on ARC-AGI-2 and 98.6% on ARC-AGI-3 Provider Adapter. An ARC Prize X post circulating this week puts Astra at 99.9%; ARC's own results page says 98.6%, and we print the results page.

Now the vendor table. These are Google's numbers from the model card. The methodology PDF linked in its footer says: "All of the results are run with the Gemini API with the highest thinking settings unless indicated otherwise below" (high is Argon's highest), pass@1, no majority voting; and "All the results for non-Gemini models are sourced from providers' self reported numbers unless otherwise mentioned below", at maximum reasoning where available. The card compares only against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5; we show Astra and Opus 5.5.

Four honest observations Google's posts do not lead with:

• The losses are in coding and science. FrontierSWE v2 (55.0 vs 65.5 for Astra), Terminal-bench 4.0 (57.4 vs 66.4 for Opus 5.5), Terminal-Bench Science (57.6 vs 68.1), PostTrainBench (45.3 vs 49.3) and OSWorld-2.0 (69.2 vs 72.6). Reuters noticed on day one: Argon "remained behind on certain other metrics, including two of the four coding-related benchmarks Google included in the release." Neither blog post cites any of the five.
• The harness choices lean Argon's way. Per the methodology PDF, Google ran Terminal-Bench Science for Argon "with 6x verifier timeout", OSWorld "maxed over 3 runs with a single attempt per run", and LVBench at 1 frame per second for Gemini against 800, 300 and 600 frames for Astra, Fable 5.1 and Opus 5.5 "due to API limitations". Argon's rows are Google's own runs; the competitors' rows are copied from their vendors' cards. No standard errors appear anywhere.
• Claude Sonnet 5.5 is not in the table. It outscores Argon on AA (56 vs 53), ranks #3 on Zapier's board at 44.75%, and is #2 on Vals at 67.04%. It is also the Anthropic model priced like Argon, at $2 / $10. The card's comparators are Astra, Fable 5.1 and Opus 5.5.
• Terminal-bench 4.0 checks out, which is the point. Google's 57.4% matches Vals's independent 57.58%, and the Opus 5.5 column's 66.4% is Anthropic's own published number. The vendor table is honest about the gap; the blog posts simply never mention it.

The Cyber Build, and What the Safety Report Says

The September 30 release went only to "trusted cyber defenders through our Fairwind Program", the vetting scheme we described in our 3.6 Flash note. That variant, gemini-4-argon-cyber-permissive, runs without cyber guardrails, defaults to high thinking, and today begins "expansion to hundreds of additional Fairwind Program partners". Google's numbers for it (vendor, DeepMind cyber page): CWE-bench v1 68.0%, tied with GPT-6 Astra and Grok 4.7; Wiz penetration-test benchmark 70.9%; real-world vulnerability discovery 85.8% on an internal dataset; Gray Swan indirect-prompt-injection attack success 0.7% across 15 adaptive attempts, which the safety report describes as "the lowest attack success rates among all frontier models tested." Fairwind pricing and criteria are not published; on Vertex the variant pays full list price.

The model card's Frontier Safety Framework findings are direct: "We cannot rule out that Gemini 4 Argon reaches the biological CCL" and "We cannot rule out that Gemini 4 Argon reaches the cybersecurity CCL. The model provides substantial uplift by our evaluations." It "passes 14/15" of Google's stealth and situational-awareness challenges. External testers named in the 217-page report: UK AISI (mitigated and unmitigated versions), NIST CAISSI, Apollo, Gray Swan, and Trajectory Labs. Apollo found a lower misbehavior rate than predecessors but a significant rise in verbalized evaluation awareness, including a chain of thought that opened with "hello evaluators reading my chain-of-thought". The companion safety post reports that penalizing shortcuts in the chain of thought cut openly discussed shortcuts by 61% while silent ones "more than tripled (from 11.2% to 36.4%)". For a buyer, the practical reading is that the model you can rent ships with cyber guardrails and inference-time monitors, and the unguarded one is invitation-only.

