AI Daily: Moonshot AI launches Kimi K2.6 with 300-agent swarms, ChatGPT ships Images 2.0

The pace of AI development makes it genuinely difficult to keep track of what actually matters on any given day. Today brought a new reasoning model out of China built around coordinating armies of AI sub-agents, a meaningful upgrade to how ChatGPT handles text inside generated images, and a fresh injection of capital into the agent research space. Here is what you need to know.

Moonshot AI has released Kimi K2.6, a new model built around orchestrating swarms of up to 300 AI sub-agents working in parallel on complex, long-horizon tasks. The Chinese AI lab, founded in 2023 and known for its work on long-context architectures and efficient mixture-of-experts designs, dropped the release this morning, 84 days after its predecessor Kimi K2.5. Benchmark figures have not yet appeared in the initial release materials, but the headline capability is agent orchestration at a scale that goes well beyond what the team had previously demonstrated. The framing is squarely aimed at software engineering, research automation, and the kind of multi-day workflows where a single model running a single chain of thought is simply not adequate.

Kimi K2.5 sat at 87.6% on GPQA Diamond and 76.8% on SWE-Bench Verified according to independent leaderboard data, placing it competitively with established Western models. Whether K2.6 improves on those numbers will become clearer as evaluations arrive over the next few days. The more interesting question is whether the 300-sub-agent architecture produces reliably useful outputs or amplifies errors at scale. Parallel agent systems are appealing in theory: break a problem into pieces, solve each concurrently, recombine. In practice, coordinating that many agents without drift or cascading mistakes is a genuine engineering challenge. Moonshot AI’s track record on long-context tasks gives them a credible foundation, but the proof will be in what developers report when they put K2.6 to work on real projects.

OpenAI has shipped an Images 2.0 model for ChatGPT, and it is drawing significant attention for how well it handles text rendering inside generated images. Text in AI-generated imagery has been one of the persistent weak points of the field for years. Words come out garbled, fonts warp, letters duplicate themselves, and captions become abstract decoration rather than readable copy. Images 2.0 appears to have made a substantial step forward: early users report that labels, signs, and short passages of text inside generated images are reading correctly far more consistently than previous versions managed. Generating accurate text requires the model to understand both what individual characters should look like and where they belong within the broader visual composition, and prior systems consistently struggled with one or the other.

If the improvement holds across more demanding prompts, it closes a gap that has limited AI image generation for anyone working in infographics, marketing materials, or branded social content. OpenAI also confirmed today that it is scaling Codex, its AI coding agent, to enterprise customers worldwide. The two announcements together point to a deliberate product push, with capability improvements on the creative side and broader commercial rollout on the developer side, ahead of what the industry widely expects will be an intensely competitive second half of the year.

A new AI research lab called NeoCognition has raised 40 million dollars in seed funding with the stated ambition of building agents that learn in a way that more closely resembles human cognition rather than brute-force pattern completion. The funding round, reported today by TechCrunch, plants the lab in a crowded but rapidly evolving space. The core claim is that AI agents should build up competence continuously from experience rather than being frozen at the point of training and then deployed as fixed systems. It is a well-established criticism of current agent architectures: they can be impressively capable on familiar tasks, but novel situations and unexpected errors tend to expose the limits quickly.

Forty million dollars at seed stage reflects how much investor appetite there is for anything positioned as the next leap in agent capability. Whether NeoCognition’s approach produces measurable differences in how agents handle novelty will only become apparent when the team publishes research or ships working products. What the round does confirm is that investment in AI agent infrastructure is accelerating in a meaningful way. The market is increasingly confident that the next wave of enterprise value from AI will come from agents capable of executing multi-step, multi-day work autonomously, and that confidence is now flowing into early-stage bets at a rate not seen since the large language model funding wave of 2022 and 2023.

At a glance, other stories from today and this week:

• Sam Altman publicly described Anthropic’s Mythos cybersecurity model as “fear-based marketing” in comments reported by TechCrunch, escalating the public rivalry between the two labs over AI’s role in national security work. [21 Apr]
• Tim Cook has announced he is stepping down as Apple CEO, with hardware engineering chief John Ternus set to take over. Apple’s AI roadmap under new leadership will be closely watched. [21 Apr]
• Clarifai, a facial recognition company, has agreed to delete three million photographs of OkCupid users that were used without explicit consent to train AI models, following a settlement with the US Federal Trade Commission. [21 Apr]
• OpenAI is expanding its Codex coding agent to enterprise customers worldwide, confirmed in a separate post today. [21 Apr]
• Anthropic has secured a further five billion dollars from Amazon as part of a broader deal under which Anthropic has committed to 100 billion dollars in cloud spending through AWS over the coming years. [20 Apr]
• Anthropic CEO Dario Amodei visited the White House this week following reports that the NSA has been using Mythos, the company’s cybersecurity-focused AI model, for classified cyber defence work. [20 Apr]
• Google has expanded its Gemini assistant inside Chrome to seven additional countries, continuing the company’s push to embed AI capabilities directly into the browser. [20 Apr]