Figure 1. Meta Muse Image generation from Meta Muse Image and Muse Video announcement. The new AI models are out of this world.TLDR - This week, two American AI labs rejoined the frontier AI model race and delivered frontier-level AI models: SpaceXAI’s Grok 4.5 and Meta’s Muse Spark 1.1. And OpenAI released GPT-5.6 and ChatGPT Work, which integrates agentic AI in ChatGPT interface.OpenAI publicly released the GPT-5.6 model family, making generally available three GPT-5.6 model versions: the flagship Sol, the balanced Terra model for everyday use, and the lower-cost Luna model. The models are frontier-level on coding, computer use, scientific reasoning and professional knowledge work. Sol state-of-the-art, and Terra on par with GPT-5.5 but less costly. OpenAI emphasizes that GPT-5.6 models are much more token efficient than prior models, making them faster and cheaper to run.Figure 2. GPT-5.6 is not only better than Claude Opus 4.8 and GPT-5.5, but GPT-5.6 Terra does it for far lower cost.GPT‑5.6 Sol is SOTA across agentic, coding, and knowledge-work benchmarks, getting 92.2% on BrowseComp, 62.6% on OSWorld 2.0, and an ELO of 1748 on GDPval-AA. Sol gets 72.7% on DeepSWE, above Claude Fable 5’s 69.9% at one-third lower cost. GPT-5.6 Sol scored 7.8% on ARC-AGI-3, the benchmark on fluid intelligence, which sounds low but it above prior models, whose scores were below 2%.Figure 3. GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, Grok 4.5 and Muse Spark 1.1 join the frontier AI model lineup this week as top 10 best AI models.OpenAI touts GPT-5.6 improvements in design and polish in outputs, including improved quality and accuracy in presentations and spreadsheets. Users have shown off its use in one-shot video games, websites, graphics, interface designs, and creating and editing videos:GPT-5.6 Sol is unbelievably good at creating and editing videos. It can do motion design, product demos, and animations like this one I made by simply giving it a screen recording. GPT 5.6 has the best design taste and significantly outperforms Fable, which relies heavily on repetitive design patterns.GPT-5.6 has undergone extensive red-teaming and safety checking as described in OpenAI’s system card. While claiming meaningful cybersecurity improvements, also described disturbing actions such as destructive virtual-machine cleanups, unauthorized credential copying, and fabricated verified research results in a small share of tasks.In addition, METR rejected its own GPT-5.6 Sol evaluation because the model showed unusually high rule-breaking and loophole exploitation during tasks. METR estimated either an 11.3-hour or 270-plus-hour task horizon depending on how cheating was treated, but they said neither figure was robust.OpenAI expressed in their launch video that users could “increase their ambition” in using GPT-5.6 on more complex agent use-cases. They introduced a new Ultra mode for Sol that coordinates four agents in parallel to more quickly complete difficult, long-running assignments. Showcasing agentic Sol use cases, OpenAI stated that Sol autonomously performed the post-training for Luna.GPT-5.6 Sol is available on paid plans, while Terra and Luna also go to free users, with listed API pricing of $5/$30, $2.50/$15, and $1/$6 per million input/output tokens.OpenAI announced ChatGPT Work, their unified ChatGPT app which has an agentic Work interface akin to Claude Cowork and where Codex features are now built-in to a combined ChatGPT desktop app. The ChatGPT desktop app that was Codex is now ChatGPT app, with users choosing between two modes, developer-oriented ChatGPT Codex and a broader ChatGPT Work. The ChatGPT Work release adds unified plugins across ChatGPT and Codex, browser multi-tab support, and faster computer use.ChatGPT Work provides an agentic interface for users to leverage plug-ins and agentic AI to generate “share-ready” work products: Slide decks, documents, spreadsheets, and more. OpenAI also includes Sites in the interface, the feature added last month to Codex that allows users to generate websites on the fly and host them on the chatgpt.site subdomain.OpenAI launched GPT-Live-1 and GPT-Live-1 mini, new full-duplex voice models that can listen and speak simultaneously instead of forcing users through rigid turn-taking. The models can respond to interruptions, acknowledge a speaker while listening and continue a conversation while delegating difficult searches or reasoning tasks to a frontier-level AI model in the background. The models are rolling out via an updated ChatGPT Voice, using GPT-Live-1 for paid users and GPT-Live-1 