Gemini 4 Argon vs GPT-6.1 Sol vs Claude Opus 5.5: Which Fits

With Gemini 4 Argon generally available since October 13, 2026, the three frontier labs each sell a current flagship on an ordinary paid API key: Google's Argon at $2 per million input tokens and $10 output until November 30, OpenAI's GPT-6.1 Sol at the same $2 and $10 for prompts up to 272K tokens, and Anthropic's Claude Opus 5.5 at $4 and $20. The prices are the easy part. The hard part is that only three benchmarks have a vendor-reported score for all three models, and on two of them Opus 5.5 has two published numbers depending on who ran it.

This post is the comparison at GA. It takes price, context, output caps, caching and the controls that break a migration from each lab's own documentation, fetched October 15. It takes benchmark scores only where at least two of the three labs reported the same benchmark at the same version, and it names the reporter in every cell. It does not pick a winner. Anthropic's own launch page says that at these capability levels benchmark margins have become a less reliable guide to real-world differences, and the footnotes we read across all three labs bear that out.

Two of our existing posts carry the arithmetic this one leans on. The $2/$10 price-point analysis works the monthly cost of a default model at that rate, and our frontier model router is the maintained decision table. What is new here is Argon's public ID, Argon's GA pricing with its December step, and the three-way benchmark overlap that did not exist before October 13.

Read the list prices first, then the footnotes, because the footnotes are where the three diverge. Google's pricing page gives Argon a standard rate of $2 input and $10 output per million tokens through November 30, 2026, and $4 and $20 from December 1; Batch and Flex are half that in both periods, and cached input is $0.10 rising to $0.20. OpenAI's model page gives GPT-6.1 Sol $2 and $10 for prompts up to 272K input tokens, with cached input at $0.10 and a cache write at $2.50; above 272K the page says input and cache rates double and output goes to 1.5 times for the full request, which is $4 in and $15 out. Anthropic's pricing page gives Opus 5.5 $4 and $20 across the whole 1M window at one rate, with cache reads at $0.20 and cache writes at $5 for five minutes or $8 for an hour.

So the two cheap models are cheap with a condition each: Argon's is a date, Sol's is a prompt size. Opus 5.5 is twice the price with neither condition. Anthropic's launch page describes its $4 and $20 as 20% less than Opus 5.

Three speed tiers sit above these and are easy to confuse. Google's Priority tier is 1.8 times standard. OpenAI's Fast mode is 2 times standard and its Ultrafast mode, added October 8 in the Responses API, is 6 times. Anthropic's Fast mode, a research preview, is $8 and $40, and is not available with the Batch API. A team comparing "fast" options is comparing a 1.8x, a 2x, a 6x and a 2x, each on a different base rate.

The failure case is budgeting Argon at October's price for a workload that starts in December, or budgeting Sol on the short rate for a retrieval pipeline that routinely passes 272K tokens of context. Both understate the real spend on paper. For the monthly arithmetic at $2 and $10, the price-point post linked above does it in full.

All three advertise about a million tokens of context. The way each counts it differs. Google's model page gives Argon 1,048,576 tokens as a unified window shared by input, thinking and output. OpenAI's model page gives Sol 1,050,000 and does not say whether that figure is unified. Anthropic's overview gives Opus 5.5 1M, at standard pricing across the whole window.

Output is where the gap is real. Google's developer guide caps a non-streaming Argon request at 32,768 tokens, down from 65,536 on Gemini 3.x, and a streaming request at 65,536; longer outputs run through a continuation loop that recent SDKs handle automatically, up to the window minus the input. Google's own model page states 65,536 as the synchronous default, so the two Google pages disagree, and we have said so in the Argon migration checklist. Sol returns up to 128,000 tokens per request. Opus 5.5 returns 128K on the synchronous Messages API and up to 300K on Message Batches with a beta header.

Modalities are a quieter difference. Argon accepts text, image, video, audio and PDF input and returns text; the model page lists the Live API and image and audio generation as not supported and computer use as supported in preview. Sol accepts text and image and returns text; its model page says audio and video are not supported, and that fine-tuning is not supported. Opus 5.5 accepts text and images and returns text. A pipeline that feeds video or audio directly to the model has one candidate among the three.

