AI News Daily 2026-07-10

01OpenAI merges ChatGPT and Codex into one app and launches “ChatGPT Work”

Published: 2026-07-10 (announced 2026-07-09) · Category: Model release / Company moves

Facts

OpenAI has folded the Codex app into the ChatGPT desktop application, consolidating Chat, Work and Codex into a single piece of software. The headline addition is ChatGPT Work, an agent feature built on GPT-5.6 that gathers information across multiple applications, breaks a stated goal down into steps, and returns finished artifacts — spreadsheets, slide decks, web applications and the like — rather than a block of prose.

• Chat, Work and Codex now live in one desktop app instead of three surfaces.
• ChatGPT Work runs on GPT-5.6 and is available on every plan, including Free.
• The Atlas browser is scheduled to be wound down in stages.

Background

Codex began life as a developer-facing, coding-specific product. Rolling it into the same application as general chat, and putting an agent on top that produces office deliverables, is a deliberate repositioning: the same underlying capability that writes and repairs code is being sold as a way to finish ordinary knowledge work. Extending the feature to the Free tier rather than reserving it for paid plans indicates OpenAI wants reach and habit formation before it wants margin.

Implications

For enterprises, the practical effect is that workflow automation can increasingly be attempted inside ChatGPT alone, which lowers the cost of choosing and switching between competing point tools. The flip side is concentration: an organization that standardizes on one agent for both engineering and back-office output has a single vendor dependency across two very different parts of the business. The winding down of Atlas is a reminder that surfaces in this product line are not permanent, which matters for anyone building process around them.

Competitively, this formalizes an “office-replacement agent” race against Anthropic and Google.

Sources: MacRumors, Neowin, Crypto Briefing.

02Anthropic expands “Claude Cowork” to mobile and web

Published: 2026-07-07 to 2026-07-09 (phased rollout) · Category: Company moves

Facts

Anthropic has begun rolling out Claude Cowork — an agent that keeps working on tasks autonomously in the background — to iOS, Android and the web, in stages. The beta is going to Max plan subscribers first. Chat and Cowork are being merged into a single home tab, so work can be continued and monitored as the user moves between devices.

Anthropic also published an analysis of roughly 1.2 million Cowork sessions. The distribution is the notable part:

• Software development use: 8.7% of sessions.
• Business-process work such as report writing and data aggregation: 33.4%, the largest single category.

Background

The timing places this within days of OpenAI’s ChatGPT Work announcement. Both companies are moving the same underlying agent technology off the developer’s desk and onto phones and browsers, where non-engineers actually work. Starting with Max subscribers is a familiar pattern for capacity-constrained agent features: the heaviest, most tolerant users absorb the rough edges first.

Implications

The usage figures are the most decision-relevant item in today’s set, because they come from observed behavior rather than vendor positioning. If only 8.7% of sessions on an agent originally framed around coding are actually coding, then the addressable market for these products is routine business process automation, not engineering productivity. Buyers evaluating agents on coding benchmarks alone are measuring the wrong thing for the workload their staff will bring.

Cross-device continuity also changes the governance question: background execution that follows a user from laptop to phone is harder to bound with desktop-only controls.

Sources: TechCrunch, VentureBeat, 9to5Mac.

03Meta ships “Muse Spark 1.1”, a coding and agent model

Published: 2026-07-09 · Category: Model release / Company moves

Facts

Meta Superintelligence Labs, led by Alexandr Wang, announced Muse Spark 1.1, a model the company says matches GPT-5.5 and Claude Opus 4.8 on many agent evaluations. It carries a 1 million token context window aimed at long-running tasks such as large-scale code migration and bug fixing, and became available the same day through the Meta Model API and Meta AI.

• Pricing: $1.25 per million input tokens, $4.25 per million output tokens.
• Context: 1 million tokens, positioned for long-horizon work rather than single-turn answers.
• Wang disclosed that a more powerful next model, codenamed “Watermelon”, is in training.

Background

Meta had not been a front-line participant in the AI coding market. Entering it with same-day API availability, published token pricing and a named successor already in training is the posture of a company treating agentic coding as a core business line rather than a research demonstration.

Implications

With Meta joining OpenAI and Anthropic, the three-way pattern is now clear: the major players have all placed agentic coding at the center of their product strategy. For enterprises this widens the shortlist meaningfully — more vendors, more price points, and more leverage in procurement. It also raises the switching-cost question sooner than expected, since capability claims at parity ($1.25 in / $4.25 out against comparable frontier models) push differentiation toward integration, reliability and governance rather than raw benchmark scores.

Sources: CNBC, Fortune, Axios.

