> ## Content Index > Fetch the complete content index at: https://aitrading67.com/llms.txt > Use this file to discover other available public pages before exploring further. # AI & Work — Weekly Observatory · 18–25 September 2026 - URL: https://aitrading67.com/en/osservatorio-2026-09-25/ - Published: 2026-09-25T08:00:00.000Z - Updated: 2026-09-25T07:43:48.000Z - Description: The week AI's builders asked the UN Security Council for rules and, in the same days, cut prices with Claude Opus 5.5 and GPT-6: a price war pushing agents into the office, even as Bengio calls it a race where everyone loses. - Author: Fabio Gentili - Tags: lang-en, observatory, AI & Work World Observatory · AI & Work AI & Work — Weekly Observatory · 18–25 September 2026 September 25, 2026 — Fabio Gentili observatoryAI & Work Editorial Some weeks are best read as the gap between what people say and what they do. This was one of them. On 23 September the leaders of the frontier — Sam Altman of OpenAI, Dario Amodei of Anthropic, and the Turing laureate Yoshua Bengio — sat before the UN Security Council to ask for caution and shared rules; Bengio was blunt that the dangers are “real and imminent” and the contest “a race where everyone loses.” Yet just a day earlier the same labs had done the opposite: on 22 September Anthropic shipped Claude Opus 5.5 and, within hours, OpenAI answered with two cheaper GPT-6 models. Fortune caught the paradox in a headline: “What AI slowdown?” The week’s clearest signal is this split between the rhetoric of restraint and the logic of the market — and price is the true engine of displacement. When OpenAI puts “Luna” at $0.10 and $0.50 per million tokens for “high-volume routine work,” it makes automating the first rung of many careers economically trivial. And the frontier is not only getting cheaper: in the same week Anthropic reported that about 950 Claude agents had autonomously surfaced a novel, CRISPR-like enzyme system in viral DNA — a reminder that capability, not just price, keeps moving even as its builders ask the world to slow down. Everything else turns on that axis: agents become low-cost commodities while unions try to write a human-control clause into contracts, governments argue over who sets the pace, and Europe keeps deferring the AI Act’s hardest obligations. The frontier speaks the language of safety; the price list speaks the language of acceleration. --- In this issue 1. [AI evolution — key developments](#sez-en-1) 2. [AI & work — displacement and transformation](#sez-en-2) 3. [Automation and reskilling](#sez-en-3) 4. [New professions and opportunities](#sez-en-4) 5. [Ethics and ethical issues](#sez-en-5) 6. [AI risks](#sez-en-6) 7. [Leading voices](#sez-en-7) --- Weekly thematic blocks 🤖 01 · AI AI evolution — key developments Medium tension Frontier labs cut prices — and Claude makes a DNA discovery **Claude discovers a novel enzyme system (Anthropic, 23 Sept):** a landmark for AI-led science. About 950 Claude agents searched a DNA database for 21 hours (\~210 million tokens), sifting 200,000+ reverse transcriptases down to 20 candidates until one agent flagged a CRISPR-like repeat array in bacteriophage DNA — a new “array-associated reverse transcriptase” (ART) system. Its function is still unknown. MIT/Broad’s Feng Zhang called it “an exciting example of how AI agents can contribute to biological discovery”; skeptics such as WashU’s Kevin Blake caution there is “nothing to indicate this is a rival to CRISPR-the-technology.” **Claude Opus 5.5 (Anthropic, 22 Sept):** the first model since Amodei’s “pace the frontier” call — an efficiency step, not a capability leap. 20% cheaper than Opus 5 ($4/$20 per million input/output tokens), up to 40% less on typical workloads, 30% faster output; vendor benchmarks: 66.4% Terminal-Bench 4.0, 40% AutomationBench (vs 26.9%), 67.7% Humanity’s Last Exam with tools; a claimed 85% cut in safety-boundary circumvention attempts. **GitHub, on Opus 5.5:** chief product officer Mario Rodriguez, per SiliconANGLE, said it “solved more terminal tasks than Opus 5 in less than half the steps.” **GPT-6 Sol & Luna (OpenAI, 22 Sept):** Sol at $2/$10 for coding and agent tasks; Luna at $0.10/$0.50 for high-volume routine work, with a 90% discount on cached input — 50% cheaper than the GPT-5.6 predecessors. **Qwen3.8-Omni-Flash (Alibaba, 18 Sept):** natively multimodal (text, image, audio, video in), 1-million-token window, $0.15/$0.47 per million, a claimed +25% average over its predecessor across 29 evaluations (vendor-reported). **Grok 4.7 (xAI, 21 Sept):** extended reinforcement learning on long-horizon tasks, 500,000-token window; scores vendor-reported, not yet independently reproduced. **The tell:** competition is shifting from raw power to the economics of inference. Ramp economist Ara Kharazian: “OpenAI and Anthropic are engaged in a price war that is driving down the price of AI.” 