AI & design briefing for Thursday, July 9, 2026

tldr

• OpenAI ships GPT-Live full-duplex voice; Anthropic's Managed Agents hit 96% of top-model performance at 46% cost; GPT-5.6 cleared for rollout with GPT-6 expected within a month.
• Prime Intellect raises $130M at $1B for open training stack; Nvidia loses ~$1T in market value in under two months; Tokyo University releases LLM agents into the real internet.
• Cognition ships SWE-1.7 near-frontier coding model; MiniMax targets 2.7T-parameter open-source model; Meta launches Muse Image with Muse Video preview.

Top Stories

• OpenAI Ships GPT-Live: Full-Duplex Voice That Talks and Listens at Once. OpenAI launched GPT-Live, a new voice architecture for ChatGPT that supports full-duplex communication, meaning it can listen and speak at the same time, like a phone call rather than a walkie-talkie. The model can interject with acknowledgments ("mhmm," "yeah"), be interrupted naturally, support live translation, and delegate harder questions to a frontier model behind the scenes. GPT-Live-1 rolls out globally for paid users (Go, Plus, Pro) as the default Voice model, with GPT-Live-1 mini for free users. Available on iOS, Android, and ChatGPT.com, but not yet in Temporary Chats, desktop app, Work, Codex, or custom GPTs. In OpenAI's tests, users picked GPT-Live over old Advanced Voice Mode in roughly 3 of 4 head-to-head chats, with expert science scores nearly doubling. API access is coming soon; current Realtime API pricing lists audio at $32/$64 per 1M input/output tokens. ^[1]
• Anthropic's Claude Managed Agents: 96% of Top-Model Performance at 46% of Cost. Anthropic shared two cost-optimization patterns for using Claude's most powerful model (Fable 5) without paying full price on every call. The Advisor pattern has Sonnet 5 do all the work, calling Fable 5 only at key decision points (~once per task), hitting 92% of Fable 5's SWE-Bench Pro score at 63% of cost. The Orchestrator pattern has Fable 5 plan the task and hand off heavy lifting to parallel Sonnet 5 workers, reaching 96% of Fable 5's performance on BrowseComp at 46% of the price. One example: verifying 20 facts cost $1.61 with the split team vs $4.00 solo. Each sub-agent keeps its own cache, so repeated context isn't billed twice. ^[2]
• SpaceXAI + Cursor Release Grok 4.5: Strongest Grok Yet at Undercut Pricing. SpaceXAI (formerly xAI, post-Cursor acquisition) launched Grok 4.5, the first model built jointly with Cursor after the $60B acquisition. Grok 4.5 shows strong performance on coding, agentic, and knowledge-work tasks, with benchmarks in similar ranges to Claude Opus 4.8 and GPT-5.5. It runs at 80 tokens/second (rivaling flash models) while claiming a 4x efficiency gain over Opus 4.8. Pricing: $2/$6 per million input/output tokens vs $5/$25 for Opus 4.8. Usage is temporarily free inside Cursor and Grok Build. Musk pitched it as "an Opus-class model, but faster, more token-efficient and lower cost," with an even bigger model coming next month. ^[3]
• OpenAI Audits SWE-Bench Pro: Finds 27–34% of Tasks Broken, Pulls Recommendation. OpenAI audited SWE-Bench Pro, the benchmark widely used to rank AI coding models, and found it no longer trustworthy. Their analysis flagged 200 tasks (27.4%) as broken; human reviewers found issues in 249 tasks (34.1%). Problems included hidden requirements not mentioned in problem descriptions, tests that reject valid solutions, grading criteria missing correct answers, and contradictory instructions. Correct solutions were failing, models penalized for writing good code. The benchmark hit saturation near a 70% noise ceiling, with frontier models jumping from 23.3% to 80.3% in eight months, suggesting scores were inflated. OpenAI has formally retracted its recommendation. ^[4]
• GPT-5.6 Cleared for Rollout; GPT-6 Expected Within a Month. OpenAI's GPT-5.6 was cleared for broad rollout after extra Commerce Department testing, with Sol, Terra, and Luna set to launch Thursday. According to a scoop, GPT-5.6 will be OpenAI's last release in the 5.x line, with GPT-6 landing in about a month (possibly even this month), built on a noticeably bigger pretrain than the ~4T "Spud" base under 5.5 and 5.6. The reason for the rush: OpenAI wants something that can go toe-to-toe with Anthropic's Fable 5 and the upcoming Fable 5.1. DeepSeek isn't waiting either, with an imminent V4 GA and a bigger model in the works to challenge MiniMax's upcoming 2.7T Pro. ^[5]

