The frontier price war collapsed into 90 minutes on Sept 22: Anthropic cut its flagship to $4/$20, OpenAI answered with half-priced GPT-6 Sol and Luna — and in the same week Akamai traded 5% of its stock for an $11.6B order while Oracle served force majeure on a 2.45 GW data centre. The bottleneck has moved from models to electricity
The thing worth remembering from this week is not any single benchmark, but that two clocks started at once. On September 22, Anthropic shipped Claude Opus 5.5, pushing flagship pricing down to $4/$20 per million tokens with cache reads at $0.20; per TechCrunch, OpenAI answered 90 minutes later with GPT-6 Sol and Luna at half the price of the 5.6 series. On the model side, cost is collapsing by the quarter. On the other side it is hardening: Akamai's September 24 release traded a warrant for up to 5% of its common stock against an $11.6B, seven-year Anthropic CPU commitment; the same day Oracle served force majeure on the 2.45 GW Project Jupiter in New Mexico, where air permits and power are the blockage. For medicine, each clock carries its own invoice: cheaper inference spreads ambient scribes faster, and Blue Cross's September 24 study has already booked $942M of added inpatient cost against them.
01 — Top Stories
Anthropic cuts flagship pricing to $4/$20: Opus 5.5 claims Fable 5.1-class work at 40% lower cost and 30% faster output
On September 22 Anthropic released Claude Opus 5.5, calling it "the first model in our new Claude 5.5 family." 9to5Mac's write-up lists the pricing: $4/$20 per million input/output tokens (down 20%), cache reads at $0.20 (down 60%), roughly 40% lower total cost than Opus 5 on typical workloads, and 30% faster output. Anthropic cites a tester who "completed a 680,000-line code migration in less than a day" and a 39-of-40 success rate on web optimisation tasks. Sonnet 5.5 and Haiku 5.5 follow "in the coming weeks."
Frontier inference is shifting from a consultant billed per engagement to a utility billed per unit. That is the first real opening for hospital IT to put an LLM into batch workflows — re-running a whole department's discharge summaries overnight, restructuring three years of imaging reports — rather than confining it to one clinician's chat box. The 60% cut on cache reads matters most, because the longest part of a clinical prompt (local guidelines, formulary, the patient's prior history) is exactly the cacheable part.
The 680,000-line migration, the 39-of-40 rate and the 30% speed-up are all vendor-reported, unaudited, and none of them is a clinical task. The figures here come from 9to5Mac's secondary coverage rather than the primary release page.
Ninety minutes later: OpenAI answers with half-priced GPT-6 Sol and Luna, claiming Sol "makes about half as many mistakes" as its predecessor
TechCrunch reports that 90 minutes after Opus 5.5, OpenAI shipped the GPT-6 generation's Sol (complex work and coding) and Luna (high-volume clerical work: document summarisation, extraction, quick Q&A). Both are priced at half the 5.6 series, which OpenAI attributes to caching and inference optimisation. The company says "GPT-6 Sol makes about half as many mistakes as its predecessor, reaching Astra-level reliability at much lower cost," and claims to beat Anthropic's Fable and Opus. Sol lands in ChatGPT Work, Codex and the API; Luna also reaches Free and Go users.
Luna is the line hospital administrators should watch. OpenAI states plainly that it is built for "high-volume clerical work," which is precisely where hospital headcount goes: prior-authorisation forms, referral letters, appeals, chart summaries. Once a model purpose-built for that sits in the free tier, the adoption decision stops being "do we buy it" and becomes "do we govern it."
"Half as many mistakes" and "Astra-level reliability" are, per TechCrunch, OpenAI's own internal evaluations, not independent benchmarks, and "beats Fable and Opus" is a vendor self-comparison. Both labs claimed to beat the other on the same day; neither claim should be taken at face value.
