AI News Briefing — July 10: OpenAI Ships GPT-5.6 and the Autonomous Agent ChatGPT Work, While the NYT Seeks Sanctions Over Hidden Evidence in Its Copyright Suit

Today's Headlines

OpenAI released GPT-5.6 for general availability — a three-tier family of Sol, Terra and Luna, plus an "ultra" setting that runs multiple agents in parallel
ChatGPT Work arrived, an autonomous agent with Codex built in that breaks hours-long tasks into steps; the desktop app reaches every plan, including Free
News organizations led by The New York Times filed a sanctions motion accusing OpenAI of faking its inability to search logs and hiding evidence
GPT-5.6 also becomes the preferred model in Microsoft 365 Copilot

OpenAI owned the day.

On the product side, the company shipped its new GPT-5.6 model family and the autonomous agent ChatGPT Work at once, pressing its case on performance, cost efficiency and workplace agents in a single stroke. On the same day, however, its copyright fight with The New York Times and other news organizations produced serious allegations over evidence discovery. The momentum of building and the weight of being held to account converged on one company.

Today's Top Three

OpenAI Ships GPT-5.6 — Three Tiers and a Parallel-Agent "ultra" Setting

OpenAI released its new GPT-5.6 model family for general availability.

GPT-5.6 spans three tiers: Sol, the flagship; Terra, a balanced model for everyday work; and Luna, the most cost-efficient. The number marks the generation, while Sol, Terra and Luna are durable capability tiers that can advance on their own cadence. Alongside them, OpenAI introduced "ultra," its highest-capability setting, which coordinates four agents in parallel by default to finish demanding tasks faster.

The performance pitch centers on efficiency per token. By OpenAI's own published figures, Sol sets a new high of 53.6 on Agents' Last Exam, an evaluation of long-running professional workflows across 55 fields, eclipsing Anthropic's Claude Fable 5 (adaptive reasoning) by 13.1 points. The company also claims the lower tiers, Terra and Luna, outperform Fable 5 at roughly one-sixteenth the estimated cost. These are OpenAI's own numbers, and the cost figures are simulated estimates; results split across benchmarks, so they should not be read as a settled ranking.

API pricing per million tokens is 5 dollars input and 30 dollars output for Sol, 2.50 and 15 dollars for Terra, and 1 and 6 dollars for Luna. The rollout starts today across ChatGPT, Codex and the OpenAI API and continues globally over the next 24 hours.

On safety, OpenAI says GPT-5.6's cyber safeguards block roughly ten times more potentially harmful activity than previous models. Retaining access to its most cyber-capable features will require enabling Advanced Account Security with hardware-backed passkeys by September 1. The general release follows a limited preview, which The Verge reports came with government approval. TechCrunch, for its part, notes that the review's methods and reviewers have not been disclosed, leaving the process opaque.

ChatGPT Work Arrives — An Autonomous Agent Built on Codex

On the same day, OpenAI launched ChatGPT Work, an agent that autonomously carries out tasks across a user's apps and files.

ChatGPT Work breaks a goal into smaller steps and can stay with complex projects for hours, completing them independently. It gathers context from connected apps such as Slack, Microsoft Teams and Google Drive, and turns it into finished materials — spreadsheets, slides, documents and web apps. It also offers Scheduled Tasks for recurring work and, on desktop, Computer Use, which operates apps and browsers directly on the user's machine.

The foundation is Codex, the technology that began as a coding agent for developers. According to OpenAI, more than 5 million people use Codex every week, and more than 1 million of them use it for work outside software development. With this launch, the Codex app merges into the new ChatGPT desktop app, and the existing desktop app is renamed ChatGPT Classic.

Sites, a feature for building web apps and sharing them through a URL, also arrived in public beta.

The rollout begins today on web and mobile for Pro, Enterprise and Edu plans, expanding to Plus and Business over the next few days. In the new desktop app for Mac and Windows, Chat, Work and Codex are available on every plan, including Free. On security, OpenAI says its auto-review layer, which has advanced models check important actions before they happen, blocked 100 percent of attempts to extract protected data during adversarial red teaming.

The Verge frames ChatGPT Work as OpenAI's answer to Anthropic's Claude Cowork. Autonomous agents that produce documents and working materials are becoming the next battleground beyond general-purpose chat.

News Organizations Seek Sanctions Against OpenAI Over Hidden Evidence

News organizations led by The New York Times filed a motion on July 9 asking the court to sanction OpenAI for allegedly concealing evidence in their copyright suit.

According to the motion, OpenAI spent two years telling the court that searching ChatGPT logs at scale was technically infeasible and unduly burdensome. But an April re-deposition of an OpenAI privacy engineer revealed, the plaintiffs allege, that the company had held two de-identified samples — 78 million and 10 million logs — since before the litigation began, and had already searched them as part of research into a filter to block regurgitation of copyrighted content.

The plaintiffs also point to the cost of the discovery process itself. The news organizations spent eight months searching a sample of just 20 million logs, far below the 120 million they originally requested. OpenAI then used AI to make 19 billion redactions to that sample — so many that the court found it "unusable." The plaintiffs further allege that OpenAI deleted or compressed billions of logs that a preservation order required it to keep.

