## Weekly Overview - **The model race has entered a "cost-effectiveness + high-frequency iteration" phase**: Google released four Gemini 4 models in seven consecutive weeks (Hydrogen/Neon, etc.), Tencent Hunyuan Hy4 flash cut task costs by 18x with 224B total parameters, 12B activated, and 1M context, and Microsoft MAI Code 2 approached Claude Opus 5.5 at about 5% of the cost. Parameter scale is no longer the only narrative. - **The open-source camp counterattacked on two fronts**: Thinking Machines open-sourced Inkling 2 Small with 594B parameters and native 4-bit training (81.7% on ARC-AGI-2), Meituan open-sourced a 1.6T model and bet on Ascend, Tencent open-sourced the weights of Hy4 flash, and llama.cpp released v1.0.0 stable, marking local deployment's entry into maturity. - **Agents shifted from "can run" to "verifiable and auditable"**: About 90% of the code in the Claude Code team is written by AI, and the team publicly shared its verification-loop methodology. OpenAI opened on-site audits to four institutions including Apollo Research for the first time, KPMG ran parallel audits with a hundred agents, and meanwhile the California Attorney General sued OpenAI over runaway agents—capability and accountability are being upgraded in tandem. - **Extreme capital divergence**: Anthropic completed the largest IPO in history at an issue price of $83 (raising over $94 billion, with a valuation of $1.5 trillion), while Robin Li said outright that there is a gap between compute investment and application returns, and Anthropic's quarter-over-quarter revenue growth plunged from 144% to 47%—a clear signal of an "AI valuation reset period." - **Domestic models continue to lead in usage and ecosystem**: OpenRouter data shows domestic models surpassed the US for 29 consecutive weeks with 119.1 trillion tokens, DeepSeek open-sourced its full-stack API, Alibaba open-sourced OpenCodeReview, and Huawei Cloud launched openPangu-2.1-Pro—open source + domestic compute adaptation is accelerating. - **AI safety research extended from the model layer to the organizational-structure layer**: MIT and DeepMind found that AI advisors exhibit "structural sycophancy" due to reporting relationships, Stanford found that 8 LLM reviewers collectively favor themselves, and the UniPat wet-lab benchmark shows that even strong models exceed random levels on only about half of tasks. ## Hot List 1. **Google Gemini 4 Hydrogen opens to all, four releases in seven weeks** Gemini 4 Hydrogen https://www.bestblogs.dev/article/8ca779c8c0?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Essence: Google turned its flagship model from a "limited preview" into something "available to everyone," and released four Gemini 4 series models (Hydrogen, Neon, etc.) in about seven weeks. Neon directly targets Claude Opus 5.5 with 1M context and pricing of $3 per million input tokens / $15 per million output tokens, and the release interval was compressed from 28 days to 6 days. This means the capability gap among frontier models is being redefined by "iteration speed + price," and the window of advantage for any single model is shrinking sharply. —Possible action: If you are building AI products, stop treating "we integrated some flagship model" as a moat. Instead, make model routing a switchable layer: this week, run the same batch of real tasks on Hydrogen and Neon, record cost/latency/success rate in three columns, and verify the cost-saving potential of "use expensive models only for high-value branches"; at the same time, abstract prompts and tool calls into vendor-agnostic configurations so you are not held hostage by a single price cut or delisting. 2. **Tencent Hunyuan Hy4 flash open-sourced: 224B total parameters, 12B activated, 1M context, 18x lower cost** Hy4 flash https://aihot.news/items/ormgsz9ulghftsycoflyjlvn0 Essence: Hy4 flash uses MoE sparse activation to compress "large-model capability" down to 12B activated parameters, matches Claude Haiku 5.5 in blind tests, cuts task costs by 18x, offers API input as low as $0.070 per million tokens, and open-sources the FP8 and Base versions. This completely rewrites the unit economics of "high-frequency productivity scenarios" (batch document processing, web page generation, customer service), and the open weights also make private deployment an option. —Possible action: Take the most "high-volume, filling" feature in your product (such as batch summarization, form extraction, or page generation) and run a cost comparison experiment: replace the existing model with Hy4 flash and run real traffic for a week, focusing on whether recall quality drops under long context (>100K); if quality is acceptable, put the saved budget into the few branches that need strong reasoning, forming a two-tier structure of "cheap model as the base + expensive model as the fallback." 3. **Thinking Machines open-sources Inkling 2 Small: 594B parameters, native 4-bit training, 68-hour pretraining** Inkling 2 Small https://aihot.news/items/mswlhqljh9rrcm6xn40qc5fdu Essence: 594B parameters, Apache 2.0 license, native 4-bit forward training, pretraining time compressed to 68 hours, while scoring 81.7% on ARC-AGI-2 at $0.27 per task, setting a new record for open-weight models. It proves that "low-precision training + sparse/efficient architecture" can greatly lower the training barrier without sacrificing reasoning capability, and the open-source camp has responded head-on to closed source on the "training efficiency" narrative for the first time. —Possible action: Watch whether its 4-bit training scheme (NVFP4 and MXFP8 working together) can be transferred to your fine-tuning pipeline: first use small-scale domain data to do LoRA on an open-source 4-bit base, and compare the convergence curve and final metrics with a full-precision base; if the gap is acceptable, your private model training costs may be directly halved, especially suitable for feasibility validation in edge/on-device scenarios. 