Back to Blog

Gemini 4 Neon Beats Opus 5.5: Half the Tokens?

Daily News8992
Gemini 4 Neon Beats Opus 5.5: Half the Tokens?

Google DeepMind has released Gemini 4 Neon, the fourth member of the Gemini 4 family, targeting high-complexity software engineering, knowledge synthesis and scientific research workloads. Announced by Demis Hassabis, the new model is available immediately for Gemini Pro and Ultra subscribers. It delivers competitive benchmark performance paired with substantial cost reduction for agent workflows, marking a major upgrade for production AI agent deployment.

Gemini 4 Neon is positioned as a high-efficiency reasoning model built for agent-native tasks. On core evaluation suites, it outperforms many established top-tier models including Claude Opus 5.5 and GPT-6.2 Sol across most test tracks. In a side-by-side benchmark set of 19 evaluation categories, Gemini 4 Neon secured the highest score in 16 groups. The test scope covers agent coding, automated workflows, mathematical reasoning, multimodal comprehension, long-context processing and cybersecurity challenges.

In the Artificial Analysis Agent Coding Index, Neon reaches 58.6 points, taking the top spot among all measured models. It edges out Claude Fable 5.2 (58.5) and Claude Opus 5.5 (58.2), while Gemini 4 Carbon scores 55.5 and GPT-6.2 Astra lands at 53.5. One of the most commercially meaningful metrics is task consumption: Neon completes comparable agent tasks with roughly half the token budget of Claude Opus 5.5. On the AA Intelligence Index, Neon hits 61 points, matching the high-tier score of Gemini 4 Carbon, yet its per-task cost is less than half of Carbon’s. Measured average spend for a standard benchmark task is 3.7 USD for Neon, compared with 7.7 USD for Carbon.

The pricing structure for Neon also creates advantages for large-scale agent runs. The model is priced at 3 USD per million input tokens and 15 USD per million output tokens, with a 1 million-token maximum context window. For teams running mixed-model agent pipelines, unified API routing helps streamline access to Gemini family models. 4sapi serves as an API gateway that simplifies credential management and traffic shaping for multi-model workloads, which fits well with the swarm-style agent orchestration pattern introduced alongside Neon.

Google’s Antigravity agent platform is where Neon’s swarm capability is fully exposed. Neon acts as the coordinator agent, while lighter Gemini 4 Hydrogen agents execute subtasks in parallel. Hydrogen is optimized for speed and low cost, priced at only 1.5 USD per million tokens, half the cost of Neon. Neon breaks down complex objectives, assigns independent workstreams to Hydrogen workers, aggregates results, validates output quality and triggers rework when deliverables fail checks. This two-tier coordinator-worker architecture differs from typical Super-agent designs. Super-agent frameworks usually focus on logic planning and tool calling, while worker agents handle low-level execution. The Gemini swarm design gives Neon oversight of task dependency sequencing, error recovery and final result fusion, with Hydrogen dedicated to data retrieval, file editing, rendering and unit testing.

Real-world demonstration cases illustrate Neon’s long-horizon agent stability. In one extended test, a Neon-led swarm of agents ran continuously for 12 hours to reconstruct Turing’s library at King’s College Cambridge inside Blender. The workflow included archival image retrieval, 3D modeling, asset binding, physically based material setup and lighting rendering. The agent chain preserved task intent through dozens of sequential steps without drift, reproducing fine details including the original Turing machine.

Another practical test focused on computational spreadsheet logic. Neon generated more than 100,000 spreadsheet cells with custom mathematical formulas, implementing Sobel edge detection and 3D depth mapping inside a table format. The model could adjust lighting parameters and refresh rendered outputs in real time. A third engineering demo tasked Neon with building a complete first-person 3D alpine skiing simulator. It pulled geographic terrain data, wrote 10,200 lines of custom WebGL rendering code, built a 120Hz physics engine, defined six snow trails, and implemented dynamic avalanche effects, terrain deformation and spatial audio.

Neon also posts strong results on cybersecurity and automation benchmarks. It scores 90.5% on SEC-bench Pro and 70.5% on CWE-bench v2, outperforming Gemini 4 Carbon by 3.4 percentage points on the CWE suite. On Zapier’s automation benchmark built from 67 real business process scenarios, Neon achieves a 56.3% pass rate, ranking first among tested models. It does not hold every top mark: Claude Opus 5.5 still leads on Terminal-bench 4.1 and FrontierCode 2.0, showing that niche coding environments remain a competitive frontier.

The launch of Neon continues Google’s rapid release cadence for the Gemini 4 series. The rollout sequence began with Argon on September 30, followed by Carbon on October 28, Hydrogen on November 12, and Neon on November 18. The compressed release cycle reflects Google’s strategy of shipping specialized model variants rather than waiting for a single universal flagship. Each model targets a distinct workload profile: Argon for general use, Carbon for maximum raw intelligence, Hydrogen for lightweight parallel execution, and Neon for balanced agent coding and reasoning.

For developer teams building production agents, Neon creates a new tradeoff option. It matches near-top-tier reasoning ability while cutting operational costs significantly. The swarm coordinator pattern is native to the Antigravity stack, which means developers do not need to build custom orchestration layers from scratch to run multi-agent jobs. Teams that already route LLM traffic through centralized management layers can add Neon endpoints without rewriting core agent logic.

Benchmark data confirms that Neon’s gains are not limited to cost efficiency. Its long-context stability measured on LVBench reaches 89.0%, and it scores 71.3% on HLE-Diamond for high-level reasoning. On MMLU, the model hits 56.5%, and it reaches 84.0% on RiemannBench for advanced mathematical tasks. These scores place it in the upper tier of available frontier models, while its per-task economics separate it from competitors in agent use cases.

As the agent development ecosystem matures, the industry is shifting from evaluating isolated model pass rates to measuring end-to-end task completion cost and reliability. Gemini 4 Neon is built specifically for this new evaluation standard. It combines high benchmark scores, lower token consumption, and native multi-agent orchestration, positioning it for adoption in software development pipelines, scientific simulation workflows, automated business operations and security auditing.

Tags:Gemini 4 NeonGoogle GeminiGemini APIGemini 4 HydrogenGemini 4 CarbonGoogle Antigravity

Recommended reading

Explore more frontier insights and industry know-how.