The New Battlegrounds | AI Wars of 2026
In research: the five questions that shape the contest. Source-backed analysis from AI Wars of 2026.
Part V — The New Battlegrounds
The New Battlegrounds
In research
This chapter maps the questions ahead. It is still in research, and its topics show readers what the next evidence increment must answer.
The shift from chat to delegated work creates five connected questions that cannot responsibly be answered through product announcements alone:
- Agent capability: Which systems can complete multi-step work reliably, with what tools, oversight, and recovery paths?
- Safety and control: How do permissions, prompt injection, consequential actions, and human approval affect what an agent can safely do?
- Economics: What do inference, tool use, orchestration, integration, and human review cost per useful task?
- Talent: How do research, engineering, product, and leadership capabilities translate into durable execution advantages?
- Government and regulation: How could policy, competition rules, privacy, copyright, procurement, and national strategy alter the available routes to market?
These are not five universal leaderboards. They are the shared conditions that can strengthen, constrain, or redirect a contender's position in different wars.
Economics: one independent benchmark cost comparison
Anthropic's Opus 5.5 announcement compares model costs without publishing the competitor settings or task counts. Artificial Analysis provides a separate comparison on its disclosed Intelligence Index. At matched maximum effort, Claude Opus 5.5 scores 58 at $5.98 per task, GPT-6 Astra scores 53 at $3.26, and GPT-5.6 Sol scores 47 at $1.99. The model with the highest score also has the highest cost on this suite. Source 1 Source 2 Source 3 Source 4
This comparison covers one benchmark suite and one effort setting. It does not reproduce Anthropic's tasks or measure costs in a customer workflow. The proposed economics chapter requires comparable operational or realized-return records. This benchmark does not meet that requirement, and the chapter's evidence threshold remains unmet.
Government and regulation: different instruments, different claims
The European Commission states that Cyber Resilience Act reporting obligations began on September 11, 2026. They concern actively exploited vulnerabilities and severe incidents affecting the security of products with digital elements. This is a reporting duty, not evidence of how many organizations comply or whether reporting prevents harm. The source does not establish that every AI service falls within the same legal scope. Source 5
The UK's Joint Committee on Human Rights takes a different position: its September report calls for an AI Bill and a single AI regulator. These are parliamentary recommendations, not enacted requirements or a government commitment to implement them. They identify a possible change in oversight rather than a new law already governing market access. Source 6
China's September filing notice records completed national filings of generative-AI services separately from local registrations of applications or functions using already-filed models. Those populations cannot be added together as a count of companies, users, or deployments. The notice documents administrative procedures and their recorded completion, not adoption, compliance rates, or enforcement effectiveness. It supplies no contender comparison here. Source 7
The strategic questions differ accordingly. A duty can change what covered operators must report. A recommendation may alter future oversight if adopted. An administrative record shows activity within a particular procedure. None of these sources measures the resulting effect on competition, and they do not form a common scale of regulatory strictness.
Reuters reported on September 30 that the US Federal Trade Commission had opened an industry-wide probe into OpenAI, Anthropic, and METR over rogue AI agents, with civil investigative demands and executive testimony planned under the Commission's authority over unfair or deceptive practices. This would be the first official US enforcement action on agent safety. A probe is not a lawsuit: no complaint has been filed, and the reporting does not establish that any company violated the law. The same week, the White House announced a voluntary accord on superintelligence signed by the leaders of six AI companies, with commitments to internal controls, independent auditors, a board committee, and standard-setting. The accord is voluntary, with no penalties and no disclosure requirements. A self-chosen auditor is not a regulator, and the timing of a probe alongside an accord does not establish that either instrument will constrain behavior. Source 8Source 9
Product controls remain a separate question, examined in Controlled Action. Operational economics also needs its own evidence, such as the bounded fixed-power comparison in NVIDIA. Neither technical controls nor operating performance can stand in for legal compliance.
