Weekly AI Roundup: Claude Limits Doubled, SpaceX IPO, Microsoft Model Data Contradiction

SWE-bench Pro 59%, input $0.6/million tokens. On June 1st, MiniMax released its flagship model, MiniMax M3, claiming it is the first domestically produced open-source model to integrate cutting-edge programming, million-token context, and native multimodal capabilities into one. Architecturally, it introduces MiniMax Sparse Attention (MSA), compressing per-token computation in million-token contexts to 1/20th of the previous generation, with prefill speeds over 9x faster and decoding speeds over 15x faster. The model natively supports text, image, and video input, and can directly control desktops to complete automated tasks.

In programming benchmarks, MiniMax M3's SWE-bench Pro score of 59% slightly exceeds GPT-5.5's 58.6%, and its 83.5 on the BrowseComp Agent benchmark surpasses Claude Opus 4.7's 79.3. However, a significant gap remains compared to Claude Opus 4.8's SWE-bench Pro score of 69.2%. Output pricing is $2.4/million tokens, with a concurrent 50% promotional discount. All scores are self-reported by MiniMax, achieved with agent scaffolding on their own infrastructure, and independent verification is not yet available. Open-source weights and a technical report are expected to be uploaded to HuggingFace within 10 days.

According to multiple media reports on June 3rd, DeepSeek is advancing its first-ever external funding round, sized at approximately 50 billion RMB (about $7.4 billion), with a post-investment valuation between 350 billion to 400 billion RMB (about $52-59 billion). Founder Liang Wenfeng personally subscribed to about 20 billion RMB ($2.8 billion), accounting for nearly 40% of this round's total. Tencent invested approximately 10 billion RMB, and CATL about 5 billion RMB. The National AI Industry Investment Fund also participated. DeepSeek thus abandons its previous convention of not accepting external capital, where all prior funding came from Liang Wenfeng's quantitative hedge fund, High-Flyer.

On June 1st, Unitree Technology's IPO on the STAR Market was approved by the Shanghai Stock Exchange Listing Committee. From acceptance on March 20th to approval took only 73 days, setting a new 2026 speed record. Unitree plans to raise 4.202 billion RMB, and based on a public float ratio of no less than 10%, its estimated overall valuation is around 42 billion RMB. In 2025, its humanoid robot shipments exceeded 5,500 units, with a number one global market share of 32.4%, and a core component self-research and production rate over 90%. However, in Q1 2026, its year-over-year revenue growth rate dropped from 333% in the previous year to 68%, and adjusted net profit declined 53% year-over-year, showing a significant deceleration. If successful, Unitree will become the 'First Humanoid Robot Stock' on the A-share market.

On June 3rd, Moonshot AI launched the Kimi Work Beta program, positioning it as a general-purpose local agent for knowledge workers. Unlike the previously mentioned Kimi WebBridge which controls browsers, Kimi Work processes tasks locally. In complex scenarios, it can autonomously split a task into up to 300 sub-agents for parallel execution, suitable for cross-document organization, batch data analysis, and multi-step workflows. The same day, ByteDance's Coze 3.0 went online, supporting flexible collaboration among multiple users and agents, and independent project management. Its biggest change is the ability to directly integrate with local agent platforms like Claude Code and Codex CLI, meaning Coze is no longer limited to pure cloud orchestration and can mix local and cloud usage.

On June 2nd, Alibaba Cloud's Tongyi team launched Qwen3.7-Plus on the Bailian platform. Building on the previous Qwen3.7-Max, the new model adds image and video understanding, deep reasoning, and tool-use capabilities. It has a built-in Agentic RL mechanism that allows it to continuously optimize its own strategy based on execution feedback. This week, the three companies are converging on the Agent theme: Moonshot AI focuses on local task processing, ByteDance on cross-platform orchestration, and Alibaba on multimodal reasoning.

Doubao is going to charge. On June 3rd, an official announcement revealed four subscription tiers: Free, Standard at 68 RMB/month, Enhanced at 200 RMB/month, and Professional at 500 RMB/month. Annual subscriptions are 0/688/2048/5088 RMB respectively. The Professional plan targets software development, data analysis, financial research, and other professional scenarios, offering dedicated computing resources, API calls, and priority response during peak times. The basic free tier retains daily chat, Q&A, copywriting, translation, and image generation, promising to remain permanently free, unlimited, and with no speed throttling. The official launch is expected in late June.

