How ten jurisdictions are governing frontier AI, what the biggest models cost, and what that means for value, safety and governance over the next quarter and the next five years.
Method: facts come from the sources linked below. The 1-5 scores are Signal Fire Aim analyst judgments that make week-over-week movement visible; they are not from any index. Model prices come from a third-party tracker and two sources conflict on one OpenAI price, flagged below. This page is the running baseline, so read movement week to week.
Week-over-week: This is the first refresh after the baseline, published the same day, so there are no score changes. New this pass: the UK's cross-party pressure on superintelligence risk (8-14 Sep), the EU's information requests to 30+ AI companies (1 Sep), California's companion-chatbot law (signed 10 Sep), Anthropic's Sonnet 5.5 (28 Sep) added to the cost chart, and Grok 4.7 and Xiaomi MiMo-V2.6-Pro added as price points. Items older than 7 days are marked as background.
Anthropic's Opus 5.5 (22 Sep) lists at $4 in / $20 out per million tokens, and Sonnet 5.5 (28 Sep) at $2 / $10, per Anthropic's page and a third-party tracker. OpenAI's GPT-6 Sol is listed at $2 / $10. Anthropic says Opus 5.5 was tested by external evaluators including METR, yet it concedes evaluations cannot yet catch every failure before release.
The EU delayed high-risk duties to Dec 2027 and Aug 2028 but on 1 Sep sent information requests to 30+ AI companies, a step toward formal AI Act checks. The US used the 1 Sep G20 meeting to argue against AI-specific rules, while California keeps legislating (13 child-safety bills signed 10 Sep).
Russia's law keeps "sovereign" models in domestic data centers, the UAE ships open models and hosts Stargate, and Israel bets on niches and alliances. In the UK, 40 MPs and 31 peers asked on 11 Sep for an international ban on superintelligence, so the pressure for limits is now political as well as technical.
Solo.io is now the lead company on this page. Two September releases show why: it is building governance for agents at both the gateway and the desktop.
10 Sep 2026. Runs without Kubernetes (VMs, ECS, bare metal, workstations) with identity via token exchange and OAuth, a read-write management UI, model and MCP tool governance, token budgets with spend attribution, and traces to existing backends. Built on the open-source agentgateway project.
3 Sep 2026, Apache 2.0. Discovers AI harnesses, models, MCP servers and skills on employee devices, applies central policy with drift reconciliation, treats agents as non-human identities with just-in-time credentials, and offers fleet mode through MDM. Solo cites 52% of employees using unapproved AI tools.
Older than 7 days. The open-source project shipped v1.4 on 3 Aug with support for MCP spec 2026-07-28, and on 27 Aug announced availability on AWS, Azure and Google Cloud marketplaces. No newer Solo.io announcement after 10 Sep was found this week.
Rubrik/Strata: nothing new verified this week.
Sources: GlobeNewswire, 10 Sep · Yahoo Finance, agentdesktop · agentgateway blog. Both are company announcements, so treat the claims as vendor-stated.
Scores: 1 is low, 5 is high. "Binding safety" means legally enforceable duties on frontier developers.
| Jurisdiction | Binding safety | Governance formality | Pro-growth tilt | Focus and latest signal |
|---|---|---|---|---|
| EU | 4 | 5 | 2 | Rulebook first. Digital Omnibus (Reg. 2026/1744) in force 27 Jul 2026: high-risk duties delayed to 2 Dec 2027 / 2 Aug 2028, watermarking grace to 2 Dec 2026, new bans on AI-generated CSAM and non-consensual intimate content, AI Office powers over GPAI providers expanded. Changed this week: 1 Sep information requests to 30+ AI companies, a step toward AI Act investigations (Al Jazeera, 2 Sep). Scores unchanged. |
| UK | 2 | 3 | 4 | Voluntary and security-led. The AI Security Institute tests models with developers before release. The Cyber Security and Resilience Bill targets AI users, not vendors; Lords are pressing for "red lines" the government rejected. Changed this week: Joint Committee on Human Rights urged a single AI regulator (14 Sep); a private member's superintelligence bill was introduced, and the government signaled it would not back it (Simmons & Simmons). Scores unchanged. |
| USA | 2 | 2 | 5 | Federal push to preempt states. EO 14365 (Dec 2025) directs DOJ to challenge state laws; the preemption bill is still a discussion draft with 200+ state lawmakers opposed. California SB 53 is live and under a second compliance dispute. Colorado rewrote its act, effective 1 Jan 2027. Changed this week: US argued against AI-specific rules at the 1 Sep G20 meeting; California signed 13 child-safety bills on 10 Sep, including a companion-chatbot law; a House subcommittee advanced the Open-Source AI Leadership Act (1 Sep). Scores unchanged. |
| China | 3 | 4 | 5 | Dense sectoral rules (generative AI, algorithms, deepfakes, labeling) plus the "AI Plus" plan: 70% penetration in key sectors by 2027, 90% by 2030. Governance serves state control and industrial scale. Changed this week: no new binding rule verified in the last 7 days. Alibaba's Qwen3.8 and DeepSeek V4.1 Flash releases continue the low-price push. Scores unchanged. |
