Gemini 4 Argon vs GPT-6.1 Sol: Which is better for coding?

Google's Gemini 4 Argon pairs a 1M-token output limit with DeepSWE 77.9% and a $2/$10 API launch price. Here's what its staged rollout means for coding and enterprise teams.

Google announced Gemini 4 Argon on September 30, 2026, positioning it as a frontier model for complex software engineering, knowledge work, and defensive cybersecurity. The most striking spec is not a leaderboard score: Argon can generate up to 1 million output tokens, an unusually large runway for multi-step work in a single trajectory. Google says it will expand access gradually, beginning with trusted cyber defenders and testers; broad developer, enterprise, and consumer access is not yet live.

That makes the comparison with OpenAI's GPT-6.1 Sol less about declaring a winner and more about evaluating two different promises: Argon's very long execution horizon and low announced introductory token rates versus Sol's established coding-oriented position. Model cards, identical benchmark runs, and public API availability will matter before teams can make a confident migration decision.

Benchmark caveat: these are publisher-reported results, not a controlled head-to-head. Benchmarks can differ by harness, prompting, tool access, sample selection, and scoring. DeepSWE v1.1 and CWE-bench v1 are different from SWE-bench Verified and CWE-bench v0; their percentages should not be treated as interchangeable.

Most model comparisons focus on input context: how much material a model can read. Google is emphasizing output capacity, raising Argon's output limit from 64K to 1M tokens. In principle, that gives an agent much more room to reason, draft, inspect intermediate work, or produce a long multi-step result without stopping to hand off state.

The practical question is whether teams should actually ask a model to produce a million tokens. Usually not. Long generations can increase elapsed time, cost, and the surface area for errors. For production, developers should set explicit budgets, stream partial results, checkpoint agent state, and require tests or human review before applying code changes. Argon's expanded ceiling is most compelling when the task genuinely benefits from one continuous trajectory—not as a default generation target.

The launch article does not specify a generally available API, complete input-context limit, rate limits, latency, or the exact model identifier. Treat those as unanswered questions rather than assuming that the announced output ceiling is already usable in a production integration.

Google reports 77.9% on DeepSWE v1.1, describing Argon as state of the art on this long-horizon software engineering evaluation. The company also recounts internal work on code migrations and optimization, including an Argon-assisted Rust rewrite of 32,000 lines of SIMD code in libgav1. Google says the resulting safe Rust decoder ran 2.7× faster than the existing Rust port with identical video output. This is a Google-reported internal case study, not an independently reproduced result or a guarantee for arbitrary repositories.

For developers, long-horizon competence can help with multi-file changes, debugging chains of failures, and migrations that need repeated profiling and verification. But a benchmark score alone does not tell you whether the agent will preserve repository conventions, avoid risky edits, or recover cleanly from a failed test. Evaluate it on your own issues and require branch isolation, test execution, diffs, and approval gates.

GPT-6.1 Sol and Argon should be tested using the same repository tasks, tool permissions, time budget, and acceptance criteria. Until a reliable public apples-to-apples evaluation exists, don't infer that either DeepSWE's score or a result from a different coding benchmark settles the comparison.

Google says Argon leads the Vals Index, an economic-impact-weighted evaluation spanning finance, coding, legal, and tax; reports 51.3% on AutomationBench; and cites 91.7% on LVBench for long-video understanding. It also describes results on Vals Finance Agent v2 and Harvey's Legal Agent Benchmark without publishing scores in the launch post. These claims suggest broader enterprise ambitions, but the announcement does not provide enough matching GPT-6.1 Sol results to make a fair numerical comparison.

Cybersecurity is a particular focus. Google reports a 68% tie for first place on CWE-bench v1, and says Argon can find, validate, and patch vulnerabilities. The release also describes a phased access program for trusted defenders and extensive safeguards, including misuse protections, prompt-injection testing, misalignment monitoring, and hardened evaluation environments. This is high-impact capability: security teams should use isolated, authorized test environments, review every proposed patch, and never provide production credentials to an unvetted agent.

Google lists introductory pricing of $2 per million input tokens and $10 per million output tokens, with cached input tokens at a 95% discount from the input price. That is a useful initial signal, but not a complete cost comparison: Argon's long output ceiling can make output consumption substantial, and end-to-end agent costs also include retries, tools, and verification.

The release says Argon is rolling out to trusted cyber defenders through the Fairwind Program and that wider availability will follow, starting with paid API customers and Google AI Ultra subscribers. No public launch date, rate limits, or complete usage terms are specified in the announcement. For now, the pricing is announced rather than a guarantee that every developer can provision the model today.

Friday is the workflow layer around models: teams can evaluate models against actual tasks, keep human approval in the loop, and route work through the tools and processes they already use. Argon's long output budget could be useful for substantial repository analysis or multi-stage knowledge work once access opens; its cybersecurity strengths call for especially clear permissions and audit trails.

Gemini 4 Argon is a consequential announcement: a reported 77.9% DeepSWE v1.1 result, a 1M-token output limit, broad knowledge-work claims, and focused cyber capabilities make it worth evaluating. Its announced input/output rates look straightforward, although long generations mean teams must control output budgets.

The biggest near-term limitation is access. Argon is still in a phased rollout, and important details such as public API identifiers, rate limits, latency, and independently comparable results remain unavailable in the cited announcement. GPT-6.1 Sol remains a useful incumbent to benchmark rather than a straw opponent. Keep your current production setup, build a matched evaluation, and revisit the decision when Argon access and documentation are concrete.

Want to try GPT-6 and 300+ other models today?

Run GPT-6.1 Sol, Gemini, Claude, and hundreds of other frontier models side-by-side in Friday Code — with real agent workflows, local execution, and full control over your data.

Is Gemini 4 Argon available to developers now?

Not generally, according to Google's September 30 announcement. It is rolling out first to trusted cyber defenders and testers, with broader availability planned later, starting with paid API customers and Google AI Ultra subscribers.

What is Gemini 4 Argon's context window?

Google highlights a 1 million-token output limit. The launch announcement does not fully specify the input context limit; do not conflate output capacity with context window.

Google announced introductory pricing of $2 per million input tokens and $10 per million output tokens, with cached input priced 95% below the input rate. Availability and final billing details should be confirmed in official API documentation when access opens.

Google reports 77.9% on DeepSWE v1.1. This is a publisher-reported result and should not be compared directly with differently configured SWE-bench tests.

There is not enough directly comparable evidence in the cited launch materials to conclude that. Evaluate both against the same tasks, tools, budgets, and quality checks.