04Illinois governor signs an AI safety law billed as the strongest in the United States

Published: 2026-07-06 (effective 2028-01-01; follow-up coverage since) · Category: Regulation and policy (United States)

Facts

Governor Pritzker of Illinois signed the AI Safety Measures Act (SB315). It includes the first provisions in the United States requiring frontier AI developers to publish plans for addressing “catastrophic risk” and to undergo an annual independent third-party audit. Where an incident occurs, the law obliges reporting to the state within as little as 24 hours.

• Mandatory publication of a catastrophic-risk response plan.
• Annual independent third-party audit.
• Incident reporting to the state in as little as 24 hours.
• Penalties of up to $3 million for violations.
• Effective January 2028.

Background

Federal AI regulation in the United States has stalled, and states have moved into the gap. Illinois is notable less for acting than for the specific obligations it chose: publication, external audit and a short incident clock are compliance mechanics, not principles, and they are enforceable.

Implications

Companies operating across multiple states now face a compounding compliance problem, because each state that acts is likely to specify obligations slightly differently. The eighteen-month runway to January 2028 is enough time to build an audit-ready evidence trail, and the organizations that treat it as a deadline rather than a distant date will not be assembling one under pressure. Expect the Illinois text to be borrowed from as other states legislate.

Sources: Illinois Governor’s Office, Transparency Coalition, StateScoop.

05The EU publishes an action plan on cybersecurity and artificial intelligence

Published: 2026-07-07 · Category: Regulation and policy (EU)

Facts

The European Commission published the EU Action Plan on Cybersecurity and Artificial Intelligence. It sets out a coordinated approach for member states, companies and public bodies to address the cybersecurity and resilience challenges that come with state-of-the-art AI models. The Commission also plans a call for third-party evaluators to strengthen the capacity to assess AI models before they reach the market, with a target of being operational in 2027. Separately, the AI Act’s transparency rules are due to take effect in August 2026.

Background

The EU is filling in the operating detail of the AI Act in stages rather than all at once. An action plan plus a procurement route for external evaluators is how a regulator builds the capability to enforce what it has already legislated — the rules exist before the institutional capacity to check them does.

Implications

For any company offering AI services inside the EU, the August 2026 transparency deadline is the near-term forcing function, and evaluation and reporting arrangements need to be in place before it. The 2027 target for third-party evaluation capacity signals what the steady state looks like: independent assessment as a routine precondition of market access, in the EU as in Illinois.

Source: Inside Privacy.

06Japan: neoAI signs capital and business alliances with five critical-infrastructure firms at once

Published: 2026-07-09 · Category: Japan

Facts

neoAI, a generative AI startup that emerged from the Matsuo Laboratory at the University of Tokyo, announced simultaneous capital and business alliances with five companies tied to critical infrastructure: Aozora Bank, Johnan Shinkin Bank, Kyushu Electric Power, Kyoto Chuo Shinkin Bank and The Bank of Iwate. Combined with its May 2026 alliance with Japan Post Bank, that brings the total to six partners.

Background

The composition of the list is the point. Two of the five are shinkin banks — regional cooperative financial institutions serving local businesses — alongside a regional bank, a commercial bank and a power utility. These are not the large national institutions that usually appear first in enterprise AI announcements.

Implications

Generative AI adoption in Japanese critical-infrastructure sectors such as finance and electric power is reaching beyond the largest firms into regional financial institutions. That is evidence of the domestic market for industry-specific AI deployment broadening rather than concentrating, and it suggests the constraint on adoption at that tier is now partnership and integration capacity rather than appetite.

Source: Jiji Press.

07Editor’s note

Three of today’s six items describe the same movement from different vantage points. OpenAI, Anthropic and Meta released office-work agents — ChatGPT Work, Claude Cowork and Muse Spark 1.1 — within a few days of each other. Competition that was recently about assisting programmers has expanded to white-collar work generally, and Anthropic’s session data gives the shift an empirical anchor: 8.7% coding against 33.4% business process.

The two regulatory items point the same direction from opposite sides of the Atlantic. Illinois legislated at the state level; the EU is writing the operating detail beneath an existing AI Act. Both converge on the same obligation set for model developers — accountability and audit. The mechanism differs, the destination does not.

The Japanese item sits underneath both. While the frontier labs compete over capability and regulators build audit machinery, adoption at the level of a regional shinkin bank is what determines whether any of it reaches ordinary operations. Six partnerships is a small number in absolute terms; what makes it worth noting is which institutions they are.

The practical read for a decision-maker: evaluate agents against the business-process workloads your staff will actually bring rather than coding benchmarks, and start building the evaluation and reporting evidence that August 2026 in the EU and January 2028 in Illinois will require.