GPT-5.6 Luna on Replicate, Muse Spark 1.1 on AI Gateway, and agent workflows
TL;DR
OpenAI and Meta rolled out new models for agents and high-volume tasks while platforms added access points and tools for builders.
What shipped
On 9 July new model releases and platform updates focused on agentic work and cost-efficient inference. OpenAI expanded GPT-5.6 variants across hosts while Meta introduced Muse Spark 1.1. Additional items covered transparency rules, hardware, and trending repositories.
Replicate new models
GPT-5.6 Luna on Replicate: OpenAI released GPT-5.6 Luna on Replicate for fast inference. The model targets latency-sensitive chat and classification at scale. Builders can call it through the HTTP API or playground today.
Vendor launches
Meta and OpenAI supplied the bulk of the updates with Muse Spark 1.1 and three GPT-5.6 tiers on AI Gateway. Google added ad labels while Nvidia and Anthropic posted smaller items.
- •Muse Spark 1.1 on AI Gateway Meta released Muse Spark 1.1 on AI Gateway with a 1M token window. The model handles text, image, video, PDF, and audio for agent planning. It supports parallel tool calls and works with MCP servers.
- •GPT-5.6 models on AI Gateway OpenAI added Sol, Terra, and Luna to AI Gateway in preview. The three tiers improve agentic coding and biology tasks while using fewer tokens than prior versions. Sol serves as the top performer.
Hugging Face trending
Seven text and image models climbed the Hugging Face charts. Nvidia contributed two large text models while smaller community releases focused on quantized and control variants.
- •Supra-Router-51M SupraLabs published Supra-Router-51M for text generation. The model runs on the transformers library and supports fine-tuning on the Hub.
- •MiniCPM5-1B variant GnLOLot released a GGUF quantized MiniCPM5 model tuned with Claude and Opus style data. It targets local text generation.
- •NVIDIA Nemotron Labs 3 Nvidia listed NVIDIA-Nemotron-Labs-3-Puzzle-75B for text tasks. The model uses the transformers stack and includes evaluation numbers.
- •Nemotron Labs Audex Nvidia added Nemotron-Labs-Audex-30B-A3B for text generation. Builders can download and fine-tune it directly.
- •Krea-2 depth controlnet Patil released Krea-2-depth-controlnet for image-to-image work. The model accepts depth maps to guide output.
- •Bielik entity study Researchers published a paper on activation signals that predict when Polish Bielik models lack entity knowledge. The work covers four model sizes.
- •Recursive self-improvement survey A new paper reviews loops where models revise their own outputs and run research tasks. It catalogs recent self-refine methods.
Product Hunt picks
Five tools appeared on Product Hunt. Two focus on agent control of legacy apps and SEO data while others target interfaces and prompt safety.
- •Just Ask SEORCE launched a WhatsApp bot that queries SEO and AI visibility data. Users send questions directly in chat.
- •Coasty Coasty introduced a computer-use agent that operates legacy desktop software like a human. It clicks and types through existing interfaces.
- •Monogram AI Monogram AI released a visual interface for interacting with models. The design supports drag-and-drop prompt building.
- •Constellation Gate AI Constellation Gate AI added prompt injection filters and token compression. It ranks first on several safety benchmarks.
Industry news
Meta drove hardware and model pricing stories while OpenAI paired GPT-5.6 with a new workflow agent. Government safety reviews and ad transparency updates also appeared.
- •Meta AI chips Meta confirmed its next AI chips enter production in September. The design uses modular blocks to adapt to changing model needs.
- •Estonia AI error checker Estonia built an AI system to catch legal wording mistakes after a $28 million error. The tool scans proposed laws before passage.
- •ChatGPT Work agent OpenAI launched ChatGPT Work, an agent that manages full projects across Drive, Slack, and Salesforce. It runs on GPT-5.6 and is live now.
- •Muse Spark pricing Meta set Muse Spark 1.1 output tokens at $4.25 per million, undercutting OpenAI and Anthropic rates. The price targets agent workloads.
- •llm-meta-ai 0.1 Simon Willison released llm-meta-ai 0.1 to run prompts against Muse Spark 1.1. The package adds the model to the llm CLI.
- •Meta coding push Meta positioned Muse Spark 1.1 for large code migrations and bug fixes. Enterprises can use it for agent-driven refactors.
Other
OpenRouter added multiple GPT-5.6 variants with different reasoning modes and price points. Builders can now route high-context or pro-reasoning calls through the same API key.
- •GPT-5.6 Luna Pro on OpenRouter OpenAI listed GPT-5.6 Luna Pro on OpenRouter with 1,050k context. It uses pro reasoning mode and costs $1 per million input tokens.
- •GPT-5.6 Luna on OpenRouter OpenAI added GPT-5.6 Luna to OpenRouter for high-volume tasks. The model keeps the same 1,050k context at $1 input and $6 output.
- •GPT-5.6 Luna Pro variant OpenRouter now hosts the pro version of GPT-5.6 Luna for complex tasks. Users add the model ID to existing setups for immediate testing.
- •GPT-5.6 Luna base OpenRouter activated the base GPT-5.6 Luna model. It suits chat and lightweight agents at the listed token rates.