Before any score, the rule each lab applies to the other two labs' numbers. These are the three sentences that govern every rival cell in the tables that follow, and each is quoted from the lab's own page.

Google's evaluation methodology document for Argon states: "All the results for non-Gemini models are sourced from providers' self reported numbers unless otherwise mentioned below." OpenAI's GPT-6.1 Sol launch post states: "Evaluations of competitor models were taken from publicly available reports." Anthropic's Opus 5.5 system card states: "Competitor figures are drawn from the respective developers' published system cards or benchmark leaderboards."

The practical consequence: when Google prints an Opus 5.5 score, it is usually Anthropic's number or a leaderboard's, and when Google ran the rival itself it says so. We have kept that distinction in every cell as "self" or "relayed", because the two kinds of number are not the same evidence. A relayed score with a different harness, effort level or fallback setting can be several points away from the lab's own figure, and the next section shows exactly that happening.

Across Google's model card, OpenAI's launch post and Anthropic's launch page and system card, exactly three benchmarks carry a self-reported score for Argon, Sol and Opus 5.5 on the same benchmark: Terminal-Bench Science 0.1, AutomationBench, and DeepSWE v1.1. The table gives each lab's own number for its own model, then the second published figure for Opus 5.5 where one exists.

Take Terminal-Bench Science first, because it shows the whole problem in one row. Google reports Argon at 57.6% and discloses that the run was self-computed with a six-times verifier timeout. OpenAI's chart puts Sol at 57.0% at max effort. Anthropic reports Opus 5.5 at 58.7% at max effort with a standard error it gives as 3.5 to 5 points. Google and OpenAI both print 63.3% for Opus 5.5: Google cites the official public leaderboard, and OpenAI's chart labels that run as Opus 5.5 with fallbacks. Three models within two points of each other, a fourth figure nearly five points higher for one of them, and an error bar wider than the spread. There is no ranking in that row.

AutomationBench is the row with the widest gap, and the gap has an explanation in the footnotes. Google reports Argon at 51.3% and says it ranks first; OpenAI and Anthropic both name release 1.0.6, and Google calls its set a private one sourced from Zapier's public leaderboard without a release number. Anthropic reports Opus 5.5 at 40.0% and states that the run was performed without fallback models, so safeguard interventions were counted as failures. Google and OpenAI relay 42.5% with fallbacks. OpenAI's text puts Sol 2.2 points above Opus 5.5 at medium effort, which on its own chart is 31.7% against 29.5%.

DeepSWE v1.1 is the cleanest of the three and still not clean. Google's 77.9% for Argon was self-computed in a mini-swe agent harness; OpenAI's 75.2% for Sol is at high effort, with medium and max both lower; Anthropic's 74.2% for Opus 5.5 is a five-trial average from its system card, and Google relays the same number. Each lab ran its own harness.

The honest reading of all three rows is the one Anthropic's launch page offers about its own results: at this level, benchmark margins are a weak guide to real differences. The useful signal is not which bar is longest. It is that no lab's own page shows its model leading every row it chose (OpenAI and Anthropic both put GPT-6 Astra ahead on Terminal-Bench Science, and Google's card puts Argon behind Astra and Opus 5.5 there), and that the one model all three labs printed a number for, Opus 5.5, has two published figures on two of the three rows.

A second tier of benchmarks has scores for two of the three models. Most are Google-reported pairs of Argon against Opus 5.5, because Google's model card is the longest of the three documents and prints a rival column for almost every row. OpenAI reported no score on CWE-Bench, Terminal-Bench 4.0, Chartography or any other benchmark on Google's card, so Sol is absent from most of this tier.

