Gemini 4 Argon: Benchmarks, Costs, and Choosing the Right Model
Argon’s benchmark gains, task costs, and which provider fits your work.
Google announced Gemini 4 Argon on September 30 for software engineering, professional knowledge work, and defensive cybersecurity. Access starts with trusted cyber defenders through Fairwind. Paid API customers and Google AI Ultra subscribers are next, without a confirmed public rollout date. Google
Benchmarks help narrow the choice. Access, included tools, and total cost decide whether a model fits your work.
What the tests show
This October 1 snapshot combines independent evaluations with OpenAI’s reported Sol result. AA Index covers reasoning, knowledge, and agentic tasks; AutomationBench tests business workflows; Vals Index weights finance, coding, legal, and tax work; Terminal-Bench tests demanding terminal tasks. Compare within each test, not across its scores. AA, Zapier, Vals, Terminal-Bench
| Gemini 4 Argon | 53 | 51.29% | 68.90% | 57.58% |
|---|---|---|---|---|
| GPT-6 Astra | 53 | 41.40% | 63.13% | 59.60% |
| GPT-6.1 Sol | 52 | 36.10%† | 61.15% | 55.05% |
| Claude Opus 5.5* | 58 | 42.47% | 66.97% | 65.15% |
Argon leads this group on business workflows and Vals Index. Opus leads AA Index and terminal work. Sol approaches Astra’s broad score at lower task cost.
Effort controls reasoning computation: AA uses Argon High and peers Max. *Opus includes provider fallbacks after refusals. †Sol’s AutomationBench result is OpenAI-reported, absent from Zapier’s board, and uses a research/API setup that may differ from production.
Tap, hover, or focus a point. Zoom with controls, wheel, or pinch; drag to pan.
Artificial Analysis Intelligence Index
USD per task · logarithmic cost axis · Argon introductory rates
Zapier AutomationBench
USD per task · logarithmic cost axis · Argon standard rates
Higher scores are better; farther left costs less. The panels measure different workloads. Dashed Sol is OpenAI-reported; other workflow curves come from Zapier.
A token price is only part of the bill
Argon’s introductory token rates match Sol’s, but task costs differ. Token usage, caching, tool fees, and retries affect the bill. AA compares Argon High against Sol Max at introductory rates; its reported multiple appears below. These are weighted evaluation costs, not customer invoices.
| Argon introductory | $2 | $10 | $1.99 | AA reported about 2.7 times as much per test task as Sol, despite matching token rates. |
|---|---|---|---|---|
| Argon standard | $4 | $20 | $3.98 | Estimated at standard rates; assumes unchanged token usage. |
| GPT-6 Astra | $10 | $50 | $3.26 | Measured at Max effort. |
| GPT-6.1 Sol | $2 | $10 | $0.72 | Measured at Max effort. |
| Opus 5.5* | $4 | $20 | $5.98 | Measured at Max effort, with fallback. |
Token rates are USD per million. AA task costs are weighted index averages. Standard-rate Argon cost is derived assuming unchanged usage. Google, Astra, Sol, Opus
AA’s chart uses introductory Argon pricing. Zapier’s leaderboard and our workflow chart use standard rates; its footnote separately lists promotional costs. Pricing notes
Choose the offering that fits your work
Sol is currently the workhorse for frequent coding and document work; Astra and Opus are candidates for complex problems where a stronger solution saves rework. Argon’s workflow results make it worth testing for business tasks spanning tools and documents. The offer can change even when the model does not. OpenAI, Anthropic, Google
| GPT-6.1 Sol | The everyday workhorse: coding, documents, and repeated technical work. | Pro 200: USD 200/month. For accounts notified of 20× → 10× usage, the lower allowance applies after October 29, 2026; eligible accounts retain the previous allowance through that date. |
|---|---|---|
| GPT-6 Astra | Complex analysis, difficult reasoning, and demanding deliverables. | Use when a better solution can justify higher spending. |
| Claude Opus 5.5 | Complex problems and sustained engineering across a codebase. | Faster output generation than Astra and competitive capability for demanding work; leads Astra in the benchmarks shown, with fallback included. |
| Gemini 4 Argon | Multi-app business workflows: reconcile records, update systems, and prepare decision documents. | Its workflow and economic-task results support testing this use case. Broad access remains pending. |
AA measures Opus output at 92.2 tokens per second versus Astra’s 51.3. That means faster text generation, not necessarily faster task completion: thinking and tool time are excluded.
Opus cache reads cost $0.20 per million tokens; Sol’s cost $0.10. Both reward reusing context, but cache writes cost extra.
Claude Code’s paid weekly allowance is 25% above its pre-May baseline, replacing the temporary 50% boost. Across four fully used weeks, that adds roughly one old-baseline week of capacity.
Among these providers, OpenAI alone supports device OAuth for third-party harnesses, without client spoofing. Some Claude users reportedly spoof Claude Code in custom harnesses to reuse subscription OAuth—a workaround that breaches Anthropic’s stated terms. Google’s Antigravity CLI (agy) supports browser and remote OAuth login, but prohibits third-party harnesses using its login and directs them to API keys.
The recent Pro allowance change above could alter OpenAI’s value for heavy users: a lower allowance at the same price can mean earlier limits or extra credit purchases. OpenAI confirms the date and reduction; the exact multipliers should be checked against your account notice. Claude’s shared app/Code limits also matter if you use both heavily. Argon cannot replace your current service until you can access it. These are reasons to reconsider an offer, regardless of benchmark rank.
The AI battle rages on as adoption grows. Chips, power, and data-center construction constrain how quickly providers can expand; infrastructure takes years while demand changes faster. That does not establish why any particular allowance changed. It does mean today’s best offer is provisional: compare the work you can finish before hitting limits, the cost of continuing, and the rework each model leaves you.
It remains to be seen how Google packages Argon to compete with OpenAI and Anthropic. Its own data centers and custom TPUs could give it room to offer more generous usage—perhaps an experience comparable to OpenAI’s original 20× tier. That is a possibility, not an announced allowance. Alternatively, tighter offers may signal a retreat from subsidized AI usage. For heavy users, that would be a painful shift: less experimentation and more money spent to finish the same work.
