Anthropic Launches Claude Haiku 5.5, 75% Cheaper Than Predecessor

Key Takeaways

• <ul><li>Anthropic released Claude Haiku 5.5, priced 75% lower than Haiku 4.5 at $0.10 per million input tokens</li><li>Haiku 5.5 scores 43 on Artificial Analysis Intelligence Index, up from Haiku 4.5's score of 17</li><li>The model beats OpenAI's GPT-6 Luna and Gemini 3.8 Flash but trails Kimi K3 by one point</li><li>OpenAI published a quasi-Riemann Hypothesis paper and released 722 manuscripts with over 300 math results</li><li>Google expanded its SynthID AI content detector globally, now supporting watermarks from all AI partners</li><li>A 2026 study found 81% of leaders expect AI agent integration within 12-18 months, but only 37% assess security beforehand</li><li>Vijil launched its DART module to test AI agent security using adversarial agents before and after deployment</li><li>Asia Pacific data center pipeline hit a record 26.5GW in H1 2026, but returns depend on more than capacity</li><li>Economist Paul Krugman warns AI data center spending is crowding out investment in housing and factories</li><li>Utah pilot program uses AI to prescribe acne treatments for $4.99 per patient visit</li></ul>

Anthropic launches Claude Haiku 5.5 as its fastest and cheapest small AI model

Anthropic released Claude Haiku 5.5, a small AI model designed for tasks like summaries, database queries and classification. The company says it runs about 75% cheaper than Haiku 4.5 and scores higher on coding, computer use and knowledge tests. It also beats OpenAI's GPT-6 Luna across reported benchmarks. Haiku 5.5 introduces adjustable effort settings for developers.

Anthropic unveils Haiku 5.5 as AI pricing competition heats up

Anthropic revealed its new Haiku 5.5 lightweight AI model, claiming strong gains over Haiku 4.5 in coding and knowledge work. The model is priced at $0.10 per million input tokens and $0.50 per output token. Industry pricing pressure is growing as companies seek to cut AI costs. The launch comes as Anthropic moves closer to a planned IPO.

Anthropic adds third Claude 5.5 model ahead of planned IPO

Anthropic launched Claude Haiku 5.5, the third model in its Claude 5.5 family. The model targets classification, summarization and live support tasks. It costs 75% less than Haiku 4.5 and is priced at $0.10 per million input tokens and $0.50 per output token for shorter prompts. It also includes new safeguards for certain cybersecurity requests.

Claude Haiku 5.5 scores 43 on Artificial Analysis Intelligence Index

Artificial Analysis tested Claude Haiku 5.5 and gave it a score of 43 at max effort on its Intelligence Index. That is a big jump from Haiku 4.5's score of 17. It beats GPT-6 Luna and Gemini 3.8 Flash but sits one point behind Kimi K3. The model uses more tokens than some rivals to reach similar scores.

OpenAI paper explores quasi Riemann Hypothesis breakthrough

OpenAI published a paper called The Quasi-Riemann Hypothesis, pushing back on where zeros of the Riemann zeta function can appear. The result does not prove the famous Riemann Hypothesis but narrows the possible zone for its zeros. The finding also applies to related functions like Dirichlet L-functions. The work was shared as part of a research collaboration.

OpenAI Releases 300+ Math Proofs as Sam Altman Reflects on AI Achievements

OpenAI released 722 manuscripts containing over 300 mathematical results, including the quasi-Riemann hypothesis. CEO Sam Altman expressed awe and gratitude toward the machines and generations of mathematicians. Some results have been verified by Lean, a computer proof system, but outside mathematicians have not yet confirmed all claims. OpenAI's own documentation notes that not all results have formalizations and some could have issues.

Vijil Launches DART Module to Test AI Agent Security

Vijil introduced its Diamond Adaptive Red Teaming for Agents (DART) module to test AI agent security before and after deployment. The module uses adversarial AI agents to probe workflows and adjust attacks based on responses. DART integrates with agent frameworks and CI/CD pipelines using APIs. It includes tests based on OWASP and MITRE frameworks that can be tuned for specific IT environments.

Asia Pacific AI Infrastructure Returns Depend on More Than Power

Asia Pacific's data center pipeline reached a record 26.5GW in H1 2026, with 7.1GW added in six months. However, the report from Oxford Economics says returns depend on connectivity, adoption, and policy rather than just capacity. Data center electricity consumption in Australia could rise from 5.2 TWh in FY26 to 34.3 TWh by FY36. The study says governments should measure success by what investment enables the wider economy to do, not by how many data centers are built.