Four rows are worth the reader's time. On Terminal-Bench 4.0, Google reports Argon at 57.4% and relays Opus 5.5 at 66.4% from the public leaderboard, which matches Anthropic's own figure at xhigh effort with a 2.6-point standard error. On CWE-Bench v1, Google reports Argon at 68.0%, tied for first, and relays Opus 5.5 at 67.0% from the leaderboard; Anthropic's launch page and system card carry no CWE-Bench figure, so that 67% is Google's attribution alone. On Chartography without tools, Google reports Argon at 71.6% and Anthropic reports Opus 5.5 at 64.4%; Anthropic's launch-page figure of 89.0% is with tools and cannot be set against Argon's. And on OpenAI's GDP.pdf chart, Sol scores 32.0% at its best effort against a relayed 28.8% for Opus 5.5 with fallbacks, a comparison only OpenAI reports.

Google's card also carries a long list of Argon-versus-Opus rows on benchmarks nobody else reported: Vals Index, Vals Finance Agent v2, Harvey's legal agent benchmark, FrontierSWE v2, Vibe Code Bench, PostTrainBench, LABBench 2, RiemannBench, GraphWalks, Agent's Last Exam and LVBench. Argon leads most of them and trails on FrontierSWE v2 (55.0% against a relayed 62.3%) and PostTrainBench (45.3% against 49.3%, both run by Google). We have not tabulated them, for the reason the previous section gave: a single reporter choosing the benchmarks is not a comparison, whichever way the rows point.

For a team actually switching, the behaviour differences matter more than two points on a benchmark. All three models have moved away from sampling parameters, in three different ways. Google's guide says Argon accepts temperature, top_p, top_k and candidate_count and ignores them. OpenAI's latest-model guide tells Sol users to remove temperature, top_p and top_logprobs when reasoning effort is set, which on Sol it always is, and does not say what happens if they stay. Anthropic's migration guide says a non-default temperature, top_p or top_k on Opus 5.5 returns a 400 error. Same intent, three failure modes: silent, unspecified, and loud.

Reasoning control also differs in shape. Argon exposes three thinking levels, low, medium and high, with medium the default. Sol exposes five reasoning efforts, low, medium, high, xhigh and max, with medium the default and none and minimal not supported. Opus 5.5 exposes the same five effort names with medium the default, and rejects any attempt to disable thinking or set a manual budget. Three labs use the same word "medium" for their default, and nothing on any of the three pages says the three defaults spend the same amount of thinking.

One row deserves a sentence of its own. Sol's model page says to use the Responses API for tool calling and that Chat Completions is supported without tool calling. A team on the older endpoint with a tool-using agent cannot move to Sol by changing the model string; it has to move endpoints too. Opus 5.5's equivalent trap is forced tool use: tool_choice set to any or a named tool returns a 400, so an agent loop that forces a tool on the first turn needs rewriting. Argon's is the thought-signature rule, where history-trimming code that drops unknown fields breaks multi-turn tool use. Each is a small change and each is a production outage if found late.

A fit judgement is not a winner. It is the set of conditions under which each model's documented properties, not its benchmark position, make it the obvious candidate. These four routes come from the pages above and from nothing else.

Where a job does not fall into one of those routes, the honest instruction is to run the three on your own evaluation set at comparable effort, with the sampling parameters removed, and pick on your result. That is the method our frontier model router applies, and it is the work our AI transformation team does when a client's workload does not match any vendor's chosen benchmark.

Cost per task. OpenAI's launch post states that on Terminal-Bench Science at max effort Sol costs $5.47 per task on average against $23.21 for Opus 5.5, and its AutomationBench chart gives $0.19 against $0.65 at medium. Those are OpenAI's figures from OpenAI's runs with OpenAI's token counts, and we have not reused them as our own arithmetic. A cost-per-task number is only transferable with the token mix and effort level attached, and the post does not give enough of either for a reader to reproduce it.

Rate limits. OpenAI publishes Sol's per-tier limits on its model page: 5,000 requests and a million tokens a minute on the Build tier, rising to 15,000 and 40 million on Grow. We did not check Argon's or Opus 5.5's per-tier limits for this post, so there is no rate-limit row.

GPT-6 Astra. Both Google and OpenAI print Astra at 68.1% on Terminal-Bench Science; Anthropic prints 64.6%, attributed to OpenAI. We could not resolve the difference from the pages read, so Astra does not appear in our tables. Our GPT-6.1 Sol upgrade guide and Opus 5.5 launch record carry each model's full vendor benchmark set, including the rows no other lab reported.