Issue 355

Gemini 3.5 Flash excels on several benchmarks. On the APEX-Agents-AA benchmark (testing long-horizon agentic tasks in investment banking, management consulting, and corporate law), it achieves 47.1% accuracy on the first attempt, leading second-place GPT-5.5 (37.7%) by nearly 10 percentage points. On MMMU-Pro (cross-disciplinary visual reasoning), it reaches 84% accuracy in high reasoning mode, setting a new record and surpassing Gemini 3.1 Pro Preview's 82%.

On GDPval-AA (real-world agentic tasks), Gemini 3.5 Flash high reasoning mode achieves an Elo score of 1,656, higher than Gemini 3.1 Pro Preview's 1,314 but lower than GPT-5.5 xhigh mode's 1,769. On ARC-AGI-2 (abstract visual reasoning), high reasoning mode scores 72.1%, trailing Gemini 3.1 Pro Preview (77.1%) and GPT-5.5 xhigh (85.0%).

On the AA-Omniscience knowledge benchmark (correct answers score, hallucinations penalized), Gemini 3.5 Flash reasoning mode scores 23, lower than Gemini 3.1 Pro Preview's 33 and Claude Opus 4.7 maximum reasoning mode's 26. As of May 24, 2026, it ranks ninth on the Arena.ai text arena (Elo 1,480) and tenth on the WebDev coding arena (Elo 1,506).

Gemini 3.5 Flash API pricing is $1.50 per million input tokens, $9.00 per million output tokens, and $0.15 for cached input. This makes its input price triple that of the previous Gemini 3 Flash. The model is available for free via the Gemini app, Google AI Studio, Google Antigravity (compute limit refreshes every 5 hours, weekly cap), and Google Search AI mode, but enterprise and API access require payment.

This price increase is part of a broader trend of rising AI model pricing. While performance gains are significant, the cost increase may impact adoption decisions by developers and enterprises.

Google has not disclosed details such as parameter count, training data, or methods. The model card indicates that Gemini 3.5 Flash is "based on" Gemini 3 Flash, which in turn is based on Gemini 3 Pro. It is a mixture-of-experts (MoE) Transformer, pre-trained on multimodal data including text, code, images, audio, and video, sourced from web crawls, licensed materials, Google user data, and synthetic data.

It is fine-tuned via reinforcement learning on datasets covering multi-step reasoning, problem-solving, and theorem proving. The model supports adjustable reasoning levels (minimum, low, medium, high) and thought retention (retaining reasoning tokens in context to maintain reasoning coherence across multi-turn conversations, similar to Kimi K2.6's retained thought feature), but does not yet support computer use.

Andrew Ng discusses the rise of AI Field Deployment Engineers (FDEs) in The Batch. The FDE role was pioneered by Palantir about 20 years ago, and now companies like OpenAI and Anthropic are building teams to embed FDEs within client organizations. FDEs require technical, communication, and business skills to customize and adjust agentic workflows.

Ng believes the number of AI engineer roles will far exceed FDEs. Most companies prefer their own employees to work on projects rather than rely on external FDEs. Additionally, clients worry that FDEs may deeply tie them to specific vendor products, reducing future flexibility to switch vendors. Currently, demand for AI engineers is high; they build applications using LLM prompts, agentic frameworks, evaluation tools, and AI coding agents (e.g., Claude Code, Codex, Antigravity CLI, and OpenCode). As the role matures, AI engineers may specialize into more distinct roles like LLMOps engineers, evaluation engineers, AI data engineers, etc., similar to how software engineers diversified in the past.