Akamai trades a warrant for up to 5% of its stock for Anthropic's $11.6B, seven-year CPU commitment, with options taking it near $20B
Akamai's September 24 release spells it out: Anthropic commits $11.6B over seven years, with expansion options adding up to $9B more (about $20B in all). Akamai issued Anthropic a warrant for up to 5% of its common stock (7.7 million shares at $111.33), roughly 2% vesting immediately against the $11.6B and the remaining 3% vesting in $3B increments. The workloads are CPU, not GPU. Akamai puts capex for the committed portion at about $5.5B, with a further $1.7B of 2026 capex for supply-chain components including memory. Bloomberg and TechCrunch covered it in parallel.
The counter-intuitive part is the workload: CPU. With everyone's attention on GPUs, a frontier lab is locking down distributed, edge-adjacent general-purpose compute — which usually means the periphery of agentic workflows (tool calls, retrieval, sandboxed execution, data preprocessing) has grown large enough to procure on its own. For regulated health data, "distributed and close to the user" is exactly the topology data residency wants; there is, however, no evidence this capacity is earmarked for healthcare customers. Paying for compute in equity is also a reminder that AI supply-chain financing increasingly looks like a circle of cross-holdings.
Oracle serves force majeure on the 2.45 GW Project Jupiter: what is holding AI back is not chips but air permits and the grid
Per the Albuquerque Journal on September 24, Oracle sent a force majeure notice to Stack Infrastructure, Project Jupiter's developer and a Blue Owl Capital portfolio company. The site, in Doña Ana County, New Mexico, carries 2.45 gigawatts of power plant capacity; more than 900 workers were hired in August alone and over 3,600 are on site. The New Mexico Supreme Court only recently lifted stays blocking the air quality permit application and construction water pumping. Oracle spokesperson Michael Egbert said "force-majeure notices are commonplace in developments of this scale" and that the project "remains on our planned schedule." The Wall Street Journal's framing is that the power and permitting hurdles "expose Oracle's AI risks."
This is where the week's two lines cross: a 40% price cut on the model side in one quarter, against 2.45 gigawatts held up by a single air permit. Any health system writing "AI costs will keep falling" into a five-year budget should note that the back half of that curve is set by substations, water rights and county politics — none of which obey Moore's law. A force majeure notice is a contractual protection, not a construction failure, but it is a public stress signal.
The Wall Street Journal original is paywalled; this item is written from the Techmeme headline and lede plus the Albuquerque Journal's open coverage. Neither the exact date of the notice nor the sums involved were disclosed.
Blue Cross books $942M of added inpatient cost to AI scribes: "If patients are truly sicker, we'd expect to see more treatment"
Per PYMNTS on September 24, the Blue Cross Blue Shield Association — 31 independent Blues plans covering over 100 million people — compared 2024–2025 inpatient billing against 2023 and found nearly $1B in added cost over two years: $653M from increased secondary-condition billing and $942M from higher overall care intensity. The study points at ambient scribe systems that listen to the encounter, draft the note and surface secondary conditions. BCBSA senior vice president Luke Chalker said: "If patients are truly sicker, we'd expect to see more treatment." Hospitals counter that patients genuinely are sicker. Tech Startups covered the same study.
This is the one story of the week that converts the abstract trend of cheaper general-purpose AI into a number — and the number points at more spending, not less. Ambient scribes have always been sold on reducing clinician documentation burden, but they also capture secondary diagnoses that used to go unrecorded. That may be clinically correct and still functions, in payment terms, as a price increase. Payers and providers reading the same note differently will be the most expensive argument in medical AI over the next two years.
The study comes from a payer with an obvious interest; correlation between rising bills and AI adoption is not causation, and the providers' "patients really are sicker" has not been independently tested. The $653M and $942M figures are measured differently and should not be summed. This item is written from secondary coverage; the full BCBSA study was not obtained.
