Skip to main content
Back

Cognition's Gemini 4 Neon Hits 46.2% on FrontierCode at 34% of Claude's Cost

Cognition's FrontierCode 2.0 benchmark puts Google's new mid-tier model within a point of Claude Opus 5.5 while costing roughly a third per task.

  • Gemini 4 Neon scores 46.2% on Cognition's FrontierCode 2.0 benchmark, best of any Google model.
  • Cost per task is $4.38, versus $12.82 for Claude Opus 5.5 at 47.3%.
  • Neon is now available in Devin Desktop and Devin CLI.
  • FrontierCode 2.0 grades whether a pull request is actually mergeable under underspecified prompts.
  • Google released Neon November 18 at $3 input / $15 output per million tokens promo pricing.
  • Model sits between Hydrogen and Carbon, with 1M context and four reasoning levels.

Cognition has added Google’s Gemini 4 Neon to Devin and published benchmark results for the model. Neon scored 46.2% on FrontierCode 2.0, Cognition’s best result from a Google model. It trails Claude Opus 5.5 by 1.1 percentage points while costing $4.38 per task, about 34% of Opus’s reported $12.82.

Claude Fable 5.2 leads the benchmark at 50.2%, placing Neon third overall. Developers can now select Neon in Devin Desktop and CLI.

What FrontierCode rewards

FrontierCode 2.0 evaluates whether an AI agent can turn an intentionally underspecified issue into a pull request that a maintainer would accept. Grading covers repository conventions, adjacent edge cases, cleanup behavior, tests, API stability, and the prompt’s explicit requirements.

Cognition says maintainers from the target repositories created the tasks, each of which required more than 40 hours of engineering work. For 80% of the tasks, the benchmark’s reference solution differs from the original upstream pull request, allowing evaluation beyond reproduction of a known patch. With the highest listed score at 50.2%, the benchmark still leaves substantial room for improvement.

A tight score gap, a wide cost gap

Neon ranks behind two Anthropic models on Cognition’s public leaderboard. It leads GPT-6.2 Astra by two percentage points and Google’s Gemini 4 Carbon by 3.6 points.

Selected FrontierCode 2.0 results. The harness column identifies the agent shell and reasoning setting used for each run.
Model Score Reported cost per task Harness
Claude Fable 5.2 50.2% Not reported claude-code, xhigh
Claude Opus 5.5 47.3% $12.82 claude-code, max
Gemini 4 Neon 46.2% $4.38 antigravity-cli, high
GPT-6.2 Astra 44.2% Not reported codex, max
Gemini 4 Carbon 42.6% Not reported antigravity-cli, high

Neon also improves on Gemini 4 Argon’s 27.6% result by 18.6 percentage points. That gain moves Google’s latest model from the lower half of the published results into the leading group.

Neon’s place in Gemini 4

Neon is the fourth Gemini 4 variant released in less than two months. Google positions it between Hydrogen and Carbon, combining Flash-class speed and lower pricing with coding, reasoning, and agentic performance comparable to Carbon. In the Gemini app, Neon replaces Argon and Gemini 3.1 Pro.

The model has a 1 million-token context window and four reasoning settings: Low, Medium, High, and Extra High. The larger context window can accommodate substantial repository history and code, although effective use still depends on the agent’s retrieval strategy and prompt construction.

During its promotional period, Neon costs $3 per million input tokens and $15 per million output tokens. Standard pricing will rise to $4 for input and $20 for output. Cognition’s $4.38 benchmark figure reflects the promotional rates, so production costs will vary with timing, token mix, reasoning level, and task length.

Harness and pricing caveats

  • Different agent shell: Neon’s benchmark run used antigravity-cli with high reasoning. Devin’s prompts, tools, and control loop can produce different results.
  • Aggregate score: The 46.2% result summarizes the benchmark suite and may conceal differences across languages, repositories, and task categories.
  • Variable spend: The reported task cost depends on promotional pricing and the benchmark’s input-output token ratio.
  • Higher-scoring options: Opus 5.5 leads Neon by 1.1 points, while Fable 5.2 leads it by four points.

Google separately claims Neon costs 32% less per task than Carbon at the default high reasoning setting. Neon can also orchestrate parallel Hydrogen subagents in Google Antigravity, an architecture intended to distribute work across cheaper models. The FrontierCode table does not quantify how much delegation contributes to cost or quality.

Teams evaluating Neon can compare models on representative repository issues while holding the harness, reasoning setting, and task set constant. Merge acceptance, reviewer edits, regressions, latency, and actual token spend will show whether Neon’s benchmark price-to-performance ratio carries into a specific development workflow.