The frontier model race is shifting from raw capability toward the economics of putting intelligence to work. OpenAI launched two lower-cost GPT-6 models, while Anthropic released Claude Opus 5.5 with lower serving costs. Governments and companies are also testing deployment through cyber defense, search regulation, chip design and enterprise agents. Public resistance to the data centers behind this expansion is rising.
OpenAI brings GPT-6 capabilities to lower-cost models
OpenAI introduced GPT-6 Sol and GPT-6 Luna, extending GPT-6 techniques into faster and cheaper tiers. API prices are 50% below promotional pricing for the GPT-5.6 equivalents. Sol costs $2 per million input tokens and $10 per million output tokens, while Luna costs $0.10 and $0.50. OpenAI says Sol makes about half as many mistakes as its predecessor in an internal factuality evaluation.
Lower cost per task could expand how often companies use capable models. Vendor benchmarks depend on effort settings, tools and safeguards, so buyers should test complete workflows. Even so, the price cut pressures every provider serving coding agents and other high-volume applications.
OpenAI also detailed a redesigned prompt caching system that keeps eligible shared prefixes reusable for 30 minutes and discounts cached input reads by up to 90%. A dashboard, miss diagnostics, cache breakpoints and prewarming help agents reuse instructions, tools and context. GitHub said recent improvements cut fresh prompt processing by more than half across billions of requests.
Anthropic answers with Claude Opus 5.5
Anthropic released Claude Opus 5.5, saying it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 on typical workloads. Input and output prices are $4 and $20 per million tokens, and cache reads cost $0.20 per million tokens. Anthropic reports output more than 30% faster than Opus 5 and says the model leads its evaluations in agentic coding, computer use and knowledge work.
Anthropic is pairing the performance claim with operational controls. The model screens actions, runs code in an auditable open-source sandbox and applies safeguards for cybersecurity, biology and model distillation. Anthropic says Opus 5.5 crossed containment boundaries about 85% less often than Opus 5 or Mythos 5.1 in an internal evaluation, while acknowledging that pre-deployment testing remains incomplete.
OpenAI gives Ukraine access to AI cyber defense
OpenAI will provide its Daybreak cyber defense system and GPT-5.6 Sol to the Ukrainian government at no cost, according to the BBC. The tools are intended to help protect civilian infrastructure such as hospitals and power plants by identifying vulnerabilities and supporting fixes. Ukraine's national incident response team recorded nearly 6,000 cyberattacks in 2025.
The deployment highlights the dual-use nature of frontier models. The same ability to find weaknesses can strengthen defenders or help attackers. Giving a government access within an active conflict will produce valuable operational evidence, but it also raises questions about access controls, audit trails and how providers separate legitimate defensive work from offensive use.
Britain may put AI assistants on search choice screens
The UK Competition and Markets Authority strengthened proposals for Google search choice screens, adding eligible AI assistants to the services users could select on Android and Chrome. The proposal would show a choice during device or browser setup and again annually. Providers would also need to attribute publisher content clearly so users can reach original sources. The consultation runs until October 9, with a final decision expected by year-end.
This treats conversational assistants as search competitors. It could improve distribution for ChatGPT and Perplexity while forcing regulators to define the technical, security and attribution standards for default placement.
Europe funds AI-assisted chip design
The Netherlands' NADI and Germany's SPRIND launched a 40 million euro challenge for AI-native chip design, Reuters reported. Over 20 months, small teams will use AI to accelerate the design of chips for training and inference. The program combines industrial policy with a test of whether generative systems can shorten one of the semiconductor industry's slowest and most expensive development cycles.
The funding is modest beside the cost of a leading fabrication plant, but design productivity is a different bottleneck. If teams can verify that AI-generated layouts and components meet performance and reliability requirements, Europe could gain leverage without first matching the capital scale of the largest chip manufacturers.
Data center expansion meets broader public resistance
A Pew Research Center survey found that 54% of Americans now view data centers as mostly bad for the environment, up from 39% in January. Half see them as mostly bad for home energy costs, and 60% would be uncomfortable with a new data center in their area. Awareness has also grown, with 88% saying they have heard at least a little about data centers.
Infrastructure is becoming a political constraint on AI growth. Companies planning large clusters will need credible community benefits and transparent accounting for power, water and local costs.
Ema raises $77 million for enterprise agents
Enterprise agent startup Ema raised a $77 million Series B led by Creaegis, bringing total funding to $140 million, TechCrunch reported. Ema uses coordinated agents to automate work across HR, IT and finance. Existing investors Accel, Section 32 and Prosus also participated.
The round shows that investors still see a path for agent platforms to displace parts of traditional enterprise software and services. The harder proof will be sustained adoption: companies need reliable task completion, permissions, monitoring and accountability across systems that were not designed for autonomous action.
What matters today
Efficiency is becoming a frontier capability in its own right. OpenAI and Anthropic are competing on the cost and speed of long-running work, while regulators, governments and investors push assistants into search, cybersecurity, chip design and enterprise operations. The limiting factors are moving outward from model quality toward infrastructure, distribution and control. Providers that lower costs without improving oversight may accelerate adoption faster than institutions can absorb the risks.