April 2026 AI Weekly: GPT-4o Image Generation Drops, Open Source Model Battle Intensifies

OpenAI Quietly Rolled Out Native GPT-4o Image Generation

Early April, someone spotted a new image_generation parameter in the GPT-4o API. Not DALL·E, not a Midjourney integration — GPT-4o generating images directly.

Tested it out. Better than DALL·E 3, but still behind Midjourney v7. Text rendering has clearly improved, but faces and hands still occasionally glitch. Pricing: about $0.04 per 1024×1024 image.

OpenAI wants image generation to be just another API call, not a standalone product. For developers this is great. For image-focused startups... another chain tightening.

Open Source: Qwen3-Coder vs DeepSeek V3.5

Qwen3-Coder hit 92.3% on HumanEval, versus Claude Sonnet 4's 94.1%. 72B parameters, inference on just 4 A100s.

DeepSeek V3.5 quietly updated: MoE architecture with only 4.2B active parameters, but 236B total. Inference costs roughly 5x lower than a dense model.

The clearest trend of Q1 2026: open source models have matched closed source on code capability. The remaining gap isn't capability — it's product integration and ecosystem.

The Word Agent Is No Longer Enough

At SFD Lab, our 15 Agents check task queues, distribute work, and hand off to each other every morning without human involvement. Last week, when our security agent found a CORS misconfiguration, it wrote a fix and queued it for deployment automatically.

That's not an Agent. That's a colleague.

Local Inference Clusters Are Going Mainstream

More developers and small businesses are building local inference clusters. The hidden costs of API calls — data privacy, latency, vendor lock-in — are pushing people local.

We run Qwen3.5 35B on two Mac Studios. Per-token cost is essentially zero.

SFD Editor's Note

This week we switched all paid model calls to free or alternative options (OpenRouter + local Qwen3.5 fallback). All 15/15 Agents running on this now.