Life sciences behind a whitelist: Anthropic's Mythos 5.1 goes only to vetted US cyber and life-sciences organisations, while false refusals on benign medical questions fall 85%
On Anthropic's Fable 5.1 / Mythos 5.1 page, one underlying model is split in two: Fable 5.1 is generally available, while Mythos 5.1 reaches only vetted professionals in cybersecurity and life sciences through trusted access programs. On the biology side, false positives on benign medical and elementary biology questions are down 85%; advanced life-sciences work requires the new Life Sciences Verification Program, run with the US government, initially limited to US organisations with international expansion coordinated with Washington. Cyber false positives fall about 60%, and the model may identify vulnerabilities but not generate exploits. On capability, Anthropic claims Fable 5.1 moves agentic scientific research from 24.7% to 52.6%, and that protein design hit 10× the binding affinity of the best competition entries on three targets with a roughly 50% hit rate across twelve. The system card, dated September 1, has the detail.
Two things happened at once. On capability, AI can now design binders an order of magnitude stronger than the best human competition entries. On governance, that capability has been wired into a list co-administered by the US government, with nationality as the first filter. For non-US research institutes and biotechs, the barrier to frontier biological capability is no longer only budget or ethics review — it is geopolitics. The 85% drop in false positives is the good half of the story: clinical and teaching use was routinely caught by safety machinery, and this version is visibly fixing that.
The protein-design and agentic-research figures come from Anthropic's own evaluations and system card, without peer review or independent wet-lab replication. The release dates to early September, outside this week's window; it is included because the life-sciences gate resurfaced in this week's cyber coverage.
Three labs simultaneously lock their strongest cyber models behind whitelists — and Google names healthcare among its first-priority "critical infrastructure"
The Hacker News's roundup records three parallel moves in September. Google shipped Gemini 3.8 Flash Cyber through the limited-access Fairwind Program (launched September 2), aimed at governments and national cyber authorities, critical-infrastructure operators (healthcare is named, alongside telecoms, energy and finance) and core technology platforms, with over 650 partners including CrowdStrike, Palo Alto, Snowflake and Wiz; paired with CodeMender, it generates "verified, deployment-ready patches in minutes." OpenAI's Astra was judged to meet the Critical cybersecurity capability threshold in its Preparedness Framework and now reaches selected testers via the Daybreak Blue program, with claims of 100% on ExploitBench and declining 91.5% of jailbreak requests. Anthropic, as above, put vulnerability identification into Fable 5.1 and kept the least-restricted version for Mythos 5.1.
Health systems are simultaneously the critical infrastructure most often breached by ransomware and the least able to staff a red team. Putting healthcare first in Fairwind is an acknowledgement of that gap — but "limited access" also means most small and mid-sized hospitals and clinics stay outside the list, while attackers need no whitelist at all. If CodeMender's "deployment-ready patches in minutes" holds up, the shock for an industry that schedules a three-month maintenance window to patch its HIS is procedural, not technical.
The perfect ExploitBench score, the 91.5% jailbreak refusal rate and "patches in minutes" are all vendor-reported, and the Fairwind page itself gives no numerical benchmarks. This item rests mainly on The Hacker News's roundup plus Google's own blog; the OpenAI and Anthropic primary posts were not each retrieved.
Gemini 3.8 Live's Live Avatar goes GA: a real-time video persona lip-syncing across 97 languages, every frame SynthID-watermarked
Google announced on September 24 that Live Avatar in Gemini 3.8 Live is generally available: near-real-time video generation fused with speech to produce a persona that listens, sees and speaks, with precise lip-sync and natural expressions, speech-to-speech synchronisation across 97 languages, asynchronous background tool execution while dialogue continues, and custom avatars generated from reference images for brand work. Everything is SynthID-watermarked. It is available in Gemini Enterprise only, and custom avatar creation requires allowlisting. The demonstrated use cases are customer service, interactive walkthroughs and hotel check-in — no healthcare applications are mentioned. The Google Cloud blog covers the enterprise side.
Real-time speech-to-speech across 97 languages lands directly on two of the hardest things to staff in a hospital: overnight interpretation and patient education. That Google pointedly avoids clinical use cases is itself information — a talking face in a clinic setting is hard for a patient not to read as a source of clinical opinion. Blanket SynthID watermarking is the most practical governance step available for synthetic faces today; read the other way, it is an admission that the technology needs labelling.