As of March 2026, Doubao's daily token call volume surpassed 120 trillion, a full 1000 times the volume since its launch in May 2024. In an online poll, 76% of users chose 'Will not use it anymore,' with only 8% indicating they would pay. Looking at competitors, Kimi membership is priced at 49 RMB/month, Zhipu Qingyan at 49 RMB/month, and Baidu Wenxin Yiyan at 30 RMB/month. Doubao's Standard tier starting price of 68 RMB/month already surpasses its peers. For now, Qianwen and Yuanbao have chosen to hold firm and remain free, but they too will inevitably face the same computing cost bill sooner or later.

Three years, 1 billion monthly active users. Third-party data shows that ChatGPT surpassed 1 billion global MAUs in May 2026. It took only three years from launch to reach this milestone, faster than all previous applications including TikTok, YouTube, and Google Maps. Year-over-year growth was 62%, with paying subscribers exceeding 50 million. In the same week, OpenAI launched GPT-5.5 Instant, replacing Canvas with native 'Writing Blocks' and 'Code Blocks'. Plus and Pro users can leverage conversation history and saved memories to achieve personalized responses across sessions.

On June 3rd, Codex launched the preview of its Sites feature. Available to Business and Enterprise users, it can directly generate and host interactive websites and applications. On the same day, six new job-role plugins were added, expanding Codex from a coding tool to non-technical roles like creative production, sales, and customer service, covering 62 applications and 110 skills. The ChatGPT memory system was upgraded to a 'dreaming' mechanism, which automatically updates outdated memories—for example, correcting 'You are going to Singapore next month' to 'You have already been to Singapore' after a trip. Lockdown mode was also opened to all individual users, which, when enabled, restricts web browsing and deep research to reduce the risk of data leakage from prompt injection.

On June 5th, Anthropic announced that Claude Cowork usage limits would be doubled for all paid plans, lasting until July 5th. This marks the third round of capacity expansion, following the permanent doubling of Claude Code's five-hour limit in early May and a temporary 50% weekly limit increase in mid-May. It's backed by the computing power contract with SpaceX's Colossus 1 data center, giving access to over 220,000 NVIDIA GPUs. On June 3rd, Claude Code changed the trigger word for dynamic workflows from 'workflow' to 'ultracode'. Previously, many users accidentally triggered agent orchestration scripts when mentioning 'workflow' in normal conversation. After the change, only the explicit use of 'ultracode' or /effort ultracode will initiate multi-agent coordination mode.

In the same week, Nous Research released the public preview of Hermes Desktop, supporting Windows, macOS, and Linux. The Hermes Agent, nicknamed 'Hermès' by the community, previously only had a command-line version. It already has a stable user base within the Chinese AI developer community. The desktop launch now allows those unfamiliar with the terminal to use it directly.

Jensen Huang and Microsoft CEO Satya Nadella shared the stage in Taipei, announcing the reinvention of the Windows PC using ARM. The RTX Spark is a 'superchip,' integrating a 20-core ARM CPU and a Blackwell GPU with 6144 CUDA cores connected via NVLink C2C, supporting up to 128GB of LPDDR5X unified memory, reaching 1 petaflop of AI computing power. NVIDIA claims it can locally run 120-billion parameter models, agents with million-token context, 4K AI video generation, and render 3D scenes exceeding 90GB. It launches in the fall, with the Microsoft Surface Ultra among the first 8 laptops. Over 30 OEMs will launch RTX Spark devices.

The next-generation AI superchip platform, Vera Rubin, was announced to be in full production, claiming a 10x increase in large-scale agent throughput over Grace Blackwell. The open-source Nemotron 3 Ultra, featuring 550 billion parameters, is designed specifically for always-on agents, claiming a 5x inference speed improvement and 30% cost reduction over comparable closed-source models, and is already adapted to mainstream agent platforms. The DGX Station for Windows was unveiled as 'the world's most powerful desktop AI supercomputer.' The DLSS 4.5 Ray Reconstruction feature will roll out to all RTX graphics cards in August, with support added for 11 new games. Jensen Huang teased a 'surprise product that no one has been told about' coming later in the year.

