On September 3, B.AI platform activated discounts on DeepSeek-V4-Flash and DeepSeek-V4-Flash-Vision-Exp, offering 50% off official peak pricing and as low as 25% off during off-peak hours. Additionally, four models – GLM-5.3-Flash, Qwen3.8-Flash, Hy3, and MiMo V2.5 – remain fully free.

CSTL, the team currently ranked first in the ARC-AGI-3 Kaggle competition with a score of 7.51%, has been identified as Mostik, a newly unveiled AI startup. The team reportedly has 15 members, including 12 PhDs and one Fields Medalist. The competition is still ongoing, with the second milestone due on Sept. 30 and final submissions closing on Nov. 2. Mostik is working on a method that lets AI models exchange internal states directly instead of relying on plain-text communication between agents. In a published demo, it used GLM-5.2 at 753B and Qwen3.5 at 4B, with the larger model handling prefill and the smaller model decoding the answer. GLM-5.2 reads the prompt but does not generate tokens, passing hidden states to the 4B model for answer generation. Mostik said the setup can cut the performance gap between the two models by about half, while using roughly 40% of the compute required by a medium-sized single model at the same accuracy. It also said the result still trails the 753B model on its own. The company has not disclosed which model and harness were used for the 7.51% ARC score, and said the 753B-plus-4B setup was a separate technical demonstration rather than the exact competition configuration.

Spanish AI startup Multiverse Computing touted Quasar 438B as Europe's top AI model, but the community quickly uncovered its ties to Zhipu's GLM-5.2. Artificial Analysis has labeled it based on GLM-5.2, and internal logs confirm identical capabilities. Performance comparisons show GLM-5.2 has a higher intelligence index, while Quasar offers faster output. Multiverse has not disclosed architecture or training methods, leaving the controversy unresolved.

Market updates from Sept. 2 to Sept. 3 spanned macro rates, whale activity, exchange products, AI financing, and sharp moves in token-linked speculation. TRON’s USDT supply rose by $4 billion over the past month to $94.27 billion, overtaking Ethereum. On-chain data also showed a Bhutan government-linked address moved 400 BTC worth about $30.62 million. In derivatives, a whale long roughly 45,000 ETH on Hyperliquid sold 1,500 ETH to add margin as liquidation pressure increased. Beyond crypto-native flows, U.S. and U.K. bond yields remained in focus. Former U.K. Prime Minister Liz Truss warned that surging sovereign yields could force emergency spending cuts, while JPMorgan Chase Private Bank’s Grace Peters said a move in the 10-year U.S. Treasury yield toward 5% could trigger a 5% to 8% pullback in global equities. Elsewhere, Coinbase launched regulated crypto and commodities derivatives for eligible investors in Canada, Bubblemaps flagged concentrated control risks around the CHUMP token, and several AI developments landed at once, including Lyte’s $165 million Series C, Google’s Gemini 3.8 Flash rollout, and Meta’s Muse Spark 1.3 update.

Thailand’s government launched the TH-AI Passport on August 31, offering free access to top-tier AI models for all citizens aged 15 and above, with a target of at least 5 million users. Within two weeks of pre-registration, over 1.3 million people signed up. The platform integrates 33 models from 15 providers, including GPT-5.6 Sol, Claude Opus 5, Gemini 3.1 Pro, and Grok 4.3, alongside image and video generation tools. A 'learn-to-earn' mechanism rewards users with credits for completing courses, potentially making it the largest national AI literacy experiment to date.

A wide set of crypto, policy, payments and market developments surfaced between Sept. 1 and Sept. 2. In the U.S., the Securities and Exchange Commission proposed updating registered transfer-agent rules for the first time in decades to reflect blockchain-based securities records and tokenization. At the international level, a G20 chair statement released by U.S. Treasury Secretary Scott Bessent said finance ministers and central bank governors recognized the growth potential of digital assets and pledged to create a clear path for “responsible innovation.” Separately, 21 global financial institutions said they plan to form a new company to issue fully reserve-backed stablecoins, starting with a U.S. dollar product and targeting launch in the first half of 2027. Market activity also stayed busy: U.S. spot Bitcoin ETFs pulled in about $3.5 billion in August, Felix Pago raised $200 million in Series B financing, Ethena rolled out a test version of Ethena Pay on Avalanche, and Bitfinex Securities listed five tokenized products tied to listed Bitcoin treasury companies. On the enforcement and security side, TRM Labs’ $95 million U.S. Homeland Security contract drew a lawsuit from Chainalysis, while Full Sail said it would wind down after a Switchboard oracle incident hit Virtue Money.

Zhipu AI released GLM-5.3 open weights on August 28 after a two-week safety delay. Within three days, Abliteration.ai launched a refusal-removed version, modifying weights to bypass safety mechanisms. The model's cyberattack performance improved significantly: ExploitBench rose from 24.4% to 54.4%, and ExploitGym completed 105 tasks in two hours (up from 29). The API only blocks child sexual content and self-harm by default, allowing other harmful content. The version is intended for offensive cybersecurity, red teaming, and agent testing.

Zhipu said it has reached large-scale, low-cost inference using 100,000 domestic chips, with per-token inference cost down 80% from the start of the year. The company tied that capacity to infrastructure it had already referenced before. When Zhipu released GLM-5.3-Flash, it said all online traffic for the model was carried by 100,000 domestic chips. Before that release, the anonymous model Ox-Alpha was also tested at scale on the same domestic computing cluster through OpenRouter and OpenCode. The disclosure links Zhipu’s current cost claim to an already used pool of domestic compute rather than a newly introduced setup.