Google says only "near real-time" and publishes no latency figure; pricing is undisclosed. There is no third-party evaluation of lip-sync quality across the 97 languages.
Grok 4.7 lands a day ahead of the others: a same-price upgrade with better coding, but judged "late to the frontier party"
On September 21 xAI released Grok 4.7, logged by LLM-Stats as a "same-price upgrade" with improved benchmarks; Evolink's summary and SQ Magazine both position it as xAI's most capable coding model to date, priced identically to 4.6. Decrypt's verdict is blunter: bigger, but late to the frontier party.
Read the three launches in three days together — Grok 4.7, Opus 5.5, GPT-6 Sol/Luna — and the signal is clear: the marginal gap within a capability generation is narrowing, and the axis of competition is moving from "who is strongest" to "who is cheapest and least gated." That makes it a good year for healthcare procurement, and it also means selection frameworks built around benchmark scores are close to useless.
This item rests entirely on secondary tech coverage; xAI's own release page and system card were not retrieved, so specific benchmark numbers are omitted.
02 — Product Analysis
Claude Opus 5.5
Frontier inference repriced as a utility · Anthropic (US)
Function and position. Anthropic describes it as performing at Fable 5.1's level while costing substantially less, aimed at long-running multi-step agentic work and complex code. Pricing is $4/$20 with $0.20 cache reads, about 40% below Opus 5 on typical workloads (9to5Mac). The buyer is an engineering team running an LLM as a batch backend, not a chat user.
- Strength : cache reads down 60% to $0.20, which pays off exactly where a clinical prompt reuses a large fixed prefix — local guidelines, formulary, prior history. Pricing detail here.
- Strength : the same family's Fable 5.1 moves agentic scientific research from 24.7% to 52.6% and posts concrete protein-design results (Anthropic), suggesting this generation's scientific reasoning is not just marketing.
- Concern : every claim is vendor-reported, and none of them is a clinical benchmark. What a hospital wants to know — omission rate in discharge summaries, sensitivity on drug-interaction detection — appears nowhere in the launch material.
- Concern : the cut was driven by a rival's same-day cut, and its durability is unknown. Budgeting on "it will be cheaper again next year" is risky while upstream power costs are moving the other way.
GPT-6 Luna
A cheap model built for "high-volume clerical work" · OpenAI (US)
Function and position. OpenAI defines Luna explicitly as a high-volume clerical model for document summarisation, extraction and quick Q&A, priced at half the 5.6 series and available to Free and Go users (TechCrunch). It does not exist to win benchmarks; it exists so that work too voluminous to justify a flagship has somewhere to go.
- Strength : the shape of healthcare administrative work matches its design — prior auth, referrals, appeals, chart summaries. Halving the price turns batch processing from a pilot into a budgetable norm (source).
- Concern : reaching the free tier means reaching ungoverned shadow IT. A clinician summarising a chart in Luna passes through no institutional review, and BCBSA's $942M has already shown what ungoverned documentation automation bills out at.
- Concern : OpenAI publishes no hallucination rate or extraction accuracy for Luna — only Sol's "half as many mistakes." Cheap models run precisely at the volume least likely to be human-reviewed.
Gemini 3.8 Live — Live Avatar
Real-time video persona across 97 languages · Google (US)
Function and position. Near-real-time video plus speech generation produces a persona that listens, sees and speaks, with lip-sync, speech-to-speech across 97 languages, background tool execution mid-dialogue and brand-custom avatars from reference images — all SynthID-watermarked, and available in Gemini Enterprise only (Google).
- Strength : speech-to-speech across 97 languages maps onto the hardest and most expensive interpretation shifts a hospital has to fill, and blanket SynthID watermarking at least makes the synthetic face auditable (source).
- Concern : Google's demonstrations are customer service and hotel check-in, with no mention of healthcare — a deliberate regulatory blank. Where a talking face delivering patient education in a waiting room ends and "giving medical advice" begins is written down in no regulation today.