Paul Krugman Warns AI Investment Crowding Out Other Sectors

Economist Paul Krugman wrote that massive spending on data centers is crowding out investment in other parts of the economy. He said this differs from the dot-com era because foreign investment is not cushioning the effect today. JPMorgan CEO Jamie Dimon expects hyperscalers to spend nearly $220 billion in debt this year. Data center construction is diverting funds from housing, offices, and factories, Krugman wrote, citing FRED data showing tech investment crowding out other investment since early 2023.

AI Agent Adoption Outpaces Security Preparedness, Study Finds

A 2026 review by Whitfield Research Partners found 81% of leaders expect AI agents to be integrated into corporate strategies within 12-18 months, but only 37% assess AI tool security before deployment. MSSP Security ranked first in the study with a score of 91.0 out of 100. The review found 13% of organizations reported a breach involving AI models or applications, and 97% of those lacked proper AI access controls. Extensive use of security AI and automation was associated with $1.9 million lower average breach costs.

Google launches global SynthID AI content detector

Google released an improved version of its SynthID AI content detector that is now available worldwide. The tool uses machine learning to identify AI-generated text and images by checking for invisible watermarks. It now supports watermarks from all of Google's AI partners. The detector is available on a dedicated website for anyone to use. Researchers and journalists can use it to verify if content is authentic.

CDT submits comments on NIST AI evaluation framework

The Center for Democracy and Technology joined a coalition led by the National Fair Housing Alliance to comment on NIST's TEVV-Athlon Framework for Evaluating AI Systems. The framework helps organizations test, evaluate, verify, and validate AI systems to manage risks. CDT praised the effort and recommended improvements for discovery-oriented evaluations that uncover unknown risks. The comments also suggested better guidance on real-world testing scenarios and broader public participation.

Utah pilot uses AI to prescribe acne treatments

A Utah pilot program by Nolla Health uses AI to evaluate patients and recommend acne prescriptions without a doctor visit. Patients pay $4.99, submit face photos, and receive treatment recommendations from the AI system. Licensed providers are available via telehealth, and non-topical medications are escalated to human clinicians. The first 100 patients have their AI recommendations reviewed before sending, while Utah officials monitor the program with monthly reports.

Ethereum researcher recommends 4 steps against AI security risks

Ethereum researcher Drake shared four steps to protect cryptocurrency from AI-driven security threats. He recommends moving funds to a never-used address and switching to a fresh address after each transaction. Drake warned against rushing or panicking and noted that wallets under 50 Bitcoin have partial protection. He said the move requires no new software or wallets but a break could happen within months in the worst case.

Sources

• Anthropic Launches Claude 5.5 Haiku, Beats GPT-6 Luna Across Most Benchmarks While Priced at 1/20th Of Sonnet 5.5
• Anthropic reveals Haiku 5.5 model as AI pricing war intensifies
• Anthropic launches third Claude 5.5 model, expanding AI lineup before planned IPO
• Claude Haiku 5.5 Scores 43 On Artificial Analysis Index, Just Behind Kimi K3 While Ahead Of GPT-6 Luna
• OpenAI Solves Quasi Riemann Hypothesis: Understand What It Means
• “Looking Up At Stars With Extra Awe Tonight”: Sam Altman After OpenAI Publishes 300+ Math Proofs Including Quasi-Riemann Hypothesis
• Vijil Adds Red AI Agent Team to Stress Test AI Agent Security
• Beyond the gigawatts: what determines the return on Asia Pacific’s AI infrastructure
• Paul Krugman Says AI Boom Is Crowding Out Other Investment - State Street SPDR S&P 500 ETF Trust (ARCA:SPY)
• MSSP Gartner Magic Quadrant Context: AI-Agent Adoption Outpaces Security Assurance as 81% of Leaders Plan Integration Within 18 Months
• Google rolls out improved SynthID AI content detector, now available globally
• CDT Joins Coalition Comments on NIST’s AI Evaluation Framework
• Utah pilot lets artificial intelligence prescribe acne treatments
• 4 Steps to Protect Your Coins From AI-Driven Security Risks, Says Ethereum Researcher