OpenAI's GPT-5.6 rollout stumbled as users grappled with 36 different model variants across Luna, Terra, and Sol tiers, each with multiple effort levels that replaced the familiar single model picker.

The complexity sparked immediate backlash from developers who found themselves constantly optimizing for cost rather than performance. API users now navigate combinations like Luna High for everyday coding, Terra Medium for bigger features, and Sol Ultra for orchestration tasks.

OpenAI responded with multiple usage limit resets after users complained their quotas burned down faster than expected. The company acknowledged that default settings nudged users toward overly expensive configurations and promised to restore familiar navigation patterns.

The new ChatGPT Work and Codex split also drew criticism for making chats and projects harder to find. Staff explained that Max settings mean one model spending longer on problems, while Ultra parallelizes work across subagents.

Early benchmarks show GPT-5.6 excelling in agentic coding and presentation tasks. It tied for first place in Code Arena Frontend with Claude Fable 5 while costing roughly half as much on listed pricing.

Meta challenges with Muse Spark 1.1

Meta's Muse Spark 1.1 emerged as the week's surprise release, with practitioners praising its UI generation capabilities and aggressive $1.25 per million input tokens pricing.

Artificial Analysis scored Muse Spark 1.1 at 51 on its Intelligence Index, up 8 points from version 1.0. The model offers 1 million token context and median speeds of 114 tokens per second.

The release signals Meta's compute-heavy investments translating into cost-effective inference products that could pressure OpenAI and Anthropic on pricing.

Meanwhile, open-model tooling continued advancing with Unsloth releasing Qwen3.6 NVFP4 quantizations claiming 2.5× faster inference, including a 27B variant running on 24GB VRAM.