$135 per share, a $75 billion fundraising target, a $1.77 trillion valuation. SpaceX will begin trading on the Nasdaq under the ticker SPCX on June 12th. Goldman Sachs is the lead underwriter, joined by Morgan Stanley, Bank of America, Citigroup, and JPMorgan. If successful, this will be the largest IPO in history, surpassing Saudi Aramco's record $29.4 billion raised in 2019. The roadshow began on June 4th, several days earlier than originally planned. A 30% retail allocation will be distributed directly through Robinhood, Fidelity, and Schwab—an unprecedentedly high retail share for an IPO of this scale. Musk will still hold over 82% of voting rights after the offering.

A week before the IPO (June 5th), SpaceX disclosed a new computing contract to the SEC. Starting October 2026, Google will pay SpaceX $920 million monthly for the contract lasting until June 2029, totaling about $30 billion. Google will rent approximately 110,000 NVIDIA GPUs along with associated CPUs and storage. A Google Cloud spokesperson stated this is to meet the higher-than-expected customer demand for the Gemini Enterprise and Agent platforms. The compute power comes from SpaceX's two Colossus supercomputing centers in Memphis, originally xAI facilities. Adding this to the previous Anthropic contract, SpaceX's total disclosed compute contracts now exceed $70 billion.

On June 1st, Anthropic confidentially filed its S-1 registration statement with the SEC, targeting a post-money valuation of approximately $965 billion. Morgan Stanley and Goldman Sachs are jointly leading the offering. Reportedly, its annualized revenue has surpassed $47 billion, a more than 4x increase from about $9 billion at the end of 2025. If SpaceX and OpenAI also go public this year, 2026 will see three trillion-dollar AI companies entering the public markets simultaneously.

Three days later (June 4th), the Anthropic Institute published a report titled 'When AI Builds Itself.' The report shows that Claude currently writes over 80% of Anthropic's production code. Internal engineering output is growing 8x per quarter, and the internal Mythos Preview model achieved a 52x speedup on an ML optimization benchmark, compared to just 3x for the previous generation. Co-founder Jack Clark and Institute head Marina Favaro call in the report for a coordinated global pause on frontier AI development by major AI companies, to buy time for alignment research and social institutions. They stress this isn't a unilateral pause, needing simultaneous execution and verifiable rule constraints across multiple countries, including the US and China.

Critics are calling this 'fear marketing by a $965 billion company.' In the same week, OpenAI published a rebuttal, arguing that it should be governments, not private companies, who decide on such a pause.

On June 2nd, at Build 2026, Microsoft unveiled its first self-developed reasoning model, MAI-Thinking-1. It features 35 billion active parameters in a roughly 1 trillion total parameter sparse Mixture-of-Experts (MoE) architecture and supports a 256k context window. Microsoft AI head Mustafa Suleyman emphasized at the launch that the model was 'trained entirely on enterprise-grade, clean, commercially licensed data with no distillation from third-party models.' Yet three days later, media uncovered that the paper accompanying the model explicitly states its training data includes Common Crawl—an open web crawl library that provides no copyright assurances—totaling 24.2 billion pages after filtering. Django co-founder Simon Willison was the first to spot this contradiction. As of June 5th, Microsoft had not issued a public response.

On June 1st, GitHub Copilot transitioned entirely to a pay-as-you-go billing model. One AI Credit equals $0.01. The Pro plan includes a $10 monthly credit allowance; Pro+ includes $39; and Business includes $19 per user. Code completions and edit suggestions remain unlimited, but chat, Agent mode, and code review now consume credits based on token usage. The developer community reacted furiously. Multiple people reported their monthly bills jumping from $29 to $750 or even $3,000. A single Agent coding session could burn through $30 to $40. By default, there is no spending cap; users must manually enable a hard budget limit in their billing dashboard. GitHub provided transitional promotional credits for Business and Enterprise users for June through August (an additional $30 and $70 per month, respectively), but individual developers received no such buffer.