The requested sanctions go far. The plaintiffs ask the court to bar OpenAI from using the 20-million-log sample, to treat it as established fact that OpenAI's outputs substantially reproduced their copyrighted material, and to instruct the jury that OpenAI deleted billions of logs. An OpenAI spokesperson countered that the motion is an attempt to invade users' privacy and said the company will keep defending fair use.

The episode shows the main front in generative-AI copyright litigation shifting: alongside the substantive question of whether training is fair use, the fight over evidence discovery is now center stage. If the sanctions are granted, OpenAI will find it much harder to argue that no substantial market harm occurred, dealing a serious blow to its fair use defense.

More to Note

Models

OpenAI announced that GPT-5.6 becomes the preferred model in Microsoft 365 Copilot, powering Word, Excel, PowerPoint, Chat and Cowork, with Microsoft accessing the model directly through the OpenAI API. - GPT-5.6 is now the preferred model in Microsoft 365 Copilot (OpenAI)
OpenAI confirmed it will sunset ChatGPT Atlas, the agentic browser it announced last October, targeting an August 9 deprecation date. Its browser features carry over to the new ChatGPT desktop app and a Chrome extension. - The ChatGPT browser is already dead (The Verge AI)
Meta released Muse Spark 1.1, an agentic coding model. Reuters reports pricing of 1.25 dollars per million input tokens and 4.25 dollars per million output tokens, close to Anthropic's Claude Haiku 4.5. Mark Zuckerberg promoted the release in his first post on X in three years. - Meta enters the crowded AI coding battle with Muse Spark 1.1 (TechCrunch AI)
Anthropic introduced Reflect, a beta feature that visualizes patterns in how you use Claude over time and nudges users to reconsider their working relationship with the assistant. - A new way to reflect on how you use Claude (Anthropic)
OpenAI turned its bio-misuse bug bounty into a standing program, the OpenAI Bio Bounty Program. The reward for a universal jailbreak doubles from 25,000 to 50,000 dollars; testing against GPT-5.5 ends July 27, after which GPT-5.6 becomes the sole focus. - OpenAI Bio Bounty Program (OpenAI)

Products

Adobe announced enterprise versions of Firefly Graph, which saves and shares Firefly editing workflows, and Firefly Creative Production, which lets non-designers mass-produce variations — a push to turn AI editing steps into reusable organizational assets. - Adobe adds team workflow sharing to Firefly (PC Watch, in Japanese)

Policy

Google introduced transparency disclosures showing whether an ad was created or edited with AI across Search, YouTube and Discover. The information appears under "How this ad was made" in My Ad Center; ads made with Google's own generative tools are labeled automatically, while others rely on advertiser self-disclosure. - Expanding AI transparency in ads (Google Blog)
Microsoft said it will use AI to find more vulnerabilities and expand the volume of Windows 11 security updates. As attackers speed up discovery with AI, detection moves earlier, while code review stays with humans. - Microsoft plans bigger Patch Tuesdays with AI (The Verge AI)

Business

Anthropic appointed Ben Bernanke, former Federal Reserve Chair and Nobel laureate in economics, to its Long-Term Benefit Trust, the independent body overseeing responsible AI development — a hire aimed at AI's effects on the economy and workforce. - Ben Bernanke appointed to Anthropic's Long-Term Benefit Trust (Anthropic)
Paris-based voice AI startup Gradium reopened its seed round to new investors including Nvidia, bringing the round's total to 100 million dollars. The company, a spinout of French AI lab Kyutai, counts automaker Renault among its customers. - Paris-based AI voice startup Gradium raises $100M seed backed by Nvidia (TechCrunch AI)
Meta will begin mass production of its in-house AI chip, MTIA, in September, designing with Broadcom and manufacturing at TSMC. The goal is lower GPU costs and less reliance on Nvidia; Meta expects capital expenditures of 125 to 145 billion dollars this year. - Meta's new AI chips will begin production in September (TechCrunch AI)

Other

Anthropic launched Inviting Hard Questions, an initiative that collects the public's concerns about AI and publicly tracks its responses. It builds on a survey of 52,000 Americans and interviews with 81,000 Claude users across 159 countries, alongside a new research effort, the Anthropic Institute. - Inviting hard questions (Anthropic)
A profile of SoftBank's company-wide RAG platform, used by roughly 19,000 employees, detailed how approval workflows for data uploads and identity-provider-linked access controls got the project past the security and governance hurdles that stall most enterprise deployments. - Inside SoftBank's company-wide RAG platform used by 19,000 employees (ITmedia AI+, in Japanese)
Research from AI-detection firm Pangram mapped the spread of AI-generated content on social media. On LinkedIn, more than 40 percent of long-form posts were flagged as fully AI-generated, and the platform accounted for 62 percent of all AI content detected while making up only about a third of scanned items. - AI Content Is Everywhere on Social Media, Especially LinkedIn (Pangram)
A UC San Diego team had surgeons remotely pilot a low-cost general-purpose humanoid robot to remove gallbladders from live pigs, with the results published in Nature — a step toward far cheaper remote surgery, though latency and other constraints remain. - Humanoid robots controlled by surgeons did world-first operation on live pigs (Ars Technica AI)

Source: Selected by the editorial team from the AI news inbox (collected July 10, 2026 — 30 items: 8 primary, 22 secondary).