Google's Gemini 4 Neon achieved 12.8% on the SWE-sweep coding benchmark when tested with the mini-SWE harness, but performance jumped dramatically to 39.0% when the same model operated within Antigravity. The sharp improvement highlights a critical gap in how the industry measures AI coding capabilities.

The more-than-threefold improvement raises questions about how much coding-agent performance comes from the underlying model and how much depends on the harness managing its tools, context, and execution.

Google announced both findings on Wednesday during the Neon launch. The disparity underscores a broader challenge: distinguishing whether gains in coding-agent performance stem from model improvements or from enhancements to the harness that orchestrates the model's tools, manages context, and handles execution.

A separate FrontierHarness benchmark released the day before revealed that running the same open model could incur costs varying by up to five times depending on which harness was used—further evidence that infrastructure choices shape real-world outcomes.

Same model, different scores

Google's Gemini 4 Neon model card presents a comparison of three frontier models evaluated on SWE-sweep using both their respective provider harnesses and a shared mini-SWE baseline.

Neon and Opus performed substantially better with their providers' harnesses, but Sol did not.

Claude Opus 5.5 improved from 3.6% with mini-SWE to 10.1% when paired with Claude Code. GPT-6.2 Sol moved in the opposite direction, declining from 11.6% with mini-SWE to 8.3% with Codex. Neon and Opus showed marked gains with their respective provider harnesses, whereas Sol's performance deteriorated.

Google evaluated each model across two harness configurations, though the comparison does not pinpoint which specific changes drove the differences or whether findings would replicate across other benchmarks and code repositories.

The harness does the work

Antigravity's harness underwent co-training with Gemini, and Neon has been refined to reduce unnecessary file reads and improve context management during extended coding tasks. Antigravity supports running subagents in isolated Git worktrees and can verify their modifications before integration.

Google only tested Neon against mini-SWE, so there's no data yet on how it performs with other harnesses.

The relative contribution of each enhancement to the SWE-sweep improvement remains unclear because they were not tested in isolation. Testing of Neon has been limited to mini-SWE, leaving open the question of how the model would perform when integrated with alternative harnesses.

Neon plans, Hydrogen builds

Neon joins the model-switching architecture Google rolled out with Gemini 4 Hydrogen. The system assigns planning and complex coding tasks to Neon while delegating simpler work to Hydrogen, with the ability to spawn multiple Hydrogen subagents in parallel through the /swarm feature introduced alongside Carbon.

Google has not disclosed how this architecture influences Neon's SWE-sweep results or what implications it carries for task completion costs.

Counting the cost per task

Neon achieved 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its standard high reasoning setting. The model also secured the top position on the Artificial Analysis Coding Agent Index with a score of 58.6.

Operating at its default high thinking level, Neon costs 32% less per task than Gemini 4 Carbon based on Google's calculations using the Artificial Analysis Intelligence Index. API pricing stands at $3 per million input tokens and $15 per million output tokens, with cached input priced at $0.15 per million tokens and a 1 million-token context window.

Neon is accessible via the Gemini API, Google AI Studio, and Gemini for Business. Users with Google AI Pro and Ultra subscriptions can also reach Neon through Antigravity.