💼 02 · WORK AI & work — displacement and transformation High tension Displacement arrives as a falling line on a price list **Price is the mechanism:** OpenAI’s Luna at $0.10/$0.50 per million tokens, recommended for “summarization, extraction” and routine work, makes automating first-rung cognitive tasks economically trivial. Displacement no longer arrives in a press release — it arrives as a price that falls. **Agents mature (Salesforce, 11 Sept, antecedent, live this week):** its “Fin” agent autonomously resolves 79% of Anthropic’s conversations; “Eva” closes 50% of Engine’s chat inquiries; “Paige” handles 70% of Autism Queensland’s admin requests — hollowing out the repetitive load. **The macro paradox holds:** Yale’s Budget Lab (15 Sept, antecedent) still finds no clear AI footprint in aggregate employment; Oracle’s 21,000 and Accenture’s 22,000 cuts belong to earlier periods and are context, not this-week news. **Where it’s prepared:** in the margin — one task at a time, not one department at a time. When the aggregate data speaks, it will describe a decision made months earlier, the day an agent became cheaper than an entry-level salary. 🎓 03 · SKILLS Automation and reskilling Medium tension Unions write the human-control clause; reskilling turns to policy **Collective bargaining as the tool:** the NewsGuild–McClatchy deal bars AI replacing staff or cutting pay/hours, prohibits deepfakes and requires human review; Oregon AFSCME excludes AI monitoring, hiring and evaluation and mandates 90 days’ notice; Kaiser Permanente won a joint technology task force. **Live this week:** at OHSU the 90-day bargaining period over AI implementation is under way, with a 14 October deadline. The thread is “human-in-the-loop”: oversight as a clause, advance notice as a right. **The bottleneck flipped:** with cheap inference, the scarce resource is human skill. PwC’s 2026 Global AI Jobs Barometer records +68.9% year-on-year hiring for specialist AI roles and a \~56% wage premium. **Reskilling as industrial policy:** the CA$300M for Bengio’s LawZero from Canada and Germany (16 Sept, antecedent) shows states creating public demand for safety skills. 🚀 04 · ROLES New professions and opportunities Low tension Cheaper agents multiply the orchestration and oversight trades **Orchestration spreads:** every agent sold needs a technician to design, integrate, train and supervise it. After the 22 Sept price cuts, more firms can afford agents — lifting demand for the “agent orchestrator,” the agentic-platform administrator and the forward-deployed engineer. **Oversight by regulatory necessity:** the UN’s call for evaluations, independent audits and shared standards (23 Sept) creates a market for model evaluators, algorithmic auditors and alignment specialists; LawZero’s funding gives it early infrastructure. **Two distinct paths (PwC 2026 Barometer):** rewarded skills — a \~56% wage premium and +68.9% specialist hiring — on one side; eroded skills on the other. The doors that open are high. **Human skills rise:** judgment, relationship and accountability grow more valuable even as automation advances, PwC finds — the counterweight to the vanishing lowest rungs. ⚖️ 05 · ETHICS Ethics and ethical issues High tension Who sets the pace? The power question reaches the Security Council **UN Security Council (23 Sept):** the builders sat before the states on the governance of risk. Altman: “No one person or company or country should be able to use the most powerful AI models to impose their worldview on everyone else.” **The tension:** the plea for common standards coexists with a price war pushing adoption everywhere, sooner and faster than any rule. **Europe vs the US:** the AI Act’s “Digital Omnibus” (spring, antecedent) defers high-risk obligations to December 2027, criticized by rights groups as a rollback; California’s Newsom has until 30 September to act on the final AI bills, after enacting SB 1050 (16 Sept) on disclosure