GPT-5.6 Sol Lands, Meta Goes Cheap, and AI Learns to Fix Itself

OpenAI's GPT-5.6 Family Turns the Screw on Cost

OpenAI has launched GPT-5.6 in three tiers, Sol, Terra and Luna, rolling out across ChatGPT, Codex and the API. The headline isn't a benchmark, it's the bill: same list pricing as GPT-5.5 but a big claimed jump in dollars-per-task. Sol lands roughly second on independent intelligence testing behind Claude Fable 5, yet at around a third of the cost per task, and it tops a leading coding-agent index at 80. It also set a new state of the art on the ARC-AGI-3 reasoning game at 7.8%, the first frontier model to actually beat one. Alongside the models came ChatGPT Work, a merged Codex-and-ChatGPT desktop app, a Sites beta, and new programmatic tool-calling and multi-agent features. The shadow over it all: UK safety auditors reported universal jailbreaks enabling long-horizon agentic misuse, described by some as the highest-stakes issue of any release yet.

Meta's Muse Spark 1.1 Chases the Bottom of the Cost Curve

Meta launched Muse Spark 1.1 and its first hosted Meta Model API, with a 1M-token context, video understanding, and pricing near $1.25 in / $4.25 out per million tokens. Independent takes were mixed but broadly positive, strong on legal, tax and medical tasks, and possibly the cheapest frontier-class agent going.

The Open and Local Wave

Ollama announced a major raise and says it now serves over 9 million active builders, positioning itself as the ownership layer for open-model AI. And a research method called TRACE let a Qwen3.6-27B model diagnose its own failures and train itself to fix them, reportedly hitting 73.2% on SWE-bench Verified and beating far larger models.

Faster, Bigger, Cheaper Under the Hood

Mirai Labs showed a hybrid draft model for speculative decoding claiming 4.37x faster generation. Sparse Delta Memory introduced sparse addressing into recurrent state, claiming a working memory around 3000x larger at the same compute. And fal reported image generation at 0.45 seconds, while Reve 2.1 climbed to number two on a text-to-image arena.

AI Learns to Read Your Wristband

Google Research unveiled SensorFM, a wearable-data foundation model trained on roughly 1 trillion minutes of data from about 5 million consenting participants, aimed at heart, metabolic, sleep and mental-health signals. It points to a future where the AI understands the sensor already on your wrist.