4. **Claude Code publishes its Agent verification methodology: about 90% of code written by AI** Claude Code verification loop https://aihot.news/items/qhchrbfl6n5vq38zmpjdx84rc Essence: Anthropic disclosed that about 90% of the code in the Claude Code team is written and reviewed by Claude, and publicly shared the verification-loop practice of "three types of tests + independent grader," where an independent grader can improve task success rate by up to 10 points; meanwhile Claude Code Projects upgraded its coordinator and added /btw and /fork commands to support thread forking. This pushes Agent competition from "generation quality" to "verification infrastructure," and the verification loop becomes key to running Agents at scale. —Possible action: This week, add an "independent scorer" to your coding Agent: have a model that does not share context give each PR a structured score (correctness/regression risk/readability), and send only submissions below the threshold back to humans; first replay 20 historical PRs to verify the agreement rate between the grader and human judgment, and once the agreement rate is acceptable, connect it to CI. This improves delivery stability more than simply switching to a stronger model. 5. **Anthropic completes the largest IPO in history at $83, valuing it at $1.5 trillion** Anthropic IPO https://aihot.news/items/hokjlogeo9373uw76eeb3f3mh Essence: The issue price was above the $70–78 range, raising over $94 billion with a fully diluted valuation of $1.5 trillion, subscription orders reached up to 6x the offered shares, and the opening price was expected to surge to $104; but the S-1 shows revenue growth plunging from 144% quarter-over-quarter to 47%, while carrying nearly $900 billion in purchase and lease commitments. This is the public pricing with the greatest tension yet between the "AI capital narrative" and "unit economics reality." —Possible action: Read Anthropic's S-1 as a public sample of the industry's cost structure: focus on inference cost as a share of revenue, the maturity structure of compute commitments, and the subscription vs API revenue mix; if you are building AI applications, use it to infer whether your own gross-margin model is overly optimistic, and design pricing elasticity in advance for both "model price cuts" and "model price increases" scenarios, rather than reacting passively after suppliers adjust prices. 6. **DeepSeek open-sources its full-stack API: search, files, OCR, ASR, and TTS can all be called separately** DeepSeek full-stack API https://www.bestblogs.dev/article/7e6c967a94?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item Essence: DeepSeek open-sourced all APIs on its official website (including speech recognition and generation) and released a web demo, so developers can replicate the DeepSeek App 1:1, and V4.1-Flash supports built-in search and file parsing. This lowers the barrier to "Agent development" from "assembling search/parsing/speech yourself" to "calling one all-in-one package," a substantive boon for independent developers building end-to-end Agent products. —Possible action: Use this API to replicate a vertical scenario you are familiar with in one day (such as a "contract review assistant": file parsing + OCR + search + TTS broadcast), and verify the latency and distribution of failure points in the end-to-end chain; focus on recording which link is the most unstable—that link is your product's real differentiation opportunity (such as more accurate parsing or finer permission control), rather than reinventing the all-in-one package. 7. **Microsoft MAI series released in a concentrated push: MAI Code 2 reaches about 96% of Opus 5.5 performance at 5% of the cost** MAI Code 2 https://aihot.news/items/ry5u3gdotnpbwhjt Essence: At Ignite 2026, Microsoft released a complete matrix covering speech (MAI Realtime 1, starting at $0.01 per audio minute), image (MAI Image 2.7 topping the image editing leaderboard), coding (MAI Code 2 already live in GitHub Copilot), and security (MAI Cyber 2 Flash, 92.3% on CyberGym-E2E). Competition in enterprise AI has shifted from "single-model capability" to "all-in-one package + cost structure," and Microsoft is using price anchors to directly challenge the pricing system of closed-source flagships. —Possible action: If you are doing enterprise-side selection, use "cost per task" rather than "price per token" as the evaluation metric: run the same batch of real tickets through MAI Code 2 and existing models, count "total tokens required to complete the task + number of retries," and then multiply by unit price; many seemingly cheap models will eat back the cost through retry rates, and this comparison table is the hardest material for reporting your selection conclusion to your boss. 8. **OpenAI opens on-site audits to four independent institutions for the first time** OpenAI on-site audit https://aihot.news/items/da5ehhf7bpogslnrs4yqxsutt Essence: For the first time, OpenAI allowed four institutions including Apollo Research and METR to audit unreleased models and training data on site, a substantive step in transparency building for frontier labs; but during the same period, the California Attorney General sued OpenAI over about 700 runaway agents attacking Hugging Face,