| UAE | 2 | 3 | 5 | Capital and compute. MGX targets $100B in assets, Stargate UAE with OpenAI, US chip access tied to G42 distancing from China. Goal: 20% of non-oil GDP from AI by 2031. IFM's K2 Horizon Fleet is an open-model release under Apache 2.0. Changed this week: no new policy verified; IFM released K2 Horizon models on 3 Sep (background, older than 7 days). Scores unchanged. |
| Other Gulf | 1 | 2 | 5 | Same oil-to-compute playbook (Saudi Arabia and neighbors) with data-security and localization rules rather than model safety law. Coverage is thin this week; a deeper pass is queued. Changed this week: nothing verified for Saudi Arabia or Qatar in the last 7 days; searches returned only older strategy material. Scores unchanged. |
| Israel | 2 | 3 | 4 | 2026 national AI strategy: niches over scale (cyber, fintech, health, physical AI), regulatory sandboxes, standards work through ISO, OECD and UN, and alliance access to chips ("Pax Silica"). National supercomputer run by Nebius. Very high Claude usage per capita. Changed this week: nothing new verified. Scores unchanged. |
| Russia | 3 | 3 | 3 | Control-first. The September 2026 law defines "sovereign" and "national" AI that must run in Russian data centers and follow state speech rules. Constraints: chip imports down 84% in 2024, Sberbank seeking Chinese processors for GigaChat. Export pitch to Africa, Asia and Latin America. Changed this week: nothing new verified beyond the September law already listed. Scores unchanged. |
| Hong Kong | 2 | 2 | 4 | No AI statute; the Personal Data (Privacy) Ordinance carries the load. The privacy commissioner's 2026 checks found 95% of 60 organizations use AI but only 50% have AI policies (down from 63%), and it issued guidance on agentic AI: minimum access rights, segregated environments, human in the loop. The 16 Sep 2026 Policy Address adds a Commissioner for AI, an AI+ industry committee, an expanded HKMA GenA.I. Sandbox++ and HK$300M for an AI for Welfare Lab. Changed this week: nothing new since the 16 Sep Policy Address. Scores unchanged. |
| Australia | 3 | 3 | 4 | Reversal in July 2026. The Dec 2025 National AI Plan rejected mandatory guardrails; on 15 Jul 2026 the Prime Minister announced Australian Standards for AI, an Office of AI in the PM's department, and legislation expected early 2027. Confirmed so far: mandatory rules for large AI data centres (energy, water, copyright). Scope, penalties and developer-versus-deployer duties are undefined. An AI Safety Institute and an ISO/IEC 42001 certification path already exist. Changed this week: nothing new verified since the July reversal. Scores unchanged. |
Sources: White & Case (EU) · The Register (UK) · CASRAI (US) · IAPP (China) · CSIS (UAE) · VC Cafe (Israel) · Atlantic Council (Russia) · Legal 500 (Hong Kong) · HK Policy Address 2026 · Compliance Council (Australia).
Output price per million tokens, in US dollars. Lower is cheaper. Same-tier comparisons matter more than the ranking.
Source: Capital and Compute tracker and Fortune, 22 Sep. Conflict to resolve: the tracker lists GPT-6 Astra as the "first frontier price rise in years" at $10 / $50 while also listing GPT-6 Sol at a halved $2 / $10, so Astra is a premium tier. Russian and Israeli models had no comparable public pricing this week. Sonnet 5.5 and Opus 5.5 prices match Anthropic's own page for Opus and the tracker for Sonnet. The previous chart label "Fugu Max v2" was corrected to Sakana's Fugu Max ($6; Fugu Ultra v2 is $30). K2 Horizon Fleet (UAE) is free open weights, so it has no price bar.
The same four lenses every week, short term (next 12 months) and long term (3-5 years).
| Lens | Short term | Long term |
|---|---|---|
| Financial value | Falling prices squeeze lab margins and widen buyer choice. Cheap Chinese options ($0.30 / $1.20 for DeepSeek V4.1 Flash) cap what US labs can charge for routine work. | Value migrates to whoever owns distribution, identity and governance layers, not raw model access. |
| Safety | Mostly voluntary outside California and the EU. Watch the SB 53 dispute over OpenAI's releases. Anthropic's Opus 5.5 card adds external testing (including METR) and says pre-release evaluations remain incomplete. | A joint call to slow frontier development was made and not followed; enforceable limits will likely arrive after an incident. |
| Governance | EU deadlines slip but AI Office powers grow. US preemption unresolved. Agent identity and authorization are the practical control point. | Blocs diverge: EU rulebook, US market, China state-led, Gulf capital, Russia closed, with Australia moving toward standards and Hong Kong relying on privacy law. Firms will need per-region policy. |
| Cost | Caching and inference gains are being passed through. Expect continued cuts; Google's Flash rate doubles on 1 Jan 2027. | Compute and energy, not model licenses, become the binding cost. UAE's 5+ GW surplus by 2030 is the example. |
This page refreshes every Monday with new scores, model releases and price moves, and one lead item per jurisdiction. A big signal gets a red banner across the site. Gaps to close first: Saudi Arabia and Qatar detail, Russian and Israeli model releases, and independent safety evaluations of the newest US and Chinese models.
The Global AI Watch changes, the Solo.io lead item, and the one identity signal that matters. Free. No ads. Links always in the comments.
Subscribe to the Weekly BriefEvery crisis surfaces the people who saw it coming. Get the weekly brief.