- •GPT-5.6 Terra Pro on OpenRouter OpenRouter added GPT-5.6 Terra Pro with 1,050k context. Pricing starts at $2.50 per million input tokens.
- •GPT-5.6 Terra base OpenRouter listed GPT-5.6 Terra for everyday balanced work. The model uses the same context window at standard rates.
- •GPT-5.6 Sol Pro on OpenRouter OpenRouter released GPT-5.6 Sol Pro at the highest tier. It costs $5 input and $30 output per million tokens for flagship performance.
What this means for you
For Vibe Builders: You can now test GPT-5.6 Luna and Muse Spark 1.1 through Replicate, AI Gateway, and OpenRouter without new accounts. Coasty and Constellation Gate AI give ready-made agents for legacy apps and prompt safety. Add the model IDs to existing keys and run a small workflow today to compare speed and cost.
For Non-techies: ChatGPT Work and Muse Spark 1.1 now handle full projects across Drive and Slack, so you can hand off routine tasks instead of typing every step. Google ad labels make it easier to spot AI-made promotions. Check subscription tiers before trying the new agents on your own accounts.
For Developers: OpenRouter and AI Gateway added GPT-5.6 variants with 1,050k context and pro reasoning modes. Compare token prices against Muse Spark 1.1 before updating production routes. Watch the new Nemotron and Supra models on Hugging Face for local fine-tuning benchmarks this week.
What to watch next
Track OpenAI and Meta API pricing changes and any new agent benchmarks. Watch for production rollout notes on ChatGPT Work and Muse Spark 1.1. Note any Hugging Face model updates that reach stable downloads.
Harsh’s take
The day showed repeated launches of the same GPT-5.6 family across hosts while Meta undercut prices on agent features. Most noise came from access points rather than new capabilities. Builders face a growing list of similar endpoints that differ mainly in cost and context limits.
The practical step is to pick one host, run a fixed agent task on both Luna and Muse Spark, and record latency plus token spend. Skip the rest until clear differentiation appears in reliability metrics.
by Harsh Desai
Sources
Replicate new models
Vendor launches
- •Muse Spark 1.1 is now available on AI Gateway
- •GPT 5.6 Sol, Luna, and Terra now available on AI Gateway
- •Expanding AI transparency in ads
- •GeForce NOW Turns Up the Heat With New GeForce RTX 5,080-Powered Toronto Server
- •Ben Bernanke appointed to Anthropic’s Long-Term Benefit Trust
Hugging Face trending
- •Supra-Router-51M by SupraLabs trends on HuggingFace
- •MiniCPM5-1B-Claude-Opus-Fable5-Thinking-GGUF by GnLOLot trends on HuggingFace
- •NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 by nvidia trends on HuggingFace
- •Nemotron-Labs-Audex-30B-A3B by nvidia trends on HuggingFace
- •Krea-2-depth-controlnet by Patil trends on HuggingFace
- •Does Bielik Know What It Doesn't Know? Activation Dispersion Separates Entity Familiarity from Factual Reliability Across Model Scale
- •Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
Product Hunt picks
Industry news
- •Meta’s new AI chips will begin production in September
- •The $28 Million Mistake That Inspired Estonia’s AI ‘Fuckup Finder’
- •OpenAI pairs its GPT-5.6 public rollout with ChatGPT Work, a new agent that handles entire workflows
- •Meta's Muse Spark 1.1 API pricing squeezes OpenAI and Anthropic as the AI price war heats up
- •llm-meta-ai 0.1
- •Meta enters the crowded AI coding battle with Muse Spark 1.1
- •Google will now disclose which ads are made with AI
- •How did the government decide OpenAI’s frontier model was safe to release?
- •Elon Musk praises Mythos/Fable, promises not to ‘cut off’ Anthropic
- •Can AI answer the $3 trillion question?
Other
- •OpenAI: GPT-5.6 Luna Pro now available on OpenRouter (1,050k context, $1.00/M in, $6.00/M out)
- •OpenAI: GPT-5.6 Luna Pro now available on OpenRouter (1,050k context, $1.00/M in, $6.00/M out)
- •OpenAI: GPT-5.6 Terra Pro added on OpenRouter
- •OpenAI: GPT-5.6 Terra added on OpenRouter
- •OpenAI: GPT-5.6 Sol Pro added on OpenRouter
More AI news
- FeatureOpenAI shuts down Atlas AI browser and moves features to desktop app
OpenAI ends Atlas after less than a year. It moves agentic browsing features to its desktop app and a Chrome extension.
- PricingOpenAI launches GPT-5.6 family Luna Terra Sol with new per-token pricing
OpenAI released the GPT-5.6 model family in three sizes with pricing of $1/$6 for Luna, $2.50/$15 for Terra, and $5/$30 for Sol per million input/output tokens.
- PricingGPT-5.6 Sol nearly matches Fable 5 on aggregated benchmarks at one-third the cost
OpenAI's GPT-5.6 Sol scores 59 points on the Artificial Analysis Intelligence Index, one point behind Claude Fable 5. At $1.04 per task it costs one third as much as Anthropic's top model and leads in agentic coding.