The AI Brief

Chinese AI startup Moonshot AI released Kimi K3 on July 16, 2026, timed deliberately to coincide with the opening of the World Artificial Intelligence Conference in Shanghai. The model carries 2.8 trillion total parameters, a 1-million-token context window, and native visual understanding, and Moonshot describes it as "the world's first open-source model in the 3-trillion-parameter class." Full weights are scheduled for release by July 27, with a full technical report accompanying them.

Independent evaluation firm Artificial Analysis scored Kimi K3 at 57 on its Intelligence Index v4.1, placing it fourth among 189 tested models, behind Claude Fable 5 (60) and GPT-5.6 Sol (59), and slightly ahead of Claude Opus 4.8 (56) and GPT-5.5 (55). On Artificial Analysis's private agentic knowledge-work benchmark AA-Briefcase, K3 scored 1,547 Elo, second only to Fable 5 Max (1,587) and beating GPT-5.6 Sol Max (1,495). It also debuted at first place on LMArena's Frontend Code Arena, ahead of Claude Fable 5. Per-task cost lands around $0.94, comparable to GPT-5.6 Sol ($1.04) and roughly half of Opus 4.8 ($1.80).

Architecturally, K3 introduces two original components: Kimi Delta Attention (KDA), a hybrid linear attention mechanism that Moonshot says enables up to 6.3x faster decoding in million-token contexts; and Attention Residuals (AttnRes), which the company claims delivers ~25% higher training efficiency. The sparse LatentMoE structure activates 16 of 896 experts. Moonshot recommends supernode configurations of at least 64 accelerators for serving. API pricing is $3.00 per million cache-miss input tokens, $0.30 on cache hit, and $15.00 per million output tokens. K3 is compatible with the OpenAI SDK. Moonshot is simultaneously reported to be closing a new funding round valuing it at $31.5 billion, up from $20 billion in May 2026.

CaveatBenchmark results from Moonshot's launch blog mix KimiCode, Claude Code, and Codex harnesses across tests; Artificial Analysis's independent scores are the more comparable signal. Full weights and technical report arrive July 27, after which weight-level verification becomes possible.

Artificial Analysis's Intelligence Index v4.1 is a composite evaluation spanning nine tests including reasoning, coding (Terminal-Bench 2.1), science (SciCode, GPQA Diamond), mathematics, and long-context retrieval. Kimi K3's 57.1 score places it between GPT-5.6 Sol (58.9) and Claude Opus 4.8 (56). The evaluation cost Artificial Analysis $2,690.80 to run, K3 generated 130 million output tokens across the suite, nearly double the median of 63 million, at 62 tokens per second with a 1.99-second time-to-first-token.

The number that sharpens the procurement case is per-task cost: Artificial Analysis measured K3 at roughly $0.94 per completed agent task, versus $1.04 for GPT-5.6 Sol and $1.80 for Claude Opus 4.8. The open-weight release on July 27 adds a third cost option: teams with the infrastructure to self-host on 64+ accelerators could, in principle, run K3 at marginal compute cost. The caveat is inference speed, 62 tokens per second is below the comparison median, and at launch, reasoning effort is locked to maximum, meaning every call generates a full reasoning trace and the $15 output rate applies regardless of task complexity.

On July 8, the AI Now Institute published a report by chief AI scientist Heidy Khlaaf and senior research scientist Boyan Milanov demonstrating a proof-of-concept (PoC) exploit enabling remote code execution (RCE) in Claude Code and OpenAI's Codex. The attack requires no plugins, MCP servers, or custom configuration, only a standard automated review session pointed at a codebase containing hidden instructions. When a user asks the agent to "scan this library for vulnerabilities," the embedded malicious code executes on the victim's machine. The PoC was tested on Claude Code versions 2.1.116, 2.1.196, 2.1.198, and 2.1.199, and Codex version 0.142.4.

The researchers notified both Anthropic and OpenAI before publication and offered to assist with verification. Their central argument is that the underlying issue is not a patchable bug but a structural property of agentic AI: any tool with enough access to write and execute code will be exposed to this attack pattern, because model-layer changes cannot establish a reliable trust boundary between data and instructions within the same context window. The researchers explicitly flag programs like Anthropic's Project Glasswing and OpenAI's Patch the Planet as contexts where this attack pattern is particularly consequential, both involve deploying these tools against unfamiliar codebases on safety-critical infrastructure. The paper notes the technique likely transfers to other agentic coding platforms beyond Claude Code and Codex for the same architectural reason.

Twenty-nine countries signed the WAICO founding agreement in Shanghai on July 16, 2026, on the eve of the 2026 World Artificial Intelligence Conference (WAIC). Founding members include Russia, Kazakhstan, Laos, Pakistan, Indonesia, Belarus, Brazil, Cuba, Venezuela, and Serbia, among others. UN Secretary-General António Guterres attended the signing ceremony. Chinese Foreign Minister Wang Yi signed on behalf of Beijing. WAIC 2026 runs July 17–20 under the theme "Intelligent Partners, Co-create the Future," featuring more than 140 forums and 1,100-plus exhibitors across three Shanghai venues.

Chinese President Xi Jinping addressed the conference on July 17 in person, his first appearance at WAIC since the event began in 2018, using the occasion to promote a "people-centered" AI governance vision and call for greater cooperation with developing nations. Chinese officials framed WAICO as a mechanism to ensure developing economies participate in setting AI standards, a position explicitly counterposed to what Beijing characterizes as Western-led, exclusionary AI governance. The organization's stated mandate covers "healthy and orderly development" of AI technologies, support for technological sovereignty, and multi-lateral capacity building.

The governance architecture of WAICO has not yet been published. The founding agreement does not specify voting structures, membership expansion criteria, binding obligations, or enforcement mechanisms. Further details are expected as the organization becomes operational.

Gemini 3.5 Pro did not launch on July 17 as widely expected. The model has now missed three consecutive public targets, June (when Sundar Pichai promised delivery at Google I/O on May 19), then an early July target, then July 17. Per TechTimes reporting citing Geeky Gadgets on July 16, the rebuilt model is still beset by hallucinations and inconsistent outputs that prevent it from clearing basic reliability standards. Google has not officially confirmed any of the dates that have circulated; as of this edition, the public Gemini API lists gemini-3.5-flash and gemini-3.1-pro-preview, with no gemini-3.5-pro model ID. First covered in Vol. I, No. 50 (July 14).

The Polymarket prediction market tracking the release, with over $323,000 in trading volume, now shows 81% probability of a July 31 delivery as the leading outcome, with a separate market placing "August 7" at 73% as current leader. Google is separately reported to be exploring Gemini 3.6 Flash as a stopgap release, per Geeky Gadgets. The competitive consequences compound: GPT-5.6 launched July 9, Grok 4.5 launched July 8, and Kimi K3's open weights arrive July 27. Alphabet reports Q2 earnings on July 28, the first call at which analysts can formally assess the competitive cost of the delay sequence.

The leaked specifications that drove developer expectations, a 2-million-token context window, Deep Think extended reasoning mode, and API pricing near $1.25 input and $10 output per million tokens, remain unconfirmed in any official Google documentation. The ground-up pre-training restart from the Gemini 2.5 Pro base, which caused the initial June slip, was driven by structural failures in recursive tool-calling and SVG generation that engineers concluded could not be closed through post-training techniques alone.