- Concern : no latency figure and no pricing are published. Clinical dialogue is acutely latency-sensitive, and "near real-time" is not enough to evaluate against.
03 — Companies & Competition
| Company | Recent state & numbers | Position & moat |
|---|---|---|
| Anthropic Frontier lab with tiered access |
Shipped Opus 5.5 on Sept 22 at $4/$20 with a 40% cost cut (9to5Mac); signed a seven-year $11.6B CPU contract with Akamai on Sept 24 in exchange for a warrant on up to 5% of its stock (Akamai); Mythos 5.1 and the Life Sciences Verification Program are US-only (Anthropic). | The moat is tiered access co-administered with government: keep the strongest capability inside a whitelist and trade it for regulatory trust. The weakness is that outside the US the same moat is friction — non-US health and biotech customers are structurally excluded from the top tier. |
| OpenAI Covering the whole demand curve on price |
Shipped GPT-6 Sol and Luna on Sept 22 at half the 5.6 series' price, with Luna pushed down to Free and Go (TechCrunch); Astra was judged to meet the Critical cyber threshold and moved into the limited Daybreak Blue program (The Hacker News). | The moat is distribution: putting Luna in the free tier pre-installs ungoverned documentation automation on every hospital employee's desktop. The weakness is that this path bypasses procurement and security review, so when something goes wrong the liability lands on the hospital, not OpenAI. |
| Google / Alphabet Allocating capability by programme |
Live Avatar went GA on Sept 24 with 97 languages, SynthID and Gemini Enterprise gating (Google); Fairwind launched Sept 2 with over 650 partners, naming healthcare a priority critical-infrastructure recipient (Google). | The moat is an existing enterprise and security channel — 650 partners is distribution, not model capability. The weakness is that any allocation requiring application, vetting and allowlisting is always a step behind an attacker, and small and mid-sized providers usually never get in. |
| Akamai CDN pivoting into distributed AI cloud |
Announced Anthropic's seven-year $11.6B commitment on Sept 24 with up to $9B of expansion options; issued a warrant for 7.7 million shares at $111.33, roughly 2% vesting immediately; about $5.5B of capex against the committed portion plus $1.7B more in 2026 for components (Akamai). | The moat is a distributed, edge-adjacent node topology — exactly the shape that data residency and low latency both want, and in theory favourable for regulated health data. The weakness is customer concentration in a single contract, plus the valuation interdependence created by paying for compute in equity. |
| Oracle The bet on hyperscale self-build |
Served force majeure on the 2.45 GW Project Jupiter to developer Stack Infrastructure on Sept 24; more than 3,600 workers are on site, and the New Mexico Supreme Court only recently lifted stays on the air permit and construction water pumping (Albuquerque Journal). | The moat was meant to be scale: build 2.45 GW at once and nobody catches up. The weakness is that the completion date sits with a county government, a water-rights court and the grid — while Akamai's route of assembling compute from existing nodes never faces that gate at all. |
| xAI Holding position with same-price upgrades |
Released Grok 4.7 on Sept 21 at parity pricing with 4.6 and improved benchmarks, positioned as its strongest coding model (Evolink); the outside verdict was "late" (Decrypt). | Barely present in healthcare, with no corresponding biological or clinical safety programme. In a market where the axis of competition has moved to access governance, having no governance story is itself having no moat. |
In one sentence: this week's competitive structure is a scissors — price going down, thresholds going up. Two labs each halved model pricing within three days, while the strongest cyber and life-sciences capability was locked behind three separate whitelists, and the physical compute underwriting all of it hangs on a single air permit. For healthcare buyers the lesson is not whose score is higher: the cheap tier is available to you at any time, the expensive tier may not have you on its list at all, and the supply of both depends on a substation you have no seat at the table for.
