Gemini 4 Argon arrives in Devin with a 75% launch discount

Cognition has added Google’s Gemini 4 Argon to Devin Desktop, formerly Windsurf, and the Devin CLI. According to the Argon announcement, it is the highest-scoring Gemini model on Cognition’s internal coding benchmark. The integration gives developers another model for long-running coding tasks and a temporary 75% discount through October 27.

A 60.5% score needs context

Cognition reports that Argon scored 60.5% on FrontierCode 1.1, its internal coding benchmark. That result places Argon ahead of every other Gemini model Cognition has tested.

A separate public benchmark analysis evaluated 23 models for correctness and code-review quality, with its top model passing every blocking check in 57.7% of attempts. Differences in model versions, agent harnesses, prompts, tools, sampling settings, and scoring limit comparisons between separate runs. Cognition’s 60.5% result establishes Argon’s position on its own leaderboard, while a controlled head-to-head test would be required to rank it against the public run.

The launch price has two layers

Google’s introductory list price is $2 per million input tokens and $10 per million output tokens. Cached input receives a 95% reduction, bringing that rate to $0.10 per million tokens. Google plans to raise the standard rates to $4 for input and $20 for output after the introductory period.

• Devin promotion: 75% off through October 27.
• Illustrative discounted rate: $0.50 per million input tokens and $2.50 per million output tokens if Devin applies the discount directly to Google’s introductory list price.
• Billing detail to confirm: Cognition’s announcement does not explain whether the Devin promotion stacks with cached-input pricing.
• Deadline detail to confirm: The quoted promotion date does not include a year, so users should check the current price shown in Devin before starting a large run.

Model selection takes one setting

Developers can select Argon from Devin Desktop’s model picker. The CLI supports model selection through the --model flag, the /model command, or the configuration file. Cognition lists short model names such as opus, sonnet, swe, codex, and gemini; users should confirm the current Argon identifier in the CLI’s model list.

A larger ceiling for long runs

Google describes Argon as its new flagship model, with stronger coding, reasoning, and multimodal capabilities. Its maximum output rises from 64,000 to 1 million tokens, an increase of roughly 16 times. Tokens are chunks of text or code, so the higher limit allows an agent to produce substantially more output before reaching the model’s cap.

The larger allowance may reduce forced handoffs during migrations, multi-file refactors, and repository audits. The figure describes output capacity; context-window size, tool limits, execution time, and the model’s ability to remain coherent across a long run remain separate constraints.

Google also cites two internal case studies. In one, Argon agents replaced 32,000 lines of SIMD code, which uses vectorized processor instructions, in the libgav1 video decoder. Google says the resulting Rust implementation ran 2.7 times faster than the previous Rust version. In another case, the model helped quantum-computing researchers improve on a published optimization baseline by 40%. These are vendor-reported projects rather than general performance guarantees.

Strengths vary by workload

Google’s reported results place Argon near the front on several coding, automation, video, and security evaluations, while other models retain higher scores on terminal use and software-engineering tasks.

Argon’s reported profile favors long-horizon refactoring, long-video analysis, automation, and security remediation. Claude Opus 5.5 has the stronger reported result for terminal-heavy work, while GPT-6 Astra leads the cited FrontierSWE v2 comparison. Benchmark versions and harness settings should remain fixed when teams use these scores to choose a model.

Devin turns the choice into a repository test

Devin provides a shared agent harness across models from Anthropic, OpenAI, Google, Cognition, and open-source projects including DeepSeek, Kimi, and GLM. The harness supplies the repository, terminal, tools, and workflow around each model, making it easier to compare models on the same task and route work by cost, speed, or benchmark fit.

A controlled repository trial will provide more useful evidence than a cross-vendor leaderboard alone. Teams evaluating Argon can keep the following conditions consistent:

• Use the same repository commit, environment, prompt, tools, and token budget.
• Choose representative tasks such as a bug fix, migration, test repair, and security patch.
• Record completion rate, wall-clock time, token use, test results, regressions, and review effort.
• Inspect generated diffs for unnecessary changes, weak tests, and unresolved failures.
• Compare total task cost after retries instead of relying only on per-token pricing.

The temporary discount lowers the cost of those trials, especially for jobs that can use Argon’s larger output allowance. Its practical value will depend on whether longer runs produce mergeable changes with fewer retries and less reviewer intervention.