of AI-generated performers. **The unresolved stake:** is declared transparency enough, or is external, independent verification required? No venue has yet answered. ⚠️ 06 · RISKS AI risks High tension “A race where everyone loses”: loss of control meets everyday cyber-risk **Bengio to the UNSC (23 Sept):** the dangers are “real and imminent”; the contest is “a race where everyone loses.” Loss of control is a challenge “that none can contain alone and that does not respect the borders we defend.” **Amodei:** “If managed poorly, I even believe that AI could be a risk to humanity as a whole,” pledging Anthropic will slow “as much as necessary.” **Already here:** autonomous agents lower the barrier to cyber-attack, letting under-resourced operators run campaigns that once required whole teams — the near-term risk Bengio flagged, of agents breaking through security barriers and penetrating companies. **Price cuts cut here too:** making AI cheaper makes it cheaper for those who use it to do harm. The distant catastrophe divides experts; the everyday attack does not. 🎙️ 07 · VOICES Leading voices Medium tension Altman’s caution · Amodei’s alarm · Bengio’s “everyone loses” **Sam Altman (OpenAI):** “As the process of building AI becomes more automated, the pace of AI progress could accelerate rapidly. This moment calls for extreme care.” And: “Being companies in a competitive race is not a reason to make rash decisions.” **Dario Amodei (Anthropic):** “If managed poorly, I even believe that AI could be a risk to humanity as a whole”; Anthropic will slow “as much as necessary… \[so\] every successive AI technology that we release is actually safe”; “No leader, no company and no nation can manage this alone.” **Yoshua Bengio (LawZero, Turing laureate):** “It is also a race where everyone loses”; “no one, no company and no country, should develop an AI system that could cause large-scale harm.” **Also present (reported):** Clem Delangue, co-founder of Hugging Face, representing the open-source camp. **Geoffrey Hinton (background):** the “godfather of AI” did not speak this week, but his 16 Sept Senate warning — “maybe a year” to regulate — kept framing the debate; his thesis that “AI is designing better AI” is exactly Altman’s automatic-acceleration fear. --- Focus / In-depth analysis Focus The week the builders preached caution — and cut prices A coordinated safety plea reaches the UN Security Council while, a day earlier, the same labs open an all-out price war that pushes agents deeper into work. This was a week of contradiction. At the UN Security Council on 23 September, Altman, Amodei and Bengio asked the world’s states for restraint and shared standards. A day earlier, Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Sol/Luna had cut prices sharply, extending a price war that makes putting AI to work cheaper by the week. The throughline to labor is direct: the cheaper the agent, the lower the marginal cost of automating an office task. Claude/Anthropic: \~950 agents, 21h → novel CRISPR-like “ART” enzyme system in viral DNA, function unknown (23 Sept) · Claude Opus 5.5: $4/$20, 20–40% cheaper, 30% faster (22 Sept) · GPT-6 Sol $2/$10, Luna $0.10/$0.50, 50% cheaper (22 Sept) · Qwen3.8-Omni-Flash 1M context (18 Sept) · Grok 4.7 500k context (21 Sept) · UN Security Council: Altman, Amodei, Bengio (23 Sept) · PwC 2026 Barometer: +68.9% AI-specialist hiring, \~56% wage premium · Salesforce agents: Fin 79%, Eva 50%, Paige 70% (11 Sept, antecedent) It is also a race where everyone loses. — Yoshua Bengio, UN Security Council, 23 September 2026 Watchpoints for the next seven days: (1) how far the price war runs and whether a frontier-capability release follows the efficiency ones; (2) independent replication of Opus 5.5, GPT-6, Qwen and Grok benchmarks; (3) which California AI bills Newsom signs by 30 September; (4) any follow-through from the UN session toward real standards; (5) whether agent deployments start to register in aggregate labor data; (6) new union AI clauses as the OHSU bargaining window runs to 14 October. --- Conclusions What the week tells us, and what to watch next The week of 18–25 September 2026 exposed the gap between what AI’s builders say and what they do. Before the UN Security Council on 23 September they asked for caution and shared standards; a day earlier they had cut prices in an all-out war, shipping Claude Opus 5.5 and GPT-6 Sol/Luna. The rhetoric