GPT-6.1 Sol vs Gemini 4 Argon: buy now or wait?

Two $2/$10 frontier models present very different buying decisions

OpenAI's GPT-6.1 Sol and Google's Gemini 4 Argon appeared one day apart with the same announced introductory rate: $2 per million input tokens and $10 per million output tokens. That superficial price match hides the most important difference. GPT-6.1 Sol is available through the OpenAI API now. Argon is initially limited to trusted cyber defenders and testers, while Google says paid API customers and Google AI Ultra subscribers will gain access later. A model that can be evaluated and deployed today is not economically interchangeable with one whose broad launch date, API documentation, and input context limit remain unpublished.

GPT-6.1 Sol is also a replacement, not merely another tier. OpenAI shipped it seven days after GPT-6 Sol, keeping the standard input and output rates while cutting cached input from $0.20 to $0.10 per million tokens. The official model page confirms a 1.05-million-token context window, 128,000 maximum output tokens, text and image input, and five reasoning settings from low through max. Requests above 272,000 input tokens carry a two-times input and cache rate and a 1.5-times output rate for the full request, so the headline price is not the price of every long-context job.

Argon is a more ambitious but less complete proposition. Google announces a one-million-token maximum output, up from 64,000 on its previous frontier generation, plus introductory $2/$10 rates and cached input at a 95% discount. Google says those rates will later double to $4/$20. The launch post does not publish the input context window or a general API model identifier, and the expected developer documentation page is not live. Buyers should record those omissions as unknowns rather than filling them with specifications from earlier Gemini models.

Independent testing makes GPT-6.1 Sol an unusually strong value case

Artificial Analysis scores GPT-6.1 Sol at 52 on Intelligence Index v4.3.2 at maximum effort, four points above GPT-6 Sol and one point below GPT-6 Astra. Its measured cost per Intelligence Index task was $0.72, compared with $1.05 for GPT-6 Sol and $3.26 for Astra. Those figures come from an independent evaluator using a fixed benchmark mix and current API prices. They do not guarantee the same ratio on a private codebase or document workflow, but they are stronger purchasing evidence than a vendor launch table alone.

The coding evidence is similarly useful and slightly untidy. Artificial Analysis reports that GPT-6.1 Sol gains three points over GPT-6 Sol at maximum effort on its Coding Agent Index. Its xhigh setting scored one point above Astra while costing less than 15% as much per evaluated task. Max did not beat xhigh in that test, which is a practical warning against treating the largest reasoning budget as the default. OpenAI reports additional gains on DeepSWE, OSWorld, AutomationBench, GDP.pdf, and scientific terminal tasks, but those exact comparisons are vendor-run even where the underlying benchmarks are public.

The limitation is token use. Artificial Analysis measured GPT-6.1 Sol using roughly 10% to 30% more output tokens than GPT-6 Sol across effort settings. Better results made low and medium Pareto-efficient in the evaluator's analysis, but a workload with long answers or repeated retries can still reverse the expected saving. Production tests should therefore compare at least medium and xhigh, record cache hits and writes separately, and calculate cost per accepted task. The upgrade is compelling; switching every route to max effort is not.

Argon's early signal is credible, but most of the evidence is still Google's

Google reports that Argon scores 77.9% on DeepSWE v1.1, 51.3% on AutomationBench, and 91.7% on the long-video LVBench evaluation. It also describes large internal code migrations and optimization work, including Rust ports and data-center memory savings. These examples suggest where Google expects the model to excel: long-horizon engineering, enterprise knowledge work, visual analysis, and defensive cybersecurity. They remain vendor-reported results produced or selected for a launch, not independent proof that Argon will outperform available models under another harness.

Arena supplies the first independent counterweight. Its current release summary ranks Gemini 4 Argon high first for text preference. The Agent Arena places Argon high eighth overall with a 7.92% estimated success rate and a wide confidence interval, while showing an estimated $0.65 per task. That early agent result is promising but not dominant: Claude Fable 5.1, Claude Opus 5.5, GPT-6 Astra, and the older GPT-6 Sol appear above it in the current table. Different Arena sections measure different behavior, so first place in text preference should not be reported as first place in autonomous work.

Artificial Analysis has created an Argon model page but does not yet expose a completed independent Intelligence Index result. Combined with restricted access, missing API documentation, and unpublished input context, that leaves too many operational variables for a production recommendation. The correct status is watch and evaluate when access arrives. Teams in Google's trusted defender cohort have a reason to test now; ordinary API buyers should not redesign routing or procurement around an announced model they cannot yet run.

Adopt Sol selectively; keep one clean Argon evaluation slot

For teams already using GPT-6 Sol, the smallest sensible change is a direct GPT-6.1 Sol trial on completed work. Use twenty to fifty tasks that include coding, computer use, long documents, tool failures, and deliverables that previously needed correction. Run the old route and GPT-6.1 at medium and xhigh with equivalent tools. Record accepted completion, omitted requirements, reviewer time, retries, input, cached input, cache writes, output tokens, latency, and billed cost. Promote the new model only where the outcome improves, even though the public evidence makes that result likely.

Keep Astra as an escalation route rather than deleting it from the evaluation. GPT-6.1 nearly matches Astra on the broad independent index, but OpenAI still reports a higher Astra ceiling in difficult scientific work and its own safety card places Astra ahead on several cyber and biology capability evaluations. Those are OpenAI-run comparisons, yet they identify the kinds of edge cases worth preserving in an owned test set. A one-point composite gap does not mean every capability differs by one point.

Reserve one challenger slot for Argon when general API access and documentation arrive. Test it against the winning GPT-6.1 route on the same tasks, with particular attention to long trajectories, output control, tool reliability, latency, safety interruptions, and the date at which introductory pricing ends. Until then, the buying rule is deliberately boring: adopt GPT-6.1 Sol where it beats the existing route, retain Astra only for proven hard cases, and treat Gemini 4 Argon as a credible upcoming test—not as a deployable default.

• Available value default: GPT-6.1 Sol, tested at medium and xhigh rather than max alone.
• Premium escape hatch: GPT-6 Astra only where owned evaluations justify the additional cost.
• Watchlist: Gemini 4 Argon until broad API access, input-context documentation, and fuller independent results arrive.
• Vendor-reported: OpenAI and Google launch benchmarks; independently measured: Artificial Analysis Sol results and Arena's early Argon rankings.