Gemini 4 Argon vs GPT-6 Astra: Same Intelligence Score, a Fifth of the Price

Google and OpenAI are now selling the same pitch: a frontier model for deep work that chews through legal briefs, financial models, and sprawling codebases. The difference is on the receipt. Gemini 4 Argon arrived on October 13 at roughly a fifth of what GPT-6 Astra charges per token, and it matched Astra on the most widely cited independent intelligence index.

That price gap is the loudest number in this matchup, but it is not the whole story. Six weeks after its September 3 launch, Astra has the maturity, the computer-use skills, and the faster task times. Argon answers with knowledge-work strength, unusually disciplined honesty, and the top spot on the Vals Index. For the full September picture of how Astra landed, see our breakdown of Astra’s launch benchmarks against Claude’s flagships.

This Gemini 4 Argon vs GPT-6 Astra comparison covers what the two models actually are, what every major independent test says, what they cost per token, per task, and per subscription, and how fast they finish real work. The two run neck and neck on average intelligence. The split happens everywhere else.

What Gemini 4 Argon and GPT-6 Astra Actually Are

Both models are aimed at the same buyer: professionals who want an AI to sustain long, multi-step work without hand-holding. Google’s pitch for Argon, the first model in its Gemini 4 generation, centers on deep reasoning across coding, finance, legal work, cybersecurity, and science, with an expanded output ceiling that lets the model think in one long trajectory instead of stopping every few thousand tokens. Google’s launch post describes an industry-leading 1 million token output limit (up from 64K), a continuation protocol for ultra-long generations, and internal wins from a browser refactor in Chrome to overnight optimizations of Google’s own open-source codebases.

Argon’s rollout was staged. Google announced it on September 30 to a small set of trusted cyber defenders through its Fairwind Program, then opened broad access on October 13 for developers and enterprises through the Gemini API and AI Studio, for Google AI Ultra subscribers in Antigravity, and for Pro and Ultra subscribers in the Gemini app and AI Mode in Search.

GPT-6 Astra has been in the wild since September 3. OpenAI’s successor to the GPT-5.6 family ships with a 1.05 million token context window, reasoning-effort settings from low to max, and a Fast mode for latency-sensitive work. Its launch rollout followed a similar playbook: Daybreak cyber-defense partners first, then ChatGPT plans, the API, and AWS. Astra’s standout behaviors are computer use (navigating real desktops and browsers), delegating subtasks to other models mid-run, and keeping notes across context windows.

The Headline Scores: Dead Even, With One Exception

On Artificial Analysis Intelligence Index v4.3.2, the two models post the same score: 53. Argon was tested at its highest reasoning setting (high) and Astra at max, so this is the closest thing to a like-for-like comparison the industry has. On the surface, Google’s cheaper model has caught OpenAI’s flagship.

The Vals Index, a separate suite weighted toward economically valuable work, disagrees on placement even as it confirms parity in spirit. Argon ranks first of 48 models at 68.90% overall accuracy. Astra ranks seventh at 63.13%. Five points is a meaningful gap on a suite built around finance, legal, coding, and tax tasks.

Two models can tie on a composite index and still be built for different jobs. To see where the work actually goes, you have to open up the suites.

Where Each Model Wins, Benchmark by Benchmark

Pull the averages apart and the pattern is close to a clean split. Argon leads written and analytical work: legal, finance, tax, security auditing, and anything measured by reading and reasoning over documents. Astra leads on computer use, terminal-heavy engineering benchmarks, and long-horizon agent tasks.

Knowledge work: Argon sweeps finance, legal, and tax

On the Vals suite, the professional-work gaps are some of the largest in this comparison. Argon scores 65.4% on multi-step financial research against Astra’s 53.5%, 54.8% against 39.4% on legal research, 19.6% against 5.4% on Harvey’s legal-agent benchmark, and 44.6% against 20.7% on tax work. Google didn’t just price Argon to win quotes; it tuned the model for the work Vals measures.

Argon’s edge stretches into very long inputs as well: 84.2% versus 71.8% on GraphWalks’ 256K-to-1M-token range, with no long-context surcharge on Google’s side.

