OpenAI GPT-6.1 Sol, GPT-6 Astra Ultrafast, and Google Gemini 4 Argon are intensifying competition among high-end Artificial Intelligence (AI) models. For you, the important issue is not which model tops a leaderboard, but which one is actually available, fits your work, and comes with predictable costs. The three offerings differ considerably on all three points.
What is available today
GPT-6.1 Sol has been available since September 29, 2026, through the OpenAI Application Programming Interface (API), ChatGPT Work, and Codex. An API is an interface used to connect a model to another service; for most users, ChatGPT Work and Codex are the more direct routes. According to the initial report on GPT-6.1 Sol, the model approaches GPT-6 Astra on coding, computer use, and professional work.
GPT-6 Astra Ultrafast is also available through the OpenAI API and to eligible ChatGPT Work and Codex users. The sources do not explain exactly who qualifies. “Ultrafast” is not a separate model generation, but an accelerated mode for Astra that produces tokens up to eight times faster than Astra Standard, according to Nvidia’s account of the acceleration.
Gemini 4 Argon, by contrast, is not yet a generally available product. Google is initially providing access to selected cyber defenders in its Fairwind Program and to internal teams. The Verge describes the launch as a deliberately limited test phase during which Google plans to strengthen safeguards against misuse, prompt injection attacks, and model misalignment. Prompt injection refers to manipulated instructions designed to make an AI system take unintended actions.
For people in Switzerland, the immediate picture is fairly plain: Sol is accessible through the listed OpenAI services if the relevant account has access, while Astra Ultrafast also depends on an undefined eligibility requirement. The sources provide neither a date for broad Argon access nor specific information about availability in Switzerland. They also offer no concrete details about supported languages, data location, or compliance with Swiss privacy requirements.
What the models cost
GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens through the API. Tokens are small units into which a model divides prompts and responses, so their number does not directly equal a word count. Cached input costs $0.10 per million tokens, which can matter when instructions or document sections are reused frequently.
OpenAI positions Sol as a substantially cheaper alternative to Astra. According to the provider, its token rates are one-fifth of Astra’s while performance on certain coding, computer-use, and professional tasks comes close to Astra. However, the supplied sources do not give a complete current price schedule for Astra or Astra Ultrafast. You can therefore understand the claimed price difference, but not calculate a reliable total cost for Ultrafast.
Google has already announced API pricing for Gemini 4 Argon despite its restricted access. During an introductory period, the rates are also set at $2 per million input tokens and $10 per million output tokens, with a 95 percent discount for cached input. After that period, regular prices will rise to $4 and $20 respectively, according to The Decoder’s pricing overview.
These figures are API usage rates, not complete monthly costs for ChatGPT Work, Codex, or other applications. The sources do not provide subscription prices, included allowances, or possible additional fees. For you, this means that identical per-token prices do not necessarily make two models equally expensive: response length, repeated calls, and access through paid services all affect the final bill.
What they can do in practice
Sol targets work in which an AI does more than return a single answer. The main report highlights agentic coding, computer use, and professional knowledge work. “Agentic” means that the model pursues a goal over several steps, uses tools, and checks the results. The short source summary does not include independent detailed tests of the claimed near-Astra performance.
Astra Ultrafast focuses less on adding a new capability than on reducing waiting time. Nvidia gives the example of a coding agent that writes code, calls a tool, checks the result, and then fixes errors. When this cycle repeats many times, token generation that is up to eight times faster can reduce pauses between steps. For one short answer, the difference may matter less than it does during a lengthy automated workflow.
Argon is intended to cover coding, research, writing, enterprise knowledge work, and cyber defense. Reports say Google employees already use it for debugging and large code migrations. Ars Technica provides concrete internal examples: Argon reportedly helped migrate C and C++ code to Rust, including more than 800,000 lines in the Zircon kernel, and used telemetry data to save 300 TiB of memory across Google’s data centers. These results come from Google and were not independently verified in the supplied material.
For less technical roles, Argon’s handling of visual material may be more relevant. The model is said to analyze long videos and charts and to assist with complex workflows in areas such as legal and financial work. One practical professional example would be reviewing a long video or a collection of charts before producing a summary. For now, however, that is a stated use case rather than a feature you can test through general access.