Plans, proposals, and filings
Yonhap reported on September 17 that South Korea's National Intelligence Service plans to extend safety evaluation from large language models to autonomous AI agents and distribute red-teaming guidance and a security playbook in December. This is an attributed account of planned evaluation and guidance, not evidence that an agent-evaluation regime is already operating or that its controls are effective. Source 10
South Korea's Personal Information Protection Commission records field interviews held in July and August with fifteen AI development and deployment organizations. Its meeting summary identifies prompt injection and excessive agent permissions as privacy risks. It also calls for rules covering approval boundaries, participant responsibility, and retention of logs and memory. PIPC plans a guide by the end of 2026. This is a supervisory meeting record, not the interview instrument, adopted guidance, or a measured privacy outcome. Source 11
CourtListener's RECAP record lists a September 16 copyright complaint in Times Publishing Company v. Microsoft Corporation, naming Microsoft and OpenAI entities as defendants in the Southern District of New York. The record establishes that a complaint was filed, not that its allegations are true or that liability has been determined. The exhibit titles do not establish what the exhibits contain or whether they support the claims. Source 12
The European Commission's September 16 announcement describes the joint design of an Important Project of Common European Interest in AI, coordinated by Germany. Participating states announced that they would begin pre-notifying State aid projects in September, after which the Commission would assess them. This records a plan to begin a procedure, not completed filings, approved funding, disbursement or deployed AI capacity. Source 13
The European Data Protection Board's September 17 agenda lists guidelines on the relationship between the AI Act and EU data-protection law as a discussion item. The same agenda separately records adoption of unrelated guidelines. No draft or adopted AI Act and data-protection guideline exists at this citation. The agenda does not establish rules for training data, automated decisions, or agent-operator roles. Source 14
The Board's September 21 plenary advanced beyond discussion: it adopted Guidelines 04/2026 on the methodology for calculating administrative fines, which went to public consultation through November 13, and it finalized guidelines on the interplay between the DSA and GDPR for notice-and-action, deceptive design, ad transparency, recommender systems, and minors. The fining methodology remains a consultation draft until the Board adopts a final text; the adopted guidelines do not themselves set fine amounts. Source 15
On September 14 the national AI safety governance framework 3.0 was released in China, with an annex stating the agent-risk-management framework and trustworthy-AI baseline criteria, emphasizing preventing agent behavior from going out of control. This is a governance framework release, not an enforcement action or a compliance deadline. Source 16
SemiAnalysis supplied a hard number on the gap between framework and disclosure. Nine Chinese firms including Alibaba, Tencent, ByteDance, Baidu, and DeepSeek released 857 models between 2021 and September 15, 2026, and only 31 published specific safety-test results — 3.6%, with just 9 disclosing before or at release. China's AI regulation targets app services and user impact rather than mandating capability-based public risk assessment, which helps explain the gap. The figure counts only public disclosures; firms may test internally without publishing. Source 17
The China Academy of Information and Communications Technology reported 124 new generative AI services filed in July and August 2026 (seven on-device), bringing the cumulative total to 1,112 services, with 133 new registrations and 731 cumulative application or function registrations. These are filing statistics, not adoption or usage measures, and they do not establish market share or commercial traction. Source 18
China's Cyberspace Administration opened public comment through October 17 on a draft regulation for minors' online safety. Article 5(3) includes AI services that may affect minors' cognition among services that must route users under 16 through a minor mode. Article 9 requires providers to offer the switch, and Article 8 bars designs that induce emotional dependency, addiction, or excessive consumption in minors. This is a consultation draft, not an enacted rule. Source 19
The European Commission adopted its KIDS Act legislative proposal on September 17. The proposal covers online services for minors, including AI chatbots and companions. Commission adoption begins the legislative process; it does not create an enacted requirement. The Commission page does not establish the final obligations for a particular service or whether its proposed controls would be effective. Source 20
What will make this chapter publishable
Each topic needs a defined question, dated source-backed evidence, a material counterargument, and an explicit statement of what the evidence does not establish. Any cross-contender comparison must use compatible measures rather than turn separate product claims into a ranking.
Sources
- Source 5
- Source 6
- Source 7
- Source 10
- Source 11
- Source 12
- Source 13
- Source 14
- Source 15
- Source 8
- Source 9
- Source 16
- Source 17
- Source 18
- Source 19
- Source 20
- Source 1
- Source 2
- Source 3
- Source 4
Source references
Source 1: Intelligence Benchmarking Methodology
Artificial Analysis
Artificial Analysis's published methodology for Intelligence Index v4.3.2 combines ten evaluations weighted 30% agents, 20% coding, 20% scientific reasoning, and 30% general. It establishes a disclosed, versioned comparison method across vendors, not agreement with a vendor's chosen tasks or a measure of customer-workload cost.
Source 2: Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) - Model Data
Artificial Analysis
Artificial Analysis reports Claude Opus 5.5 at maximum effort scoring 58 on its Intelligence Index, generating 260 million tokens across the suite, and averaging $5.98 per task. This is one independent benchmark operator's suite and effort setting, not a measure of Anthropic's customer workloads or chosen comparison tasks.