mini the default for free users, with API access planned later.SpaceXAI and Cursor launched Grok 4.5, a 1.5T mixture-of-experts model that achieves frontier-level performance on software engineering and agentic tasks. Grok 4.5 was trained with Cursor on real agent-interaction data, which has helped it achieve Opus 4.8 level coding benchmark scores: 64.7% on SWE Bench Pro, 62% on DeepSWE, 83.3% on Terminal Bench 2.1. Grok 4.5 costs only $2/$6 per million input/output tokens and is token efficient, making Grok 4.5 a highly attractive “daily driver” AI coding model. Grok 4.5 is available through Grok Build, Cursor and the SpaceXAI API, and developer tool environment Warp added Grok 4.5 support after the model’s launch.Meta’s Superintelligence Labs launched Muse Spark 1.1, a multimodal reasoning model with frontier-level benchmarks that gets Meta back in the frontier AI race. Muse Spark 1.1 is designed for computer use, coding, tool calling and multi-agent orchestration. Overall benchmarks put it in the class with Claude Sonnet 5 or GLM 5.2. It has a one-million-token context window and supports parallel subagent delegation and computer use across desktop, browser, and mobile.Meta is making Muse Spark 1.1 available in the Meta AI app and through a new Meta model API at $1.25/$4.25 per million input/output tokens to compete for third-party developers.Meta released Muse Image and previewed Muse Video, the first media-generation models from its new superintelligence organization. These models use agentic generation and test-time compute to improve accuracy, with Muse Spark reasoning and calling web search or code execution during generation. Muse Image can combine multiple visual references to generate an image and refine its own output.Meta released Muse Image in Meta AI, Instagram Stories, and WhatsApp. However, one Instagram feature that allowed people to generate images using photographs from public Instagram accounts created a backlash, with critics saying the feature could facilitate nonconsensual digital replicas. Meta quickly discontinued it.Muse Video is designed to generate video with native audio and is in early preview. Meta claims Meta Muse Video ranks third on text-to-video Arena rankings.Figure 4. Meta’s Muse Image model can compose an image from multiple input images and prompts.Reve released Reve 2.1, an update on their visual image generation model that is ranking No. 2 on the Text-to-Image Arena with a score of 1302. (The new Muse Image model ranks number 3 at 1280.) The update promises greater prompt understanding, world knowledge, and stronger foreign-text rendering. The model builds images through an underlying layout engine so that generated elements appear as editable layers, and it allows edits such as changing text inside an image while rebuilding the composition around the change.ByteDance launched Seedream 5.0 Pro, a multimodal image creation model focused on “complex information visualization and native multilingual input and rendering. image-text alignment, structural coherence, text rendering, and visual aesthetics. The release adds features including point-and-lasso editing, sketch rendering, material and color replacement, layer separation, multi-image fusion, and multilingual generation. Seedream 5.0 Pro scores 1231 and ranks 11 on the text-to-image leaderboard.Cognition released SWE-1.7, a coding model that the article says runs at 1,000 tokens per second on Cerebras. Cognition disclosed that the model is based on Kimi K2.7, notable because other coding products had used the same base. Cognition’s reinforcement-learning recipe improved its FrontierCode benchmark score from 30.1% to 42.3%.Mistral introduced Robostral Navigate, an 8B parameter model for ‘embodied navigation.’ The model uses a single RGB camera and plain-language instructions to navigate environments, and Mistral reports a 76.6% success rate on unseen R2R-CE environments, outperforming some systems that use depth sensors or multiple cameras. The model was trained in simulation on approximately 400,000 trajectories and is intended for robots used in logistics, manufacturing, delivery and hospitality.OpenAI released GPT-Realtime-2.1-mini for developers using the Realtime API, bringing reasoning and tool use to the Realtime API’s mini tier.Developed by Synthetic Sciences, OpenScience has been launched as an open-source alternative to Claude Science. OpenScience includes more than 250 research skills and works with any model rather than being tied to one proprietary model.Shanghai AI Lab released Agents-A1, a 35B parameter agentic mixture-of-experts