04 — Taiwan Angle
(1) Oracle is stuck at 2.45 GW; Taiwan is stuck at 1 GW. On September 1, former digital development minister Huang Yan-nan said Taiwan's sovereign AI faces a double shortage of compute and electricity: South Korea has bought 260,000 GPUs from NVIDIA and Japan has over 40,000, against Taiwan's roughly 30,000 — and "30,000 GPUs require over 1 GW of electricity, equivalent to one nuclear power plant's output." The same piece notes Taiwan ranks 16th on the Global AI Index, behind Singapore, South Korea and Japan. Set the Oracle story beside it and the point sharpens: if America's largest software company can be held up on permits for 2.45 GW, the 1 GW Taiwan needs for medical AI will not appear simply because the budget did. The practical consequence for hospitals is that building an in-house inference cluster stays uneconomic for the foreseeable future; hybrid cloud and API procurement will dominate, which puts data governance in contracts and audits rather than in the server room.
(2) The immediate beneficiary of halved pricing is documentation automation — and Taiwan's governance paperwork is already written. GPT-6 Luna targets high-volume clerical work explicitly and reaches the free tier, while Opus 5.5 cuts cache reads to $0.20. Together, these will materially lower the cost of deploying ambient scribes and automated chart summarisation in Taiwan within a year. The Ministry of Health and Welfare has already issued its Guidelines on the Use of Generative AI in Medical Institutions, a governance framework of six risk categories and nine key points, and stood up three national AI centres. What deserves attention is that BCBSA's $942M study points not at hallucination but at payment inflation caused by AI capturing secondary diagnoses that used to go unrecorded. That mechanism is more sensitive under Taiwan's National Health Insurance, because in a single-payer global-budget structure a systematic rise in coding intensity presses directly on the cap. Taiwan's guidelines currently focus on clinical safety and patient rights and do not yet address what AI-assisted coding does to reimbursement — a gap that can be prepared for in advance.
(3) The item that deserves the most attention is that Taiwan is not on the list. Anthropic's Life Sciences Verification Program and Mythos 5.1 are initially limited to US organisations, with international expansion coordinated with the US government; Google's Fairwind names healthcare a priority but is equally limited and vetted. Taiwan holds a rare combination in Asia — comprehensive single-payer health data and a dense biotech cluster — yet sits structurally outside the first round of allocation for frontier biological design and cyber defence capability. For policymakers this is not a matter of waiting for the door to open: South Korea and Japan have made themselves unignorable counterparties through sovereign compute procurement, and Taiwan's roughly 30,000 GPUs carry no leverage in an allocation logic denominated in nation-states. For providers, the pragmatic near-term move is to govern well the tier they can actually reach — generally available Fable 5.1, Opus 5.5, GPT-6 — rather than planning around the tier they cannot.
05 — Further Reading
-
Introducing Claude Fable 5.1 and Claude Mythos 5.1 — Anthropic (2026-09)
The primary document most worth reading end to end this week. The reasoning behind splitting one model into two access tiers, the exact magnitude of the biology and cyber safeguard adjustments, and the terms of the government-coordinated verification programme are all on this one page.
-
Akamai Announces $11.6 Billion Multi-year Agreement with Anthropic — Akamai Newsroom (2026-09-24)
Read it for how the warrant vesting is written: 2% immediately, the rest in $3B increments. This structure — binding procurement to equity — is spreading across the AI supply chain, and it is worth remembering the first time it was spelled out plainly.
-
AI-Generated Medical Coding Adds Nearly $1 Billion to Blue Cross Costs — PYMNTS (2026-09-24)
One of the few analyses to connect AI adoption directly to actual payment amounts. Even allowing for its payer origin and interested methodology, the question it models — are patients sicker, or merely better recorded — is one every hospital deploying an ambient scribe should ask itself once.
-
Fairwind Program: Cyber defense tools for trusted partners — Google Blog (2026-09-02)
Read the eligibility terms rather than the product description: participating organisations must confine access to internal security, incident response or penetration testing teams and deploy protections such as multi-factor authentication. Those conditions effectively define which size of provider qualifies — and the answer is: not many.