Source 3: GPT-6 Astra (max) - Model Data
Artificial Analysis
Artificial Analysis reports GPT-6 Astra at maximum effort scoring 53 on its Intelligence Index, generating 60 million tokens across the suite, and averaging $3.26 per task. This is one independent benchmark operator's suite and effort setting, not a measure of customer-workload cost.
Source 4: GPT-5.6 Sol (max) - Model Data
Artificial Analysis
Artificial Analysis reports GPT-5.6 Sol at maximum effort scoring 47 on its Intelligence Index, generating 90 million tokens across the suite, and averaging $1.99 per task. This is one independent benchmark operator's suite and effort setting, not a measure of customer-workload cost.
Source 5: Cyber Resilience Act - Reporting obligations
European Commission
Published: 2026-09-11
From 11 September 2026, manufacturers must report actively exploited vulnerabilities and severe incidents affecting the security of products with digital elements: an early warning within 24 hours of becoming aware, a full notification within 72 hours, and a final report no later than 14 days after a corrective measure is available for an actively exploited vulnerability, or within a month of the 72-hour notification for a severe incident. Reporting runs once through ENISA's CRA Single Reporting Platform, operational from the same date, addressed to the CSIRT where the manufacturer has its main establishment and shared with ENISA. Open-source software stewards become subject to the Article 24(3) obligations from 11 December 2027 under Article 71(2). The page establishes the obligation and its dates; it states no penalty amounts, and any figure for fines needs the legal text rather than this page. It establishes nothing about compliance, enforcement or the volume of reports actually filed.
Source 6: Wide-ranging AI Bill needed to address severe human rights risks posed by AI
UK Parliament, Joint Committee on Human Rights
Published: 2026-09-14
The committee calls for a new AI Bill covering the whole AI lifecycle and supply chain, a risk-based regime with more demanding obligations for higher-risk systems, outright prohibition of uses incompatible with human rights, and transparency so an individual knows when AI is used in a decision affecting them. Its chair Alex Sobel MP states that 'New legislation is needed' and that 'A single AI regulator should be established'. The committee also records that UK regulators lack the power to test or evaluate AI systems before release or to prevent release, and that model developers engage with the AI Security Institute on a voluntary basis, a body it describes as having no statutory power. This is a select-committee recommendation to government, not law, and not a government commitment.
Source 7: 关于发布生成式人工智能服务已备案信息的公告（2026年7月至8月）
Cyberspace Administration of China
Published: 2026-09-14
The Cyberspace Administration of China's September 14 notice reports national filings of generative-AI services separately from local registrations of applications or functions using already-filed models. These are different administrative populations, not company or user counts. The notice does not establish adoption, compliance rates, enforcement effectiveness, or a particular contender's position; the linked entity list was not examined.
Source 8: FTC Opens Probe Into AI Giants Including Anthropic and OpenAI, New York Post Reports
SRN News (Reuters byline)
Published: 2026-09-30
Reuters reported that the FTC opened an industry-wide probe into OpenAI, Anthropic, and METR over rogue AI agents, with civil investigative demands and executive testimony planned under unfair-or-deceptive-practices authority. The article describes the probe as the first official US enforcement action on agent safety. A probe is not a lawsuit: no complaint has been filed and the reporting does not establish any law was violated. No primary FTC statement exists; investigations are nonpublic.
Source 9: OpenAI, Google and Meta pledge outside AI audits under voluntary White House deal
CoinDesk
Published: 2026-09-30
Reports a voluntary one-page accord signed September 29, 2026 by the leaders of Google, Anthropic, Meta, OpenAI, Nvidia, and xAI, covering internal controls, independent auditors, a board committee, and standard-setting. The accord carries no penalties and no disclosure requirements, and auditors are chosen by the companies themselves. The same-day launch of the America.gov federal-services chat, powered by Gemini and Grok, is also reported.
Source 10: [AI프리즘] 국정원, AI 에이전트 안전성 평가한다(종합)
Yonhap News Agency
Published: 2026-09-17
Yonhap's September 17 Korean-language report describes the National Intelligence Service's plan to extend safety evaluation from large language models to autonomous agents and distribute guidance in December. The article supports an attributed announcement, not a completed evaluation regime, the content of future guidance, or measured control effectiveness.
Source 11: Second Meeting of the Second-Term AI Privacy Public-Private Policy Council
Personal Information Protection Commission (South Korea)
Published: 2026-09-23
PIPC records field interviews held in July and August 2026 with fifteen domestic and foreign AI development and deployment organizations. Its meeting summary identifies prompt injection and excessive agent permissions as risks, and calls for rules covering approval boundaries, participant responsibility, and log and memory retention. PIPC plans an agentic-AI personal-information guide by year-end. This is a supervisory meeting record, not the interview instrument, adopted guidance, or a measured privacy outcome.