The honesty split: fewer right answers, far fewer confident mistakes

The most interesting behavioral difference sits in Artificial Analysis’s knowledge test. Astra answers more questions correctly, 63% to Argon’s 50%, but it does something very different when it does not know the answer. Astra’s non-hallucination rate is 49%, meaning it gives a confident wrong answer about half the time it lacks knowledge. Argon’s is 85%. In pipelines where a fabricated statistic is worse than an admission of ignorance, that is a real operational advantage, and it shows up in production settings more than any single benchmark score.

Coding: a narrow split at the top, a widening gap below

On Artificial Analysis’s Coding Agent Index, which measures full agent runs inside a coding harness, Argon scores 64 in Antigravity’s CLI and Astra scores 62 in Codex. Argon also holds the headline software-engineering number, 77.9% on DeepSWE v1.1 against Astra’s 74.1%. But the picture flips in the terminal. Astra wins Terminal-Bench 4.0 at its xhigh setting (59.6% to 57.1%), nearly doubles Argon on Terminal-Bench Science (62.9% to 44.3%), and leads FrontierCode 1.1, 47.0% to 42.6%. On Google’s own launch table, Astra even takes the FrontierSWE v2 row, 65.5% to 55.0%. Astra is also dramatically better at computer use: 19.2% to 4.8% on Vals’ CUA-bench, the measure of operating a desktop environment without help.

Long-horizon agents: Astra’s clearest edge

The Time Horizon Index at Vals is the single largest spread between the two models: 90.5% for Astra against 52.5% for Argon on maintaining coherent behavior through very long task sequences. Combined with its lead on document-heavy workflows (32.2% to 21.8% on GDP.pdf) and office-task automation (93% to 89% objectives completed on AutomationBench-AA), Astra looks like the stronger pick for autonomous, many-step pipelines that run unattended.

Security: Argon’s home turf

Google trained Argon specifically for cyber defense, and the independent numbers back the positioning. Argon cut prompt-injection attack success to 0.7% on Gray Swan’s 2026 indirect prompt injection benchmark, ahead of Claude Opus 5.5 and Fable 5.1 at 1.0% and far ahead of Astra at 8.5%. On Vals’ CyberBench, Argon scores 77.9% against Astra’s 41.1%. On CWE-bench v1, which measures vulnerability remediation, the two tie at 68%.

The two labs also draw their own safety lines differently: OpenAI classifies Astra’s cybersecurity capability at the Critical tier of its Preparedness Framework, while Google keeps Argon under its Frontier Safety Framework with activation-level monitoring during deployment.

Robotics: two benchmarks, two opposite stories

Google positions Argon as a robotics model too, citing an 87% score on the Embodied Reasoning (ERQA) benchmark and a demo where the model watched a human toss beanbags, built a simulation from the video, and transferred the skill to an Apptronik Apollo humanoid that sank seven shots in a row. Against that, Vals’ Robo-bench shows Argon at 0.83% and Astra at 23.3%, one of Astra’s largest wins anywhere in the suite. Robotics capability clearly depends on which slice of the stack a benchmark probes, so treat both camps’ robotics claims as early evidence rather than settled fact.

The vendor view: Google’s own comparison table

Google’s launch materials include a head-to-head table of their own, and even on home turf it shows the same pattern: Argon leads the knowledge, long-context, and remediation rows, while Astra holds terminal-driven science, computer use, and one of the software-engineering suites. The table runs to nineteen rows, and Argon takes 14 of them to Astra’s four, with one tie. Here is the selection where the two models diverge most.

Source: Google’s Gemini 4 Argon launch materials, published September 30 and October 13, 2026. Vendor tables run inside each lab’s native harness, so small swings against the independent suites above are expected; read the pattern, not the decimals.

Below is the full data picture, grouped the way Artificial Analysis and Vals organize their suites.

Source: Artificial Analysis v4.3.2, models run at their highest available reasoning settings. Elo percentages are calculated as (Elo minus 500) divided by 2000. *Astra’s max-reasoning result on Terminal-Bench 4.0 is 59.1%, while 59.6% is its xhigh run.