Argon’s announced output limit of one million tokens is also notable, compared with 64,000 tokens in previous Gemini models. Google says the larger limit allows extensive tasks to be completed in one run. A higher limit does not guarantee a more accurate answer or deeper reasoning; it primarily permits much longer outputs and workflows.
Pros and Cons of the new model choices
Pros:
- Lower entry price – According to OpenAI, Sol offers selected Astra-like capabilities at one-fifth of Astra’s token rates.
- Less waiting – Astra Ultrafast can save time when workflows involve frequent tool calls and repeated steps.
- Long tasks – Argon’s one-million-token output limit could support extensive workflows without an early cutoff.
- More choice – OpenAI and Google emphasize different combinations of cost, speed, length, and controlled access.
Cons:
- Restricted access – Argon is limited to internal teams and a small group of selected cybersecurity partners.
- Incomplete costs – The sources do not provide full pricing for Astra Ultrafast or the relevant end-user services.
- Provider benchmarks – Many performance claims come from the companies themselves and cannot replace testing with your own work.
- Security risks – Google’s phased release indicates that advanced cybersecurity capabilities could also be misused.
The benchmark rankings do not produce one consistent picture. Google and a report on Argon’s benchmark results place the model ahead of GPT-6 Astra and Claude Opus 5.5 in most tests. The Decoder instead describes Argon as being on Astra’s level overall and behind Claude Opus 5.5 in the Artificial Analysis index, while TechCrunch points to Google’s lead in the Vals index. These statements do not necessarily conflict on every individual test, but they demonstrate how the chosen benchmark and aggregation method can change the outcome.
What this means for you
If you are getting started, GPT-6.1 Sol is the most obvious first model to assess because it is already offered through ChatGPT Work and Codex and its API rates are known. Use a clearly defined task whose result you can judge, such as explaining an error or structuring an existing work document. Compare not only answer quality, but also completion time, required corrections, and the actual number of tokens consumed.
If you are an advanced user, a comparison based on repeatable workflows will reveal more. Astra Ultrafast is most relevant when an agent calls tools frequently and each delay lengthens the entire process, while Sol is aimed more directly at a lower balance of performance and token cost. Argon remains something to monitor unless you belong to the limited group of Fairwind partners.
Argon’s limited release should not be mistaken for a lack of capability. TechCrunch reports that Google trained the model specifically for defensive cybersecurity and says it can autonomously find, validate, and patch critical software vulnerabilities. Those same capabilities help explain why Google is restricting access while it tests additional safeguards before a wider release.
The current competition does not give you one clearly superior model, but three different priorities: Sol emphasizes a comparatively affordable level of performance, Astra Ultrafast focuses on speed, and Argon targets long, complex workflows with a cautious release process. Public pricing makes Sol easier to budget for immediately, while Argon’s introductory rate remains theoretical for most people. The central open questions are how these models perform beyond provider benchmarks and which security, privacy, and access restrictions will apply when their use expands.
Sources
- OpenAI Releases GPT-6.1 Sol: Near-Astra Coding and Computer Use at One-Fifth of Astra’s Token Price – MarkTechPost, 2026-09-30
- How NVIDIA GPUs Help Accelerate OpenAI’s GPT-6 Astra Ultrafast – NVIDIA Blog, 2026-10-01
- Google releases Gemini 4 Argon, called its most powerful model yet – TechCrunch, 2026-09-30
- Google Gemini 4 Argon schliesst zur KI-Spitze auf, bleibt aber hinter Anthropics Claude Opus 5.5 – The Decoder, 2026-09-30
- Google DeepMind Unveils Gemini 4 Argon with 1M Output Tokens for Coding, Knowledge Work and Cyber Defense – MarkTechPost, 2026-09-30
- Google announces Gemini 4 and says it’s so capable that only trusted cyber defenders can have it right now – The Verge, 2026-09-30
- Google announces Gemini 4 Argon AI model, but you can’t use it yet – Ars Technica, 2026-09-30
Image: Matheus Bertelli via Pexels