Another packed week in AI. OpenAI ended its 12-day restricted preview and opened GPT-5.6 to the world — three models, a new durability concept, and the ChatGPT Work + Codex integration that signals where the company is headed. Meta's Muse Spark 1.1 landed with enough firepower to pull Mark Zuckerberg back to X after three years of silence. And NVIDIA and Hugging Face took a big swing at open-source robotics.
Let's dig in.
OpenAI officially released the GPT-5.6 series on July 9, ending 12 days of restricted government preview. The three-model family — Sol, Terra, and Luna — introduces a new "durable capability tier" concept: the names identify capability levels, not versions, meaning Sol can be upgraded to a future GPT-5.7 while keeping its tier identity.
Sol, the flagship, sets new state-of-the-art across coding (80 on the Artificial Analysis Coding Agent Index, beating Claude Fable 5 by 2.8 points), cybersecurity (73.5% on ExploitBench vs GPT-5.5's 47.9%), and knowledge work. It runs in three effort modes: default for cost efficiency, max
for extended reasoning, and ultra
which coordinates 4 parallel agents by default (scalable to 16). Pricing runs $5/$30 per million input/output tokens for Sol, $2.50/$15 for Terra, and $1/$6 for Luna.
Alongside the model launch, OpenAI merged Codex into the ChatGPT desktop app and introduced ChatGPT Work, a unified interface for chat, coding, and long-running agent tasks. A new Programmatic Tool Calling feature in the Responses API lets GPT-5.6 write and run lightweight programs that coordinate tools inline. According to internal benchmarks, GPT-5.6 Sol improved the RSI Index by 16.2 points over GPT-5.5 on AI research acceleration tasks.
— OpenAI · ChatGPT Blog
🔗 OpenAI GPT-5.6 · ChatGPT Work The same week, OpenAI launched GPT-Live, a full-duplex voice model that listens and speaks at the same time. Two versions — GPT-Live-1 and GPT-Live-1 mini — started rolling out globally on July 8.
Previous voice systems either chained three models together (cascaded) or worked in rigid turn-based mode where the model waited for silence before responding. GPT-Live's full-duplex architecture processes input continuously while generating output, making interaction decisions many times per second — whether to speak, listen, , interrupt, or invoke a tool. It handles backchannel cues ("mhmm", "yeah"), stays quiet when you need a moment, and can perform real-time simultaneous translation.
When a question requires deeper reasoning or search, GPT-Live delegates to GPT-5.5 behind the scenes and brings results back into the conversation without breaking flow. In head-to-head evaluations, both models are strongly preferred over Advanced Voice Mode for pleasantness, turn-taking, and natural flow. GPT-Live-1 substantially outperforms Advanced Voice Mode on GPQA (expert-level science reasoning) and BrowseComp (agentic web search). A demo showed it translating live between languages with no perceptible delay.
— OpenAI
🔗 OpenAI GPT-Live Meta released Muse Spark 1.1 on July 9, and the event was significant enough that CEO Mark Zuckerberg posted on X for the first time in three years. "An incredibly capable agent and coding model at a very low price," he wrote.
Muse Spark 1.1 is purpose-built for agentic tasks — planning, tool calling, subagent delegation, and computer use. On agent benchmarks, it scores 54.7% on JobBench (beating Claude Opus 4.8's 48.4%) and 88.1 on MCP Atlas (ahead of Opus 4.8's 82.2). Its 1-million-token context window and context compaction mechanism allow it to maintain state across long sessions. The model supports a main-agent/sub-agent delegation pattern, zero-shot generalization to new tools and MCP servers, and three computer-use execution modes (scripts, clicks, or batched actions per step).
On coding, gains are dramatic but mixed. Vibe Code Bench jumped from 19.7% to 72.2%, but on SWE-Bench Pro Muse Spark 1.1 scores 61.5% — behind Claude Opus 4.8's 69.2%. DeepSWE 1.1 shows a similar gap at 53.3% vs Opus 4.8's 59.0%. Meta positions the model less as a pure coding leader and more as an agent orchestrator — capable of managing multi-agent workflows, maintaining context across sub-tasks, and completing projects faster than its predecessor.