is safety; the commercial logic is acceleration. For work, the mechanism is price. OpenAI’s Luna at $0.10/$0.50 for routine tasks, and maturing agents that already resolve most of a service queue, make automating first-rung cognitive work economically trivial — even as Yale’s Budget Lab still sees no clear aggregate footprint. The worker-side response has moved to the contract: “human-in-the-loop” clauses and 90-day notice, with OHSU bargaining live this week. Watch next: how far the price war runs; independent benchmark replication; California’s 30 September signings; and whether the UN plea turns into standards — or stays rhetoric while the price list keeps falling. --- Sources & references [ 01 SiliconANGLE — Anthropic releases Claude Opus 5.5 and OpenAI counters with two cheaper GPT-6 models 22 September 2026 https://siliconangle.com/2026/09/22/anthropic-releases-claude-opus-5-5-and-openai-counters-with-two-cheaper-gpt-6-models/ ](https://siliconangle.com/2026/09/22/anthropic-releases-claude-opus-5-5-and-openai-counters-with-two-cheaper-gpt-6-models/) [ 02 CNBC — Anthropic and OpenAI roll out cheaper models in first release since call for slowdown 22 September 2026 https://www.cnbc.com/2026/09/22/anthropic-openai-cheaper-ai-models.html ](https://www.cnbc.com/2026/09/22/anthropic-openai-cheaper-ai-models.html) [ 03 Fortune — What slowdown? OpenAI, Anthropic release dueling models as AI price wars heat up 22 September 2026 https://fortune.com/2026/09/22/what-ai-slowdown-openai-anthropic-release-dueling-moreaffordable-models-as-ai-price-wars-heat-up/ ](https://fortune.com/2026/09/22/what-ai-slowdown-openai-anthropic-release-dueling-moreaffordable-models-as-ai-price-wars-heat-up/) [ 04 Engadget — Anthropic and OpenAI announce more powerful (and cheaper) AI models 22 September 2026 https://www.engadget.com/2265801/anthropic-and-openai-announce-more-powerful-and-cheaper-ai-models/ ](https://www.engadget.com/2265801/anthropic-and-openai-announce-more-powerful-and-cheaper-ai-models/) [ 05 DigitalApplied — AI Model Releases: September 2026 Tracker (Qwen3.8-Omni-Flash, Grok 4.7) 18–21 September 2026 https://www.digitalapplied.com/blog/ai-model-releases-september-2026-tracker ](https://www.digitalapplied.com/blog/ai-model-releases-september-2026-tracker) [ 06 CNN Business — Sam Altman, Dario Amodei urge UN Security Council to adopt international AI standards 23 September 2026 https://www.cnn.com/2026/09/23/tech/altman-amodei-ai-safety-un-security-council ](https://www.cnn.com/2026/09/23/tech/altman-amodei-ai-safety-un-security-council) [ 07 France 24 — AI leaders urge caution at UN, with Anthropic chief pledging to slow down 23 September 2026 https://www.france24.com/en/americas/20260923-ai-leaders-urge-caution-at-un-with-anthropic-chief-pledging-to-slow-down ](https://www.france24.com/en/americas/20260923-ai-leaders-urge-caution-at-un-with-anthropic-chief-pledging-to-slow-down) [ 08 CNBC — Altman pushes for AI cooperation at UN after Trump rebuffs controls 23 September 2026 https://www.cnbc.com/2026/09/23/altman-amodei-un-ai-safety.html ](https://www.cnbc.com/2026/09/23/altman-amodei-un-ai-safety.html) [ 09 Bloomberg — Altman, Amodei Call on UN, World Leaders to Boost AI Safety 23 September 2026 https://www.bloomberg.com/news/articles/2026-09-23/altman-amodei-call-for-global-cooperation-on-ai-to-boost-safety ](https://www.bloomberg.com/news/articles/2026-09-23/altman-amodei-call-for-global-cooperation-on-ai-to-boost-safety) [ 10 Yoshua Bengio — Professor Yoshua Bengio's speech at the UN Security Council 23 September 2026 https://yoshuabengio.org/en/blog/professor-yoshua-bengios-speech-un-security-council ](https://yoshuabengio.org/en/blog/professor-yoshua-bengios-speech-un-security-council) [ 11 BNN Bloomberg / AFP — AI's dangers 'real and imminent,' Yoshua Bengio tells UN 23 September 2026 https://www.bnnbloomberg.ca/business/artificial-intelligence/2026/09/23/ais-dangers-real-and-imminent-canadian-ai-pioneer-yoshua-bengio-warns/ ](https://www.bnnbloomberg.ca/business/artificial-intelligence/2026/09/23/ais-dangers-real-and-imminent-canadian-ai-pioneer-yoshua-bengio-warns/) [ 12 UN News — OpenAI and Anthropic brief Security Council amid 'real and imminent' threat 23 September 2026 https://news.un.org/en/story/2026/09/1168414 ](https://news.un.org/en/story/2026/09/1168414) [ 13 UN — Ongoing Efforts to Create Powerful AI a 'Race Where Everyone Loses' (SC/16462) 23 September 