Access before benchmarks
Google is rolling Argon out through Fairwind before broader availability. Verify your account access before designing a production dependency. R1 weights are available under MIT; a self-hosted deployment still needs a serving engine, capacity planning and operational ownership. Preview access and downloadable weights answer different deployment questions.
Context capacity is not recall accuracy
The Argon announcement describes a 1M output-token limit. It does not establish the 2M input-context claim in this comparison request. The unconfirmed input-context capacity is left blank in our table. Build a retrieval test with known file references and distractor documents before paying to send an entire repository.
Published price versus historical price
Argon introductory API rates are $2 input and $10 output per million tokens. The original R1 release quoted $0.55 uncached input and $2.19 output; those are historical rates. DeepSeek later announced retirement of legacy API names. Budget a current R1 host quote or GPU rental rather than assuming the old official endpoint is available.
Measure an agent loop before scaling
Run the same patch task through your actual agent and serving adapter. Record tool-call parsing failures, retries, test pass rate and accepted-patch cost. For self-hosting include idle GPU time and engineering effort; for preview APIs include access constraints. No same-harness tool reliability or long-context recall score is asserted here.