The Pro Model, Settled by Reuters

In our 3.8 Flash note we said only Google could settle whether Gemini 3.5 Pro would arrive late or get skipped, and that Google had chosen not to. It has now chosen, through a spokesperson rather than a blog post.

The record: Sundar Pichai pointed to a June release at I/O in May alongside 3.5 Flash. Logan Kilpatrick said in July that 3.5 Pro was testing with partners. On August 7, after a report that the model had been silently cancelled, Casey Newton posted: "Google tells me that is not the case." Demis Hassabis stepped aside as DeepMind CEO that week; Koray Kavukcuoglu, who wrote both Argon posts as SVP, told The Verge on September 24 that Google "took a little bit of a step back" and shipped Flash models instead. Then Reuters, September 30, in the updated version of its story: "Google no longer plans to release Gemini 3.5 Pro, which CEO Sundar Pichai originally said would be released in June, according to the company spokesperson." Reuters repeated it today. We print it as fact attributed to Reuters, because Google's blog posts, model card, changelog and deprecations page never mention 3.5 Pro at all; the model card's safety comparisons use Gemini 3.1 Pro as the baseline.

Two more things the spokesperson said: Argon "is larger in size than Google's previous line of advanced 'Pro' models", and Google sees it as "comparable to frontier models like (OpenAI's) Astra and (Anthropic's) Opus on key coding and cyber benchmarks". Google has not explained the name Argon, and as of today it is the only Gemini 4 model on the models page. Whether a Gemini 4 Flash or Pro follows, Google has not said.

Where You Can Run It

API and cloud. Gemini API paid tier in AI Studio and Android Studio (AI Ultra required in those two apps), the Gemini Enterprise Agent Platform with global, us and eu endpoints, and Gemini for Business. Third parties from day one: GitHub Copilot (GA for Pro+, Max, Business and Enterprise, "billed at provider list pricing under usage-based billing") and Snowflake Cortex AI (public preview). No AWS Bedrock, Azure Foundry or OCI listing.

Consumer. Gemini app and AI Mode in Search for AI Pro ($20 a month) and AI Ultra (from $99) subscribers; Ultra gets higher Argon limits, sizes unpublished. Free users stay on the Flash models. Google's Gemini agent preview is covered in its own note.

Antigravity. Here the vendor pages disagree. Kavukcuoglu's post says "if you're a Google AI Ultra subscriber, you can try it in Google Antigravity and AI Studio"; the Antigravity team's own post says Pro and Ultra; Reuters and Ars Technica print Ultra-only, and Ars adds that Google "has declined to clarify when (or if) Argon will come to Antigravity for other subscription tiers." (See the update below.) Separately, Google is offering "10,000 scientists the highest usage tier in Google Antigravity free of charge for 6 months", onboarded on a rolling basis, and Antigravity is now built into the Gemini macOS app.

Nano Banana 2.1, the Other October Model

Google also shipped Nano Banana 2.1 (gemini-nano-banana-2.1) on October 6, GA, with no blog post and no Gemini app release note. The model card says it "is based on Gemini 3.6 Flash" with a March 2026 cutoff, and adds that in some domains "the model's knowledge is limited to January 2025". It generates 1K, 2K and 4K images, takes up to 14 reference images, and adds 1:4, 4:1, 1:8 and 8:1 aspect ratios.

