OpenAI pulled the trigger on GPT-5.6 general availability today, ending a government-mandated limited preview that lasted nearly two weeks. Meanwhile, Chinese labs are eating token market share from the bottom, and Mistral just taught a robot to navigate with nothing but a cheap RGB camera. The frontier is getting wider at the top and cheaper at the bottom simultaneously. Both things are true. Progress is not linear.

GPT-5.6 Arrives: Three Tiers, Government-Approved

OpenAI shipped GPT-5.6 to general availability Thursday, rolling out three tiers: Sol (flagship, $5/$30 per million input/output tokens), Terra (balanced, $2.50/$15), and Luna (cheapest, $1/$6). The model family clears the previous GPT-5.5 on virtually every benchmark OpenAI reports, with Sol hitting 53.6 on Agents Last Exam, 13.1 points ahead of Claude Fable 5 in adaptive reasoning. On the Artificial Analysis Coding Agent Index, Sol hits 80, beating Fable 5 by 2.8 points while using less than half the output tokens and costing about a third less.

The most interesting claim from Sam Altman on CNBC: GPT-5.6 Sol is 54% more token-efficient on agentic coding tasks. “Every enterprise now is thinking about spend and the value they are getting in exchange for AI,” Altman said. That is not a capability pitch. It is a cost pitch. The model also introduces “ultra” mode, which coordinates four agents in parallel by default for complex tasks, trading higher token use for faster results. On Terminal-Bench 2.1, Sol ultra hits 91.9%.

The release was not normal. OpenAI previewed GPT-5.6 on June 26 but held it to a “small group of trusted partners” at the U.S. government is request. Altman said the company worked with Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent, and National Cyber Director Sean Cairncross during the approval process. Washington is now on the release calendar.

OpenAI also used the launch to double the Bio Bug Bounty reward from $25,000 to $50,000 for universal jailbreaks against GPT-5.6, and separately published a deep audit of SWE-Bench Pro finding that roughly 30% of its 731 tasks are broken due to overly strict tests, underspecified prompts, and low coverage. When the benchmark you use to measure progress is 30% noise, every leaderboard position is suspect. That is an uncomfortable finding for a field that has been racing on these scores.

ChatGPT Becomes ChatGPT Work

Alongside GPT-5.6, OpenAI launched ChatGPT Work, an agent layer that gathers information across your apps (Slack, Teams, Google Drive, SharePoint, email, calendars, CRMs) to produce finished spreadsheets, slides, docs, and web apps. It can stay with complex projects for hours, breaking them into smaller steps and completing them independently. Scheduled Tasks let it run on a recurring schedule or trigger on events. Nearly 100% of teams inside OpenAI, including finance and sales, already use it internally. The finance team reduced month-end close from days to hours.

This is not a chatbot update. It is a productivity-suite play. ChatGPT Work rolls out today for Pro, Enterprise, and Edu plans, with Plus and Business following in days. GPT-5.6 also becomes the preferred model in Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat, and Cowork. The OpenAI-Microsoft pipeline now runs both ways: natively in Copilot and via the API for direct Microsoft 365 integration.

GPT-Live: Full-Duplex Voice

OpenAI also shipped GPT-Live, a voice model built on a full-duplex architecture that can listen and speak simultaneously. It says “mhmm” and “yeah” while you talk. It can interrupt, pause, and resume naturally. When a question needs deeper reasoning or web search, GPT-Live delegates to GPT-5.5 in the background and keeps the conversation going while it waits. Two versions (GPT-Live-1 and GPT-Live-1 mini) are rolling out globally today.

Full-duplex is the real breakthrough here. Previous cascaded systems chained speech-to-text, a language model, and text-to-speech, losing information at each step. Turn-based models like Advanced Voice Mode were smoother but rigid. GPT-Live processes input and output continuously, making interaction decisions many times per second. This is the architecture that could finally make voice AI feel like talking to a person.

