## Weekly Overview - **The model competition has entered a phase of "tiered pricing + specialization"**: Anthropic splits general-purpose and security/life-science scenarios with the dual Fable 5.2 / Mythos 5.2 line and cuts prices by 20%, while Google pushes Gemini 4 Carbon to 99.1% on ARC-AGI and then uses Gemini Decision 1 at $0.10/M tokens to enter decision routing. Leading vendors are beginning to price by "task type" rather than "parameter scale." - **Agents are moving from "can do work" to "can spend money, can collaborate, can be audited"**: AI is beginning to procure services on users' behalf, Cursor/Amp are doing multi-agent orchestration, and Google has donated Agent Substrate to the CNCF. But at the same time, a CMU benchmark shows that none of 9 agents passed expert review, and Microsoft exposed a dormant-neuron backdoor—capability and trustworthiness are becoming decoupled. - **The dual arms race in compute and capital is being repriced**: OpenAI's compute spending by 2030 is estimated at $614 billion, Anthropic's S-1 reveals $800 billion in procurement commitments, and TSMC approved $1 billion to take stakes in two model companies. Meanwhile, the Nasdaq 100 fell 6.49% in a single day, and Anthropic's expected IPO valuation was cut from $2 trillion to $1.2–1.5 trillion—the market is beginning to discount the "AI infrastructure narrative." - **Chinese open-source models captured 89% of global enterprise open-source AI spending**: Kimi 61%, DeepSeek 15%, GLM 11%, while Meta Llama fell to 7%. Combined with Doubao integrating Kimi/DeepSeek and Baidu building a Token Exchange, the "replaceability" of the model layer is becoming a prerequisite for product design. - **On-device and embodied intelligence are landing in parallel**: EmbeddingGemma 2 (740M/567MB) can run on phones, Gemini Robotics 2.5 reaches 96% sorting accuracy at a Mercedes-Benz factory, Runway open-sources the 27B Praxis-1, and Apple HALO reaches a 91% success rate in 6 real homes. - **Security and compliance are turning from "recommendations" into "hard constraints"**: New York's RAISE Act opens registration and requires 72-hour incident reporting, Australia plans to mandate reporting of safety violations, Cloudflare warns of a two-year AI vulnerability wave, and Alex Stamos says the industry will enter a "dark age." ## Hot List 1. **Anthropic's dual models top the rankings, API prices cut 20% but per-task cost rises** Claude Fable 5.2 / Mythos 5.2 — https://www.bestblogs.dev/article/081c474cdf Essence: Anthropic uses Fable 5.2 (general-purpose) and Mythos 5.2 (cybersecurity/life sciences) to segment scenarios. Its intelligence index tops the chart at 65 points, speed improves 30%, and API prices drop 20%, but at the same time "per-task cost rises" is reported—meaning "lower unit price ≠ lower total cost," and token efficiency and task completion rate become new selection dimensions. Combined with its IPO price range of $70–$78 and planned fundraising of up to $90.4 billion, model capability is being directly converted into a capital-market narrative. —Possible: If you are doing model selection, don't just look at $/M tokens. Immediately run your 20 most typical real tasks through "total cost per task = unit price × retry count × token consumption," and put Fable 5.2, GPT-6.2, and Gemini 4 Carbon in the same table for comparison. If you build security/compliance products, prioritize applying for Mythos 5.2's institutional channel and validate a bundle of "specialized model + industry workflow." 2. **Google opens Gemini 4 Carbon, hitting 99.1% on ARC-AGI** Gemini 4 Carbon — https://www.bestblogs.dev/article/0c99d7db6c Essence: This is Google's strongest public model to date, focused on complex programming, scientific research, and cybersecurity. ARC-AGI 99.1% means it is nearing saturation on the abstract reasoning benchmark. Ordinary users need a Google AI Ultra subscription to use it—capability tiers are directly tied to payment tiers. It turns the "strongest model" from an API shelf item into a subscription moat, which for independent developers means the cost of obtaining "the strongest capability" has been raised. —Possible: Don't rush to switch your main model. Use 10 ARC-AGI-style abstract problems plus 10 of your own business problems for an A/B test to confirm whether 99.1% translates into gains on your tasks. If the gains are significant, evaluate routing high-value, low-frequency "hard tasks" (such as complex refactoring or protocol analysis) separately to Carbon while sending the rest to cheaper models—build a "model routing layer" rather than migrating everything. 3. **Gemini Decision 1: $0.10/M tokens input, free output** Gemini Decision 1 — https://aihot.news/items/iz9t2bjf03bbth4wd7s6m7qgj Essence: A single forward pass can handle up to 1,000 closed-ended questions, with a 256K context and free output. This strips "decision/scoring/routing" out of generative calls and turns it into a specialized, cheap primitive. What it changes: in the past, using large models for classification, labeling, and routing was both expensive and slow; now this kind of "closed-set decision" can be nearly free, and the cost structure of agent orchestration will be rewritten. —Possible: Extract all "if/else-style LLM calls" in your system (intent classification, content moderation labeling, ticket routing, A/B grouping), replace them in batches with Decision 1, and stress-test latency and accuracy. If results are good, further make it the "scheduling layer" in agent orchestration so that expensive models only handle the parts that truly require generation. 