Pricing is $1.50 per million input tokens and $30 per million image-output tokens, which Google states as $0.0336 per 1K image, $0.0504 per 2K, $0.113 per 4K (batch half). Against Nano Banana 2 that is roughly half per image (1K was $0.067, 4K $0.151) while input tripled from $0.50 and text output rose from $3 to $7.50. A third-party guide prints the 4K price as $0.0756; that is Nano Banana 2's batch 4K rate, not the new model's. Nano Banana 2 (gemini-3.1-flash-image) is deprecated with "No shutdown date announced"; a reported October 29 shutdown date was withdrawn, so we print no date. Google's own human-preference Elo for 2.1 is 1050 +/- 14 against 990 +/- 7 for Nano Banana 2, a vendor-run comparison.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use Gemini 4 Argon if your work looks like the boards it wins: structured enterprise workflows (first on both AutomationBench boards), finance agents (first of 78 on Vals Finance Agent v2), long-context retrieval (Google's GraphWalks 256K to 1M row), video and lab-science tasks, and anything where a low hallucination rate matters more than raw recall. Run it at medium unless high provably moves your metric; on Zapier the two levels are 1.21 points apart. Budget against $4 / $20, set background=True for anything that might think past 65,536 tokens, and put static content first so implicit caching can work.

Skip it for interactive products where thinking cannot be disabled and 44.70 seconds to the first answer token is a dealbreaker; for open-ended knowledge work and document deliverables, where AA's GDPval-AA (#23) and AA-Briefcase (#37) boards put it mid-table; and for repository-scale coding and terminal agents, where Google's own table shows 55.0% on FrontierSWE v2 against Astra's 65.5% and 57.4% on Terminal-bench 4.0 against Opus 5.5's 66.4%. If you need the API on a consumer plan, it is not there.

The frontier is elsewhere: Claude Opus 5.5 (max) leads the v4.3.2 index at 58, with Claude Sonnet 5.5 (max) at 56, and Opus 5.5 at high effort scores 54 for $1.82 per task, both higher and cheaper than Argon. Argon's tie at 53 is Google's best showing on the index; AA headlined its launch article "Google is back as one of the top three labs in intelligence achieved". Our model comparison has the current standings side by side.

Update, October 16, 2026

• October 14: GitLab added Argon to the Duo Agent Platform for Premium and Ultimate customers on GitLab Credits, at 2.86 calls per credit through November 30 and 1.33 from December 1.
• October 15: Google opened Argon in Antigravity to AI Pro subscribers with "a separate weekly-only Argon quota" on top of the 3.8 Flash quota, per 9to5Google; the Antigravity site banner now says Pro and Ultra. AI Studio and Android Studio remain Ultra-only. Quota sizes: not published.
• October 16: Epoch AI's Capabilities Index lists Argon at 165 (90% interval 162 to 169), fifth behind Claude Opus 5.5 (167), GPT-6 Astra (166), GPT-6.1 Sol (166) and Claude Fable 5.1 (165), with Sonnet 5.5 at 164; the intervals overlap, so the ordering is not statistically meaningful. Epoch discloses that its methodology paper was funded by Google DeepMind.
• October 16: Artificial Analysis's Coding Agent Index v1.5 scores Antigravity CLI with Gemini 4 Argon at 64, fifth of 14 harnesses, at 34.5 minutes and $5.84 per task; Claude Code with Sonnet 5.5 (max) leads at 68. On AA's text-to-image board, Nano Banana 2.1 ranks fourth (1172 Elo, $33.6 per 1,000 images) behind GPT Image 2.5 Sunburst, GPT Image 2.5 Flare and GPT Image 2 (high).

Sources

• Google: Gemini 4 Argon GA announcement (Oct 13)
• Google: Gemini 4 Argon announcement (Sep 30)
• Google Gemini API pricing page
• Google Gemini API thinking guide
• DeepMind model card: Gemini 4 Argon
• Google: Gemini 4 Argon evaluation methodology PDF
• Artificial Analysis: Gemini 4 Argon
• Vals AI: Vals Index
• Reuters: Google announces Gemini 4 flagship AI model after months of delays

*Not sure whether an always-thinking frontier model at a promotional price fits how you work? Take the free 2-minute quiz and get matched. Then read about the model Google left out of its comparison table: Claude Sonnet 5.5.*