DeepSeek Is Eating the Bottom of the Market

While OpenAI and Anthropic push prices up at the frontier, Chinese labs are grabbing developer share from below. DeepSeek costs about $0.87 per million output tokens, compared to $30 for OpenAI and $25 for Anthropic. On Vercel, DeepSeek traffic share jumped from under 1% in May to 17% by July. OpenRouter reports DeepSeek usage doubled between January and June 2026. Lindy, a San Francisco AI assistant startup, moved off Claude to DeepSeek after AI costs for its 25-person team exceeded personnel costs. The CEO said the switch saved “millions.”

US labs still build the strongest models. But many routine business tasks (sorting email, drafting responses, summarizing documents, customer service) do not need the strongest model. They need a cheap one. Airbnb and Cursor have already added Chinese models to their mix rather than fully replacing US providers. The routing market is here: cheap tiers absorb bulk work, frontier models handle the hard edge cases. That is not a temporary phase. That is the structure.

Fable 5 Returns With a New Classifier and a Industry Framework

Anthropic redeployed Claude Fable 5 globally on July 1 after the US government lifted export controls imposed on June 12. The model was offline for 19 days after Amazon researchers found a jailbreak that let Fable 5 identify software vulnerabilities and produce exploit code. Anthropic trained a new safety classifier that blocks the reported technique in over 99% of cases, though it also flags more benign requests during routine coding. The model holds the top SWE-Bench Pro score at 80.4% and costs $10/$50 per million tokens, the steepest rate in the market.

More significantly, Anthropic proposed an industry-wide framework for scoring jailbreak severity, developed with Amazon, Microsoft, Google, and other Glasswing partners. The framework aims to give AI developers a consistent standard for triaging new findings, launching capable models with greater safety, and communicating risk levels to government and industry. This is the response to a real governance gap: when a jailbreak report can trigger export controls, the industry needs a shared definition of what actually counts as severe.

Mistral Ships a Robot Navigation Model and Formal Math Prover

Mistral released Robostral Navigate, an 8B parameter model that enables robots to navigate complex environments using a single RGB camera. No LiDAR, no depth sensors. It hits 76.6% on R2R-CE validation unseen, beating the best multi-sensor approach by 4.5 points. The model is trained entirely in simulation across 6,000 scenes and 400,000 trajectories, then generalizes to real-world obstacles it never saw in training. It runs on wheeled, legged, and flying robots. The training trick is prefix-caching: compressing an entire episode into a single sequence reduces training tokens by 22x, turning months of training into days.

Mistral also shipped Leanstral 1.5, a formal math verification model that saturates miniF2F completely (100% on both validation and test), solves 587 of 672 PutnamBench problems, and found 5 previously unknown bugs across 57 open-source repositories. It is Apache-2.0 licensed, 119B total parameters with only 6B active. The model proved AVL tree time complexity (O(log n)) over 2.7 million tokens and 22 context compactions. Formal verification just became practical for real codebases.

Quick Hits

OpenAI – GPT-5.6 is now the preferred model in Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat, and Cowork. Direct API integration means Microsoft pulls from OpenAI is API, not just the native hosting path.

Hugging Face – LeRobot v0.6.0 ships world model policies (VLA-JEPA, FastWAM, LingBot-VA) that learn to imagine the future before acting, plus six new simulation benchmarks and a deployment CLI with DAgger-style human-in-the-loop corrections. The vLLM transformers backend now matches or beats native vLLM throughput across dense, tensor-parallel, and MoE configurations with a single flag.

NVIDIA – Nemotron 3 Ultra tuned with LangChain Deep Agents achieves the highest accuracy among open models on LangChain is benchmark, at 10x lower inference cost per run than leading closed models. No model retraining. Every gain came from engineering the harness around the model. NemoClaw for LangChain Deep Agents is the open reference blueprint, available now.

Mistral – Studio now gives prompts and skills a system of record: versioned, owned, and traceable. Immutable versions, rollback, audit logs, and classification labels. The closed loop between what runs in production and what you version is the difference between cataloging your AI and governing it.

Anthropic – Claude Science, an AI workbench for scientists, is now available. It integrates the tools and packages researchers use most, produces auditable artifacts, and provides flexible access to computing resources.

Rundown for July 10, 2026. Sources: Yellow, OpenAI, Anthropic, Mistral, NVIDIA, Hugging Face.