Source: Vals.ai, October 2026. Beyond the overall table, Argon holds the number one spot on Finance Agent v2 (of 78 models) and Code Migration (second of 77), while Astra’s standout ranks include first of 41 on Terminal-Bench Science and first of 35 on SRE Bench.

The Cost Picture: Tokens, Tasks, and Plans

API pricing is where the two models look least alike. Argon’s introductory rates are $2 per million input tokens and $10 per million output tokens, with cached input at $0.10. Astra sits at $10, $50, and $1.00. That is a fifth of the price on every line, and a tenth on cache reads. Google’s footnote matters, though: the introductory window closes on November 30, 2026, after which Argon moves to $4 and $20. Even then it stays about two and a half times cheaper per token.

Two details shape real bills. Astra’s standard rates cover requests up to about 272,000 input tokens; past that, the entire request is billed at double the input and cached rates and 1.5 times the output rate. Argon’s pricing stays flat regardless of request size. And the gap compounds with volume: a 200,000-token input with 50,000 tokens of output costs roughly $0.90 on Argon at its introductory rates versus $4.50 on Astra.

Per-token pricing is only half of a real budget, so both testing suites also publish what a full task costs. On the intelligence suite, Argon lands at $1.99 per task against Astra’s $3.26, and it stays ahead on coding agent work and across the Vals suite.

Both models sit alongside a broader field of frontier APIs, which we track in our LLM API pricing comparison.

The cost advantage survives despite a big efficiency gap in the other direction. Argon generates 61,558 output tokens per Intelligence Index task, more than twice Astra’s 27,206, and its reasoning chains are longer than its answers on both models. The verbosity is exactly why Google could not simply match Astra’s per-task price with a lower token rate; the low rate has to absorb the extra tokens, and so far it does.

One more cost input deserves attention: caching. Artificial Analysis models cache hit rates of 89% for Argon and 99% for Astra in its task pricing, billed at each platform’s cached-input rate. Google’s API documentation lists caching for Argon as implicit only, meaning the discount arrives automatically when repeated content appears, while explicit cache management is not supported. How that lands in a monthly bill depends on how repetitive your prompts are, so the headline numbers above are best read as standardized comparisons rather than forecasts.

Coding plans: what agentic access actually costs

Most people will not touch these models through per-token billing. They will meet them inside coding agents and chat subscriptions, and there the entry prices invert the API story. Codex includes Astra on plans starting at ChatGPT Plus, $20 per month, with metered usage (5 to 45 Astra messages against 15 to 150 for Sol, with weekly limits possibly applying). To drive Argon inside Antigravity, Google’s agentic coding environment, you need Google AI Ultra at $99.99 per month, though AI Pro subscribers at $19.99 get Argon in the Gemini app and Canvas, and Antigravity itself has a free tier with basic weekly limits.

The practical math: running Astra through Codex costs $20 a month at the entry point, while running Argon through Antigravity costs $99.99. When the work moves to the API, Argon flips the bill the other way. Which side of that line you sit on depends on whether you are buying a subscription or a workload.

Speed: Tokens per Second, and Time per Task

Two measurements point in opposite directions. At task level, Astra is the quicker model on Artificial Analysis’s standardized runs: it finishes the Coding Agent workload in 29.4 minutes against Argon’s 34.5, and the full Vals suites in 27 minutes 48 seconds against Argon’s 46 minutes 33 seconds. Artificial Analysis lists Astra at 44 output tokens per second and has not yet published an equivalent figure for Argon. Service-level medians from OpenRouter show the other half of the picture: there, Argon streams at a median 51.0 tokens per second to Astra’s 37.0, with lower median latency (3.33 seconds versus 5.33). Argon is faster on the wire; Astra is faster to the finish line.

The gap is not really about network speed. It is volume: Argon generates far more output to finish the same task, and those tokens take time to produce. The effect compounds over long agent runs, where Astra also reaches its answers in fewer steps. If your workflow is interactive, or you are paying for agent time in infrastructure hours, Astra’s speed is worth real money. If you are running batch documents overnight, it rarely matters.