-
拚主權 AI 台灣算力、電力雙缺 — 聯合新聞網 (2026-09-01)
The piece that pulls Taiwan's AI conversation back from "do we have a model" to "do we have the power." The conversion of 30,000 GPUs into 1 GW is the constraint that belongs on page one of any five-year medical AI plan.
06 — References
- Anthropic upgrades Claude with new Opus 5.5 model, details here. 9to5Mac, 2026-09-22. 9to5mac.com
- Claude Opus 5.5 launch. LLM-Stats, 2026-09-22. llm-stats.com
- OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes. TechCrunch, 2026-09-22. techcrunch.com
- OpenAI's New GPT-6 Sol and Luna Models Bring Astra Improvements to Cheaper Tiers. MacRumors, 2026-09-22. macrumors.com
- Akamai Announces $11.6 Billion Multi-year Agreement with Anthropic to Support Growing Demand. Akamai Newsroom, 2026-09-24. akamai.com
- Anthropic to pay Akamai $11.6 billion over seven years in cloud deal. TechCrunch, 2026-09-25. techcrunch.com
- Anthropic Strikes $12 Billion Deal With Akamai for AI Computing. Bloomberg, 2026-09-24（付費牆）. bloomberg.com
- Oracle invokes 'force majeure' on Project Jupiter, but says data center still on track. Albuquerque Journal, 2026-09-24. abqjournal.com
- Project Jupiter faces power and permitting hurdles, exposing Oracle's AI risks. Techmeme / Wall Street Journal, 2026-09-25（WSJ 原文付費牆）. techmeme.com
- AI-Generated Medical Coding Adds Nearly $1 Billion to Blue Cross Costs. PYMNTS, 2026-09-24. pymnts.com
- AI tools are driving up health insurance costs by nearly $1 billion, Blue Cross study finds. Tech Startups, 2026-09-24. techstartups.com
- Introducing Claude Fable 5.1 and Claude Mythos 5.1. Anthropic, 2026-09. anthropic.com
- System Card: Claude Fable 5.1 & Claude Mythos 5.1. Anthropic, 2026-09-01. www-cdn.anthropic.com
- Google, Anthropic, and OpenAI Unveil Cyber AI Models, Safeguards, and Access Programs. The Hacker News, 2026-09. thehackernews.com
- Fairwind Program: Cyber defense tools for trusted partners. Google Blog, 2026-09-02. blog.google
- Fairwind Program. Google DeepMind, 2026-09. deepmind.google
- Introducing Gemini 3.8 Live with Live Avatar. Google Blog, 2026-09-24. blog.google
- Gemini 3.8 Live with Live Avatar is now generally available. Google Cloud Blog, 2026-09-24. cloud.google.com
- Google Brings Live Avatar Visual Presence to Gemini 3.8 Live. Unite.AI, 2026-09. unite.ai
- xAI Launches Grok 4.7. It's Bigger, But Late to the AI Frontier Party. Decrypt, 2026-09-21. decrypt.co
- Grok 4.7 Release Date: Out September 21, 2026, What's New. Evolink, 2026-09-21. evolink.ai
- xAI Launches Grok 4.7, Its Most Capable Coding Model Yet. SQ Magazine, 2026-09. sqmagazine.co.uk
- LLM News Today — AI Model Releases. LLM-Stats, 2026-09. llm-stats.com
- AI News Today, September 24: Top Stories. AI Weekly, 2026-09-24. aiweekly.co
- 拚主權 AI 台灣算力、電力雙缺. 聯合新聞網, 2026-09-01. udn.com
- 高醫大論壇揭示 AI 醫療新局！衛福部推「333 政策」國家 489 億預算力挺. 聯合新聞網, 2026-06-27. udn.com
- 衛福部頒布「醫療機構應用生成式人工智慧指引」. 理律法律事務所. leeandli.com
- 臺灣智慧醫療三大中心. 衛生福利部. aicenter.mohw.gov.tw