2026 https://press.un.org/en/sc/16462.doc.htm ](https://press.un.org/en/sc/16462.doc.htm) [ 14 Salesforce — Agentforce: a New Portfolio of Job-Ready AI Agents (antecedent) 11 September 2026 https://www.salesforce.com/news/stories/agentforce-job-ready-ai-agents/ ](https://www.salesforce.com/news/stories/agentforce-job-ready-ai-agents/) [ 15 Unite.AI — Salesforce Debuts Job-Ready Agentforce Agents and Long-Horizon Runtime (antecedent) 11 September 2026 https://www.unite.ai/salesforce-debuts-job-ready-agentforce-agents-and-long-horizon-runtime/ ](https://www.unite.ai/salesforce-debuts-job-ready-agentforce-agents-and-long-horizon-runtime/) [ 16 PwC — AI reshapes global labour market into two distinct paths: 2026 Global AI Jobs Barometer (context) 2026 https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html ](https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html) [ 17 NW Labor Press — AI, meet CBA (union AI clauses) (antecedent) 2 September 2026 https://nwlaborpress.org/2026/09/ai-meet-cba/ ](https://nwlaborpress.org/2026/09/ai-meet-cba/) [ 18 The Budget Lab at Yale — Tracking the Impact of AI on the Labor Market (antecedent) 15 September 2026 https://budgetlab.yale.edu/research/tracking-impact-ai-labor-market ](https://budgetlab.yale.edu/research/tracking-impact-ai-labor-market) [ 19 Governor of California — Newsom signs SB 1050 (AI-generated advertising) (antecedent) 16 September 2026 https://www.gov.ca.gov/2026/09/16/governor-newsom-signs-new-law-to-protect-workers-require-disclosures-on-ai-generated-advertising/ ](https://www.gov.ca.gov/2026/09/16/governor-newsom-signs-new-law-to-protect-workers-require-disclosures-on-ai-generated-advertising/) [ 20 White & Case — EU agrees Digital Omnibus deal to simplify AI rules (antecedent) 2026 https://www.whitecase.com/insight-alert/eu-agrees-digital-omnibus-deal-simplify-ai-rules ](https://www.whitecase.com/insight-alert/eu-agrees-digital-omnibus-deal-simplify-ai-rules) [ 21 Anthropic — Claude discovers a novel enzyme system (ART, CRISPR-like) 23 September 2026 https://www.anthropic.com/news/claude-discovers-novel-enzyme-system ](https://www.anthropic.com/news/claude-discovers-novel-enzyme-system) [ 22 Al Jazeera — AI model Claude discovers CRISPR-like enzyme system, Anthropic says 24 September 2026 https://www.aljazeera.com/economy/2026/9/24/ai-model-claude-discovers-crispr-like-enzyme-system-anthropic-says ](https://www.aljazeera.com/economy/2026/9/24/ai-model-claude-discovers-crispr-like-enzyme-system-anthropic-says) [ 23 Gizmodo — Claude Found a Mysterious CRISPR-Like System, but Anthropic Can't Say What It's Capable Of 24 September 2026 https://gizmodo.com/claude-found-a-mysterious-crispr-like-system-but-anthropic-cant-say-what-its-capable-of-2000816906 ](https://gizmodo.com/claude-found-a-mysterious-crispr-like-system-but-anthropic-cant-say-what-its-capable-of-2000816906) **Sources & classification:** public sources from 18–25 September 2026, with dated antecedents noted in the text (Salesforce’s agents of 11 Sept; Yale Budget Lab’s update of 15 Sept; SB 1050 on 16 Sept; the LawZero grant of 16 Sept; the EU Digital Omnibus agreed in spring; PwC’s 2026 Barometer, as context). Written 25 September 2026, 08:17 CEST (06:17 UTC; 23:17 PT on 24 Sept); every event cited as having happened predates that moment. English-language statements are quoted in the original; in the Italian edition a translation is added in parentheses where useful, and no translated quotation is presented as the original. CONFIRMED: the release of Claude Opus 5.5 and GPT-6 Sol/Luna (22 Sept) and their prices; the UN Security Council session (23 Sept) and the statements by Altman, Amodei and Bengio; the release of Qwen3.8-Omni-Flash (18 Sept) and Grok 4.7 (21 Sept); Anthropic’s report that Claude surfaced the “ART” enzyme system and the method described (\~950 agents, 21 hours) (23 Sept). REPORTED: vendor-reported benchmarks, not independently verified; the GitHub CPO statement via SiliconANGLE; Clem Delangue’s participation; Salesforce agent performance figures, as reported by the company; the scientific significance and possible applications of the ART system (its function is still unknown and it has not been independently characterized or validated). EXPECTED: California signings due by 30 September; independent benchmark replication; Amodei’s 5–10-year disease-cure projection. This document is not investment, legal or regulatory advice.