model built on Qwen 3.5 35B and released under Apache 2.0. Agents-A1 supports a 256K-token context window and was trained for long-horizon work, with interestingly high scores on science-related benchmarks.LiquidAI released Antidoom, a method or model update aimed at reducing reasoning doom loops. The article says it reduced the loop rate on Qwen3.5-4B from 22.9% to 1%. It also says scores improved across the board after the change.Anthropic extended Fable 5 access on paid plans through July 12 and increased weekly Fable usage rate limits. Feeling the competition, Anthropic?We audited SWE-Bench Pro, one of the most widely used AI coding benchmarks, and found it no longer reliably measures frontier coding capability. We find 30% of SWE-Bench Pro tasks to be broken and are retracting our previous recommendation that the research community use it as a leading coding eval. - OpenAIOpenAI published a report finding that about 30% of SWE-Bench Pro problems are broken. The report suggests this caps the benchmark near 70% and should affect how readers interpret model claims based on SWE-Bench Pro. With frontier AI models scoring at that level on SWE-Bench already, other benchmarks such as Deep-SWE should be used.Anthropic published research describing a “global workspace” inside Claude, called J-space. This space is part of a model’s internal representations and contains roughly 25 active internal concepts. The J-space appears important for multi-step reasoning, because ablating it leaves basic abilities intact while reasoning collapses. The research also connected J-space to safety evaluation awareness, including a blackmail evaluation where removing test awareness changed model behavior.Google Research presented SensorFM, a health foundation model trained on one trillion minutes of wearable de-identified sensor data from five million consenting Fitbit and Pixel Watch users. The model transfers to 35 cardiovascular, metabolic, sleep, mental-health and lifestyle prediction tasks and reportedly outperformed specialized supervised baselines on 34 of them. Google also tested SensorFM as a grounding system for a personal health agent, although substantial clinical validation would still be necessary before diagnostic use.Anthropic researchers introduced GRAM, a modular training technique that routes knowledge from selected dual-use domains into components that can later be removed or enabled for authorized users. In experiments involving cybersecurity, virology and nuclear-physics information, removing a module suppressed the targeted capability without significantly degrading general model performance. The work is preliminary and has only been tested on models up to 5 B parameters.An internal memo reviewed by Reuters indicates that Meta plans to begin producing its internally designed Iris accelerator in September. Developed with Broadcom and manufactured by TSMC, the chip is intended to reduce Meta’s dependence on Nvidia and AMD as the company expands toward 14 gigawatts of computing capacity in 2027. Meta expects to spend as much as $145 billion on AI infrastructure during 2026.Tencent and Manus’ original investors are discussing buying the AI-agent company back from Meta for at least $2 billion, after Beijing had ordered Meta to unwind the acquisition while reviewing whether it violated Chinese investment rules.The Bank of England warned that the AI boom is becoming a financial risk. The central bank said that a reassessment of AI profitability might trigger falling equity prices and leveraged investment in AI companies and heavy infrastructure borrowing could amplify a market downturn. It also warned that existing financial rules were not designed for autonomous AI agents.Apple filed suit against OpenAI and two former Apple employees over alleged theft of hardware trade secrets. The lawsuit alleges that confidential information was taken by two to support OpenAI’s expansion into consumer hardware. OpenAI denies the accusations.Final thoughts: This is the most important week for AI releases this year. Two American AI labs, SpaceXAI and Meta, released top-tier AI models. GPT-5.6 sets a new standard for the AI model frontier, and we can breathe a sigh of relief that the Government hasn’t blocked access and we’ll get to use frontier AI.With AI is reaching a new level, remember to increase your ambition with AI. Take AI further than you have taken AI before. Give it harder questions, larger tasks, longer reasoning, and more autonomy.The future of the firm is a learning loop in which human capital and token capital compound. – Microsoft CEO Satya Nadella