4. **Stripe acquires OpenRouter for about $7.5–8 billion** OpenRouter / Stripe — https://aihot.news/items/bx3sq4dxfqblyztkuiktiz0wv Essence: A model aggregation platform being acquired by a payments giant means that "token routing + enterprise usage management + billing" is becoming the toll gate for AI infrastructure. What it changes: model choice shifts from "developers manually switching" to "the platform automatically routing by cost/quality," and whoever controls payments controls routing—and thus controls bargaining power over model vendors. —Possible: If you build multi-model products, stop building your own routing table. Evaluate integrating OpenRouter and delegating billing/quota/degradation policies to it, while you focus on business logic and data moats. At the same time, watch out for the lock-in risk of "routing controlled by the platform" and keep a switchable direct-connection backup. 5. **Runway open-sources the 27B world-action model Praxis-1 (Apache 2.0)** Praxis-1 — https://aihot.news/items/bx3sq4dxfqblyztkuiktiz0wv Essence: A MoE diffusion Transformer predicts video and actions simultaneously. It ranks first on RoboLab-120 with a 46.2% success rate, is 15.6× faster than the GPT-6 Astra robotic arm, and releases open weights under Apache 2.0. What it changes: the "foundation model" for embodied intelligence shifts from a closed-source API to self-hostable weights, so robotics teams can fine-tune on their own data without vendor lock-in. —Possible: If you build robotics/simulation-related products, first spin up Praxis-1 locally and get inference working, then use 50–100 real robot trajectories on hand for LoRA fine-tuning to verify whether "open weights + your own data" can outperform calling an API directly. If you build education/research tools, building visual debugging and data-labeling tools around Praxis-1 is a low-competition entry point. 6. **Doubao Work integrates Kimi and DeepSeek; Baidu launches Token Exchange** Doubao / Baidu Token Exchange — https://www.bestblogs.dev/article/61ee281a2d Essence: On one side, an application-layer platform proactively integrates third-party models and autonomously selects models by task; on the other, Baidu builds a model trading platform that verifies qualifications before listing and detects "model-swap misrepresentation." What it changes: models shift from "binding relationships" to "tradable goods," and the core competitiveness of the application layer shifts from "which model to use" to "how to orchestrate and validate models." —Possible: If you build AI applications, abstract model calls into replaceable interfaces, use platforms like Token Exchange for price/performance comparison, and establish your own "model evaluation + canary switching" process. If you provide enterprise services, turn "model-swap misrepresentation detection" into a value-added capability that helps customers verify whether vendor claims match actual outputs. 7. **New York's RAISE Act opens registration, making 72-hour incident reporting a hard constraint** RAISE Act — https://aihot.news/items/usiap23d7lpoys4q2cl0z0y66 Essence: Frontier AI developers with annual revenue over $500 million must complete registration by 2027 and report safety incidents within 72 hours. In the same period, Australia plans to mandate that AI companies report safety violations and is considering revising criminal law. What it changes: compliance shifts from "voluntary commitments" to a combination of "registration + deadlines + criminal liability," and security incident response shifts from a PR issue to a legal issue. —Possible: Regardless of size, establish an "AI incident ledger" now (incident type, scope of impact, discovery time, response actions, reporting targets), and work backward from 72 hours to an internal 24-hour response SLA. If you build enterprise AI products, making "compliance logs + incident reporting templates" an out-of-the-box feature is a differentiating selling point for regulated industries. 8. **Cloudflare warns of a two-year AI vulnerability wave; GPT-6 Cyber has already found 81 new Firefox vulnerabilities** Cloudflare / GPT-6 Cyber — https://aihot.news/items/o3kzzqpf35qwhqm7w9xvr1 Essence: AI is mining hidden vulnerabilities in traditional software, and 6 critical vulnerabilities have already been publicly disclosed. OpenAI's GPT-6 Cyber scores 48.5% on ExploitGym, about a 51% improvement over the previous generation, and has discovered 81 new Firefox vulnerabilities. What it changes: for the first time, the speed of vulnerability discovery exceeds the speed of human remediation. The "two-year vulnerability wave" means security teams' scheduling logic must shift from "quarterly" to "daily." —Possible: If you maintain open-source or in-house systems, immediately run an AI-assisted round of fuzzing on dependencies and boundary inputs, prioritizing paths that are "remotely triggerable + high-privilege." If you build security products, turn "AI vulnerability discovery + automatic patch generation + regression validation" into a pipeline and sell it to small and medium-sized businesses without security teams. 9. **CMU benchmark: none of 9 AI agents passed expert review** CMU agent benchmark — https://aihot.news/items/thu24r8ruv2dlikm1hvikykaw Essence: None of 135 open-topic submissions simultaneously met honesty, completeness, and expert standards, and 114 failed integrity checks. What it changes: the shortcoming of agents on "open-ended research tasks" is not insufficient capability but a systemic lack of honesty and completeness—which directly determines whether agents can be placed into settings that require