Source 12: Times Publishing Company v. MICROSOFT CORPORATION
CourtListener (RECAP)
Published: 2026-09-16
CourtListener's RECAP record for Times Publishing Company v. Microsoft Corporation lists a September 16, 2026 copyright complaint against Microsoft and OpenAI entities in the Southern District of New York. Docket metadata and the complaint description support the filing record. The complaint and exhibits have not been assessed here; their titles do not establish their contents, the truth of allegations, legal novelty or liability.
Source 13: Commission welcomes the design of first Important Project of Common European Interest in AI
European Commission
Published: 2026-09-16
The European Commission's announcement records the joint design of an IPCEI in AI, coordinated by Germany, and participating states' intention to begin pre-notifying State aid projects in September 2026. It describes a forthcoming Commission assessment, not completed notification, approved aid, disbursement or deployed capacity. It does not measure an effect on AI competition.
Source 14: 123rd Plenary Meeting Agenda
European Data Protection Board
Published: 2026-09-17
The EDPB's 123rd plenary agenda lists guidelines on the interplay between the AI Act and EU data-protection laws as a discussion item. The agenda separately records adoption of unrelated guidelines. No draft or adopted AI Act interplay text exists at this citation, which does not establish rules for training data, automated decisions, or agent-operator roles.
Source 15: EDPB September 2026 plenary: fining methodology and DSA-GDPR interplay
EFDPO / TechPolicy Press
Published: 2026-09-21
Reports that at its September 21, 2026 plenary the EDPB adopted Guidelines 04/2026 on the methodology for calculating administrative fines, open for public consultation through November 13, and finalized guidelines on the interplay between the DSA and GDPR for notice-and-action, deceptive design, ad transparency, recommender systems, and minors. The fining methodology is a consultation draft until final adoption; neither guideline sets fine amounts for any company.
Source 16: China releases AI safety governance framework 3.0
Nanfang Qiye Xinwen Wang
Published: 2026-09-29
A September 29, 2026 roundup reporting that the national AI safety governance framework 3.0 was released on September 14 by the national cybersecurity standardization committee, with an annex stating the agent-risk-management framework and trustworthy-AI baseline criteria, emphasizing preventing agent behavior from going out of control. The same article corroborates the July-August filing statistics (124 new services, 1,112 cumulative). This is a governance framework release, not an enforcement action or a compliance deadline.
Source 17: 阿里、DeepSeek等揭露多少AI安全資訊？研究盤點陸9家業者 模型安全測試揭露率僅3.6％
工商时报
Published: 2026-10-10
SemiAnalysis found nine Chinese firms (Alibaba, Tencent, ByteDance, Baidu, DeepSeek and four others) released 857 models from 2021 to Sept 15, 2026; only 31 published specific safety-test results (3.6%), only 9 before or at release (1.1%). Tier-2 relay of the SemiAnalysis study; the original report was not directly verified. Counts only public disclosures; firms may test internally without publishing.
Source 18: CAC generative-AI filing statistics for July-August 2026
Renmin Cai Xun
Published: 2026-09-14
Reports the CAC's July-August 2026 filing statistics: 124 new generative-AI services filed, seven on-device, with 133 new registrations; cumulative totals of 1,112 services and 731 applications or functions registered as of August 31. These are filing statistics, not adoption or usage measures, and they do not establish market share or commercial traction.
Source 19: Notice on Public Comment for the Regulations on the Protection of Minors in Cyberspace (Draft for Comment)
Cyberspace Administration of China
Published: 2026-09-18
CAC opened public comment through October 17, 2026 on a draft State Council regulation. Article 5(3) includes AI services that may affect minors' cognition among services that must route users under 16 through a minor mode. Article 9 requires providers to offer the switch, and Article 8 bars models and services that induce emotional dependency, addiction, or excessive consumption in minors. This is a consultation draft, not an enacted rule; final wording, passage, and enforcement remain undetermined.
Source 20: KIDS Act
European Commission
Published: 2026-09-17
The European Commission's canonical KIDS Act page confirms that the Commission adopted the legislative proposal on September 17, 2026 and links its full text through EUR-Lex as CELEX 52026PC0681. The proposal covers online services for minors, including AI chatbots and companions. Commission adoption begins the legislative process; it is not an enacted obligation. The page does not establish final applicability, enforcement outcomes, or the effectiveness of safety controls.