— Meta AI Blog · Zuckerberg on X
🔗 Meta AI Blog — Muse Spark · TMT Post Analysis · Kingy.ai Benchmarks
Microsoft has quietly started replacing third-party AI models — including OpenAI's and Anthropic's — with its in-house MAI series in core Office products, Bloomberg reported on July 8.
Excel and Outlook now process tens of thousands of weekly AI prompts entirely on MAI models, a deployment scale not previously disclosed. While the swap covers only a fraction of Microsoft's total AI workload, it marks a strategic inflection point: Microsoft is no longer willing to pay premium pricing to OpenAI and Anthropic at scale. Mustafa Suleyman's AI team is building toward full model independence, with the MAI series designed to handle Copilot's massive token consumption at a fraction of the cost. The current OpenAI partnership still provides discounted access, but those terms are narrowing.
— Bloomberg · Peng
🔗 Bloomberg via 163.com NVIDIA and Hugging Face announced on July 7 a major expansion of their robotics partnership, integrating NVIDIA's Isaac GR00T 1.7 vision-language-action model and Isaac Teleop data-capture framework into Hugging Face's open-source LeRobot library.
GR00T 1.7 is the first open, commercially viable robot foundation model for humanoid robots. Developers can post-train and deploy it through standard LeRobot workflows without proprietary toolchains. Isaac Teleop enables high-quality human demonstration capture in interoperable formats, feeding directly into LeRobot datasets. On the road map: Cosmos 3, a frontier world foundation model for generating synthetic robotics data when real-world data is too expensive or dangerous to collect.
The partnership connects NVIDIA's 3 million robotics developers with Hugging Face's 16 million AI builders, creating a unified pipeline: teleoperate → train on GR00T → simulate with Cosmos → deploy through LeRobot.
— NVIDIA Blog · Hugging Face Blog
🔗 NVIDIA Blog · Hugging Face Blog Z.ai (formerly Zhipu AI) launched ZCode on July 2, a free "Agentic Development Environment" powered by the GLM-5.2 model — 744 billion parameters with ~40 billion active under a Mixture-of-Experts architecture.
On SWE-Bench Pro, GLM-5.2 scores 62.1, surpassing OpenAI's GPT-5.5 at 58.6 (though trailing Claude Opus 4.8 at 66.0). On Terminal-Bench 2.1, it scores 81.0 against Opus 4.8's 85.0. ZCode's pricing is aggressive: the base tier is free, paid plans start at $16.20/month (undercutting Cursor Pro at $20), and API pricing is $1.40/$4.40 per million input/output tokens — a fraction of Claude Opus 4.8's $5/$25.
The agent-first IDE supports macOS, Windows, and Linux, including remote control via WeChat and Feishu messaging bots — a feature designed for the Chinese enterprise market. ZCode arrives three weeks after the US suspension of Anthropic's Fable 5 model, creating what some developers call "another DeepSeek moment." The model uses Z.ai's proprietary IndexShare sparse-attention technique and supports a one-million-token context window.
— Z.ai · EastFrontier
🔗 EastFrontier Mistral AI released Leanstral 1.5 under Apache 2.0 on July 2 — a 119-billion-parameter sparse MoE model specialized for theorem proving and code verification in Lean 4.
The numbers are striking: 100% on miniF2F (both validation and test), 587 out of 672 PutnamBench problems solved, and new state-of-the-art on FATE-H (87%) and FATE-X (34%) algebra verification benchmarks. At roughly $4 per problem on PutnamBench, it's far below the highest-compute comparison systems.
But the real story is practical impact: Mistral used a pipeline translating Rust into Lean, generating candidate correctness properties, and attempting to prove or disprove them. Across 57 open-source repositories, Leanstral 1.5 identified 11 genuine bugs, five of which were previously unreported. The model activates only ~6 billion parameters per token (of 119B total), making it deployable at a fraction of its full compute cost. It supports a 256,000-token context window and is available free through Mistral's Labs tier and Vibe agent environment.
— Mistral AI
🔗 Mistral AI Blog · The Agent Times · Hugging Face Model Next digest: July 13, 2026. Follow KD Agentic for daily AI coverage.