Which Should You Pick?

The choice comes down to the shape of your workload, and the data makes the trade-off unusually clear:

• For documents, research, and regulated industries: Argon. It leads every knowledge-work suite that matters here, from financial research to legal drafting to tax, and its refusal to guess when it does not know is the safest footing available for workflows where a wrong answer is expensive.
• For security work: Argon, without much debate. It takes both the prompt-injection and CyberBench tests by margins Astra does not come close to.
• For agents that use computers: Astra. Desktop navigation, terminal-heavy engineering, and the Time Horizon Index all run in its favor, and computer use is one of the widest gaps between the two models here.
• For cost-sensitive API volume: Argon, especially before November 30. The lowest input, output, and cache-read rates of the two, plus lower cost per task in both testing suites.
• For production teams that need a known quantity today: Astra has a longer track record, published speed data, and a cheaper subscription entry point. Argon is days old and still filling in its speed benchmarks.

Plenty of teams will simply route between them: Argon for reading and writing, Astra for clicking and coding through terminals. At these per-task costs, running both and letting the cheaper model handle the bulk of the work is a rational default rather than a luxury. For the consumer-assistant question underneath all of this, our Gemini vs ChatGPT comparison covers how the two ecosystems differ outside professional work.

Frequently Asked Questions

Is Gemini 4 Argon better than GPT-6 Astra?

They tie at 53 on the Artificial Analysis Intelligence Index, so “better” depends entirely on the task. Argon leads on legal, finance, tax, security, and honesty under uncertainty. Astra leads on computer use, terminal work, long-horizon agents, and speed. On the Vals Index, Argon ranks first of 48 models and Astra seventh.

How much do Gemini 4 Argon and GPT-6 Astra cost?

Argon is $2 per million input tokens and $10 per million output tokens during its introductory period, rising to $4 and $20 after November 30, 2026. Astra is $10 and $50. Cached input costs $0.10 for Argon and $1.00 for Astra. Per task, Argon comes in at $1.99 versus $3.26 on Artificial Analysis’s intelligence suite.

Which model is faster?

It depends on the measure. Astra finishes standardized tasks sooner: 29.4 minutes versus 34.5 on the Coding Agent Index, and 27 minutes 48 seconds versus 46 minutes 33 seconds across the Vals suite. But in live serving data, Argon streams faster, at a median 51.0 tokens per second against Astra’s 37.0. Argon’s longer reasoning traces explain the difference between those two pictures.

Which is better for coding?

Close at the top, split underneath. Argon wins the Artificial Analysis Coding Agent Index (64 versus 62) and DeepSWE v1.1 (77.9% versus 74.1%). Astra wins terminal-heavy work, including Terminal-Bench Science (62.9% versus 44.3%) and SRE Bench (56.9% versus 44.3%), and it finishes coding tasks faster.

Can I use Gemini 4 Argon today?

Yes. It is available through the Gemini API paid tier and AI Studio, in Antigravity and AI Studio for Google AI Ultra subscribers, and in the Gemini app and AI Mode in Search for Pro and Ultra subscribers. Cyber defenders can access it without cyber guardrails through the Fairwind Program. GPT-6 Astra has been available since September 3 across ChatGPT plans, Codex, the API, and AWS.

The Bottom Line

For the first time in this generation of frontier models, the cheaper option and the more capable option are genuinely different models rather than different price points on the same thing. Argon buys you the best knowledge-work scores on the market, the lowest hallucination discipline, the strongest security numbers, and a price per token that OpenAI does not currently answer. Astra buys you the best computer-use agent on the market, faster task completion, and a production track record with a cheaper subscription entry point.

If you have to pick one number to remember from this Gemini 4 Argon vs GPT-6 Astra breakdown, make it the $1.27 per-task gap on the intelligence suite: Argon delivers the same headline capability, better professional-work performance, and fewer confident mistakes for about 60% of Astra’s cost per task. Then remember the other number: 90.5% versus 52.5% on the Time Horizon Index, which is the reason Astra still wins whole categories of work that the averages hide.