Deeper Cuts

• Prime Intellect Raises $130M Series A at $1B Valuation for Open Training Stack. Prime Intellect secured $130M in Series A funding, crossing $100M in annualized sales in its first year, to help companies train their own agentic systems with an open stack for frontier models. The raise signals strong investor appetite for decentralized/open AI training infrastructure. ^[6]
• Microsoft Replacing Some OpenAI/Anthropic Models with In-House MAI in Copilot. Microsoft is reportedly replacing some OpenAI and Anthropic model calls with in-house MAI models in Copilot apps such as Excel and Outlook, in a cost-cutting move. This signals a potential shift in the Microsoft-OpenAI partnership dynamics. ^[8]
• Nvidia Loses ~$1T in Market Value in Under Two Months. Nvidia lost roughly $1 trillion in market value in under two months, dragging its valuation back down to pre-AI-boom levels. Separately, AI companies are racing toward custom chips, but fab, memory, and packaging bottlenecks still keep the supply chain tight. ^[12]
• Tokyo University Releases LLM Agents Into the Real Internet: One Sold an Ebook. Tokyo University released LLM agents into the real internet for 12 weeks. One of the agents autonomously sold an ebook on the open internet, demonstrating real-world agentic capability beyond controlled benchmarks. ^[13]
• OpenAI Chief Futurist Joshua Achiam Departs After 9 Years. OpenAI's chief futurist Joshua Achiam announced he is leaving the company after a nine-year stretch, calling it "a decade where centuries happened." ^[18]
• OpenAI Buys Northslope for Enterprise Deployment Muscle. OpenAI's Deployment Company agreed to buy Northslope, giving the lab more enterprise implementation capabilities. The acquisition makes enterprise rollout part of the GPT-5.6 launch. ^[19]
• FlowWM Predicts Multiple Possible Futures for Self-Driving. FlowWM is a model that predicts multiple possible futures for self-driving applications using richer visual features, improving autonomous driving safety. ^[21]
• Stop reporting UX activity and report business outcomes. Nielsen Norman Group argues that UX teams should report business outcomes (impact on revenue, cost, risk, speed, retention) rather than merely reporting UX activity, in order to demonstrate value and secure resources. ^[22]
• 98% browser support still excludes millions of users. An essay exploring why a 98% browser support metric still excludes millions of users, making the case that "98% isn't very much" when considering global web accessibility. ^[23]
• Most AI work can wait: prioritize routing over model choice. Tom Tunguz argues that most AI work runs on cheap local models, and teams should prioritize routing over model choice, you don't always need the most expensive frontier model for every task. ^[24]

Tools and Launches

• Cognition Releases SWE-1.7: Near-Frontier Coding Model at Low Cost. Cognition released SWE-1.7, an in-house model for its Devin agent built on China's open-source Kimi K2.7, nearing GPT-5.5 and Opus-level scores while costing far less. The model is part of a trend of cheaper near-frontier coding models challenging incumbent pricing. ^[9]
• MiniMax Targeting 2.7T-Parameter Open-Source Model for Q3. MiniMax is reportedly targeting a Q3 release for a 2.7T-parameter model, which would land 6x bigger than the lab's flagship and larger than any current open model. This adds further pressure to closed frontier pricing. ^[10]
• Meta Launches Muse Image, Previews Muse Video; AI Glasses Testing Reported. Meta launched Muse Image and previewed Muse Video, while Instagram reuse controls turned the launch into a privacy fight. Meta is also reportedly testing AI glasses that capture a wearer's surroundings with cameras and audio, pulling its hardware push into another privacy debate. Meta also previewed a detector for Muse-generated images and videos using invisible watermarks, though rate limits still apply. ^[11]
• Zyphra Releases Zamba2-7B with Hybrid Mamba2-Transformer Architecture. Zyphra released Zamba2-7B, a chat model combining Mamba2 and transformer architectures in a hybrid design. Available on HuggingFace with 29,100 downloads. ^[14]
• Meituan/LongCat Open-Source Talking Avatar Tool from Photo + Audio. A Chinese lab released LongCat-Video-Avatar-1.5, an open-source tool that turns a photo and audio clip into a talking avatar. Available on HuggingFace. ^[15]
• Gemini API Managed Agents Get Background Execution & MCP Support. Google DeepMind added background execution, remote MCP connections, and custom functions to Gemini API Managed Agents for longer-running developer agents. API pricing applies. ^[16]
• Cohere Releases Transcribe Arabic: Open-Weight Arabic Speech Recognition. Cohere released Transcribe Arabic, an Arabic speech-recognition model with open weights and hosted API access, free and open-source. ^[17]
• Monogram Emerges from Stealth with $40M Seed for AI-Generated iOS Interfaces. Udemy co-founder Eren Bali's Monogram emerged from stealth with a $40M seed round, pitching an iOS app that swaps text walls for instantly generated interfaces. ^[20]
• InkField Gallery: humans and agents learning from each other. A collection of JSON recordings showing humans and AI agents learning from each other, presented as an experimental art project. ^[25]
• Paper Shaders: zero-dependency shaders as React components or GLSL. Ultra-fast, zero-dependency shaders usable in projects as React components or GLSL, aimed at graphic designers and developers. ^[26]

Prompt of the day