Gemini 4 Argon vs GPT-6 Astra: Benchmarks, API Price Gap and App Access

Astra lists at five times Argon's price until November 30, 2.5 times after

The rates below come from Google's Gemini API pricing page and OpenAI's API pricing page, per million tokens. Astra's column does not change on December 1; Argon's doubles.

Read the long-prompt row first if you send big contexts. OpenAI prices a whole Astra request at twice the input and cache rates and 1.5 times the output rate once the prompt passes 272K input tokens, while Google charges Argon one rate across its 1,048,576-token window. On a 500K-token prompt, input costs $20 per million on Astra against $2 on Argon in November and $4 from December.

Caching points the same way. Argon caches implicitly at 95% off input, with no explicit cache control and no write fee listed; Astra's cache hits cost $1 per million, and OpenAI charges for every cache write. OpenAI's Fast mode gives up to twice Standard speed at twice the price, and Ultrafast costs six times Standard.

An illustrative month: 40 million tokens in, 4 million out

ToolColumn arithmetic, at standard rates for prompts under 272K tokens, with no caching and the same token counts on both models. In practice each model spends its own number of thinking tokens, and Argon bills them as output.

Even at half price through Batch or Flex, Astra costs more than Argon at its full December rate ($300 against $240). Argon's own batch rate would bring its December month to $120.

Google's model card: Argon ahead on 14 rows, Astra on 4

Google's Gemini 4 Argon model card ("results as of October, 2026") is the main table on this page because it was published with Argon and gives both models a result on 19 benchmarks. Argon leads 14, Astra leads 4, and one is tied. Every figure below is Google's.

Astra's four leads are FrontierSWE v2 and Terminal-Bench Science 0.1 (10.5 points each), OSWorld-2.0 (3.4 points) and Terminal-bench 4.0 (0.8 points). Argon's widest leads are on legal, finance and office automation and on the longest inputs: Harvey's Legal Agent Benchmark (19.6% vs 5.4%), GraphWalks at 256K to 1M tokens (84.2% vs 71.8%), Vals Finance Agent v2 (65.4% vs 53.5%) and AutomationBench (51.3% vs 41.4%). Coding splits: Argon is ahead on DeepSWE v1.1 and Vibe Code Bench, Astra on FrontierSWE v2.

Google's newer product-page table puts Astra ahead on 10 of 17 rows

Google's Gemini product page now carries a second table, from the October 28, 2026 launch of Gemini 4 Carbon, a model Google has so far opened only to a subset of cyber defenders. It uses different benchmark versions, some rows are Artificial Analysis runs, and Google's methodology says non-Gemini figures are the providers' self-reported numbers unless noted. On its 17 rows Astra is ahead of Argon on 10, Argon on 6, with one tie. Astra's leads there include the Artificial Analysis Intelligence Index v5.0 (57 vs 50), OSWorld-2.1 (80.9% vs 74.7%), SWE-sweep (9.5% vs 4.3%) and Vending-Bench 2 ($15,515 vs $13,718 final bank balance).

The count changes mainly because the two tables pick different benchmarks, not because they disagree. Where both tables carry a benchmark, they point the same way, with one exception: Agents' Last Exam, where Argon leads on the model card's pass rate (39.5% vs 34.2%) and Astra on the product page's version 1.1 (73.1% vs 69.0%). This page does not mix figures from the two tables.

OpenAI publishes no Argon comparison, and its Astra figures differ from Google's

OpenAI's Astra launch post (September 3, 2026) predates Argon, and its only Google column is Gemini 3.8 Flash; OpenAI has published no comparison with Argon since. Where OpenAI and Google both report Astra on a benchmark of the same name, the figures do not always match.

Read each column as the figure that publisher prints: Google's Carbon methodology says its non-Gemini scores are the providers' self-reported numbers unless noted, and the "(AA)" rows are Artificial Analysis runs. OpenAI's footer says its scores are the maximum at any effort; neither vendor explains the other gaps. Astra's 0.8-point lead on Terminal-bench 4.0 in Google's card is smaller than the 1.2-point spread between published Astra figures for that benchmark, so ToolColumn reads that row as level.

Astra writes 128K tokens per request; Argon's long decode runs to 1M

Both models read about a million tokens. The differences sit on the output side and in how you set reasoning.

For one long answer under 128K tokens, Astra needs no special handling; for anything longer, such as a full translation of a book-length document, only Argon's long decode goes further. Argon also takes video and audio directly, which Astra does not. OpenAI's Astra model page and latest-model guide do not state Astra's default effort, so set it explicitly when you compare costs.

Plus gets Astra only in Work and Codex; AI Pro puts Argon in the Gemini app

The two cheapest app plans buy different things. ChatGPT Plus at $20 a month gives you Astra for agentic tasks in Work and Codex but not for chatting, while Google AI Pro at $19.99 a month puts Argon in the Gemini app itself.

• ChatGPT: Plus includes GPT-6 Astra in Work and Codex, where OpenAI's help center estimates 5–45 local Astra messages per 5-hour window on Plus and Standard Business seats, as estimates rather than guaranteed counts. In Chat, Astra runs only as GPT-6 Pro, on eligible Pro, Business and Enterprise plans. Pro 100, Pro 200 and Pro 500 currently have no five-hour limit in Work and Codex, and among Plus, Pro and Business, only Pro 500 includes Astra Ultrafast at launch. Enterprise administrators enable Astra for their workspace; it was off by default at launch.
• Gemini app: Google AI Pro ($19.99 a month) and AI Ultra ($99.99 or $199.99) include Argon in the Gemini app and in AI Mode in Search, in select countries. AI Plus gets the Flash model and accounts without a plan get Flash-Lite. Google publishes no Argon message count; app limits are compute-based and refresh every 5 hours up to a weekly limit.
• API: Astra is gpt-6-astra on the OpenAI API (Responses, Chat Completions and Batch) and through Microsoft Azure and Amazon Bedrock. Argon is gemini-4-argon, a stable model on the paid tier of the Gemini API only, and on Google Cloud's Gemini platform.
• Coding agents: Astra runs in Codex on the ChatGPT plans above. Argon runs in Google Antigravity, on AI Ultra according to Google's launch post and on AI Pro and Ultra according to the Antigravity blog.