GPT-6.1 Sol, Gemini 4 Argon and Claude Opus 5.5 compared

GPT-6.1 Sol, Gemini 4 Argon and Claude Opus 5.5 differ in pricing, long-context terms and supported input formats. Sol and Argon currently share the same base API prices, but Argon’s current rates apply only through November 30, 2026. Argon accepts audio and video, while Opus 5.5 applies standard rates across its full million-token context window.

The figures below are published standard rates for the providers’ direct APIs, in US dollars per million tokens. They exclude consumer subscriptions, faster service tiers and additional services. Cache-read prices apply to reused data; creating a cache may incur a separate charge.

• GPT-6.1 Sol: $2 for input, $10 for output and $0.10 for cache reads. These base rates apply to requests with up to and including 272,000 input tokens.
• Gemini 4 Argon: $2 for input, $10 for output and $0.10 for cache reads through November 30, 2026. Published rates from December 1, 2026 are $4, $20 and $0.20 respectively.
• Claude Opus 5.5: $4 for input, $20 for output and $0.20 for cache reads. Anthropic specifies standard pricing across the full million-token context window.

For requests exceeding 272,000 input tokens, Sol’s rates are $4 per million input tokens, $15 per million output tokens and $0.20 per million tokens for cache reads. These higher rates apply to the entire request. Argon’s published pricing table does not specify a separate long-context surcharge.

A hypothetical request containing 100,000 input tokens and 10,000 billable output tokens, including reasoning tokens, would cost $0.30 with Sol or Argon at current rates and $0.60 with Opus 5.5. This calculation excludes caching, tools and additional surcharges. It is a comparison based on price lists: completing the same task may require different token counts and retries across models.

OpenAI documents a 1,050,000-token context window for GPT-6.1 Sol and a maximum output of 128,000 tokens. Claude Opus 5.5 specifies a million-token context window and a standard output limit of 128,000 tokens. Window size alone does not establish how accurately a model will find the required information in a large document collection.

Gemini 4 Argon’s input, reasoning and response share a single window of 1,048,576 tokens. Its standard generation limit is 65,536 tokens. The long decode mechanism can exceed that limit using the remaining space in the window; this requires background execution or a continuation mechanism. The advertised response of up to a million tokens must fit within the shared window alongside the input and reasoning.

Gemini 4 Argon’s model card lists text, images, audio and video as inputs, with text output. The model cards for Sol and Opus 5.5 list text and image inputs, with text output. Argon is worth including in an initial evaluation for meeting recordings or video analysis because the model directly supports those input formats.

OpenAI reports that Sol at medium reasoning effort exceeds Opus 5.5 by 2.2 percentage points on AutomationBench at roughly one-third of the cost. Google reports 77.9% for Argon on DeepSWE v1.1, which evaluates extended software engineering tasks. Anthropic reports 66.4% for Opus 5.5 on Terminal-Bench 4.0 at xhigh effort.

These figures refer to different tests and do not establish an overall ranking. Google discloses its use of a custom agent harness for DeepSWE and published competitor results. Anthropic gives a standard error of ±2.6 percentage points for its Terminal-Bench result and notes that safeguards can trigger a switch to other models on some tasks. These conditions need to be considered when assessing how the results apply to a production project.

Sol is worth evaluating for text and coding tasks when base cost matters and requests stay within its pricing threshold. Argon deserves evaluation for tasks involving audio, video or unusually long generation; budgets should also account for the rates effective December 1, 2026. Opus 5.5 is worth comparing on large documents and extended agent tasks, weighing result quality against cost at the required context length.

Before making a final choice, evaluate the models on a small set of your own tasks with acceptance criteria defined in advance. Measure completion rate, time to a usable result, total cost and the amount of manual correction required. These measures can help establish which model fits a particular workflow and where its advantages justify the expense.