4 September 2026

Meta releases faster coding model at unchanged prices

• Meta's new Muse Spark 1.3 model completes coding tasks using about 25% fewer tokens and 20% fewer tool calls than the previous version, with pricing staying the same per token.
• The model improves at sustained longer tasks by generating its own context, correcting plan gaps, and following complex instructions with fewer back-and-forth turns needed.
• Independent evaluators found different efficiency results than Meta's internal tests, suggesting gains may vary significantly depending on the specific coding workload and evaluation method used.
• Meta's pricing undercuts competitors like Anthropic and OpenAI by 4 to 6 times while delivering comparable performance, potentially appealing to enterprises running large-scale software development.

Where they differ

• TLDR AI

  focused on a jailbreak vulnerability affecting multiple AI models.

• AI Breakfast

  reported on Meta's model release and competitive pricing. The outlets covered entirely different stories.

What each one reported

TLDR AITLDR editorial team

A MATS researcher discovered a synthetic transcript generation prompt that became a universal jailbreak hitting 84-100% attack success on nine of 23 models tested. Only recent Anthropic models and Meta Muse Spark 1.1 were never fully broken.

AI BreakfastIndependent editors

Meta released Muse Spark 1.3 with competitive frontier benchmarks and aggressive pricing at $1.25 and $4.25 per million tokens, tied to its personal agent roadmap. MBZUAI also released K2 Horizon, a fully open model family up to 375B with training data and code included.

Reported by InfoWorld