Artificial intelligence performance is increasingly determined by far more than the underlying model.
OpenAI has shown that two changes to an agent’s memory architecture can triple its score on an intelligence benchmark while using six times fewer output tokens. The company has also detailed how GPT‑5.6 helped optimise the infrastructure used to serve itself. Meanwhile, the US is restricting Chinese humanoid robots over national-security concerns, Robo.ai has appointed former INTERPOL president Ahmed Naser Al-Raisi to lead Alif Holding, and myCOI’s illumend is applying explainable AI to insurance compliance.
The common thread is infrastructure. Whether the subject is context management, GPU efficiency, physical robots, sovereign industrial systems or insurance records, AI’s practical value increasingly depends on the systems built around it.
Two settings triple OpenAI’s ARC-AGI-3 score
OpenAI has demonstrated that an AI benchmark may measure the surrounding evaluation system almost as much as it measures the model itself.
GPT‑5.6 Sol initially performed poorly on ARC-AGI-3, a benchmark designed to test how effectively an agent can learn the rules of unfamiliar two-dimensional games through exploration.
The model scored only 7.8% in an initial evaluation, despite having performed strongly on other long-horizon reasoning challenges. GPT‑5.5 scored just 0.4%.
OpenAI investigated the discrepancy and identified two limitations in the benchmark harness. After every action, the system discarded the model’s private reasoning. The model retained a record of earlier moves and short notes, but lost the plans and insights that produced those decisions.
The harness also used rolling truncation. Once the conversation exceeded its context limit, earlier actions were removed entirely. The model was therefore losing both its previous reasoning and parts of its interaction history.
According to OpenAI, enabling retained reasoning and replacing rolling truncation with compaction increased GPT‑5.6 Sol’s public-set score from 13.3% to 38.3%. The changes also reduced output-token consumption by approximately six times.
For comparison, OpenAI estimated that the average human tester scored around 48% under the benchmark’s Relative Human Action Efficiency metric.
Retained reasoning allows the agent to preserve useful internal context across turns and tool calls. Instead of interpreting the game from the beginning after every action, it can continue from its earlier conclusions.
Compaction addresses context limits by summarising and preserving important information instead of simply deleting the oldest messages. This allows an agent to retain earlier discoveries throughout a long-running task.
The experiment produced a striking result in one of the benchmark games. No frontier model using the official harness progressed beyond the first level, while GPT‑5.6 Sol using OpenAI’s Responses API harness completed all six levels.
The broader lesson extends well beyond puzzle games. AI agents responsible for research, coding, operations or customer workflows need persistent knowledge of what they have attempted, what they learned and why previous strategies failed.
An agent without that continuity may repeat investigations, contradict earlier decisions or spend excessive tokens reconstructing its own history.
Developers building long-running agents should therefore treat memory architecture as a core product decision. They need to decide:
- Which information should persist between actions?
- When should detailed context be compressed?
- Which observations must remain available exactly as recorded?
- How should incorrect conclusions be revised?
- How can retained context be inspected without exposing protected internal reasoning?
- When should an agent begin a task with a clean state?
OpenAI recommends using the Responses API, retaining reasoning and enabling compaction for GPT‑5.6 deployments. The company also argues that model comparisons should use configurations that reflect how the systems operate in production.
The result does not mean every model should receive a custom evaluation designed to maximise its score. Benchmark fairness still requires consistent and transparent conditions. It does mean that an evaluation using a generic harness may underestimate a model whose capabilities depend on memory and context features absent from that harness.
HIPTHER previously examined the operational side of agent memory and control through its coverage of OpenBox AI and Temporal’s runtime governance for long-running agents.
GPT‑5.6 helps OpenAI optimise its own infrastructure
OpenAI has also published an engineering account of how GPT‑5.6 Sol is being used to improve the systems that deliver its own inference.
Model efficiency is often discussed in terms of token prices or benchmark scores. The cost of operating an AI service, however, also depends on request routing, hardware utilisation, data movement, caching, scheduling and the software kernels running mathematical operations on GPUs.
In its engineering report, OpenAI says GPT‑5.6 Sol with maximum reasoning outperforms Claude Fable 5 on the Artificial Analysis Coding Agent Index at less than half the cost. The company also positions Terra as delivering performance comparable with GPT‑5.5 at half the price, while Luna costs 80% less than Sol.
These are OpenAI’s own comparisons and should be assessed alongside independent evaluations and real-world workloads. Nevertheless, the technical work behind the efficiency claims is significant.
OpenAI used GPT‑5.6 Sol through Codex to analyse production traffic and identify imbalances in how requests were distributed between regions, clusters, model instances and computing resources.
A data centre may possess substantial accelerator capacity while still operating inefficiently if some machines are overloaded and others remain underused. Routing decisions must account for geography, available hardware, context length, cache availability and the characteristics of individual requests.
GPT‑5.6 Sol reportedly helped engineers test new routing strategies and refine those decisions.
The model was also used to improve the forward pass—the sequence of computations that turns model input into predictions. It identified operations that could be precomputed, avoided or processed in parallel.
Through Codex, GPT‑5.6 Sol autonomously rewrote and optimised production kernels using the Triton and Gluon GPU-programming languages. OpenAI says these and related improvements reduced end-to-end serving costs by 20%.
The company used verification tooling, including its open-source Floating-Point Sanitizer, to check the correctness of AI-generated kernels. This validation is essential because a faster kernel that introduces subtle numerical errors could damage model quality or reliability.
GPT‑5.6 Sol also improved its speculative-decoding system. Speculative decoding uses a smaller model to propose several output tokens, which the primary model can verify together. Accepted proposals reduce the number of expensive sequential operations required from the larger model.
OpenAI says the resulting changes improved token-generation efficiency by more than 15%.
The agentic harness used by Codex and ChatGPT Work contributes further savings. Deferred tool discovery prevents every available tool, skill and integration from occupying the context window at all times. Tool output is capped by default, while deterministic ordering and append-only history improve prompt-cache reuse.
These optimisations matter because an agent may make dozens of model and tool requests within a single user task. A small inefficiency repeated 30 times can become a noticeable delay and cost.
The most consequential aspect of the report is recursive improvement: an AI model is helping optimise the software and infrastructure used to operate that model family.
This does not amount to autonomous self-improvement without human control. Engineers still define objectives, review changes, test correctness and decide what reaches production. It does show how AI-assisted optimisation can accelerate development across a highly specialised technical stack.
HIPTHER has explored a related direction through Fractal’s PiEvolve agent for autonomous machine-learning and scientific discovery.
US restrictions turn humanoid robots into a geopolitical issue
The United States has restricted new imports and sales of certain foreign-made humanoid and quadruped robots, with the measures primarily targeting products from China.
The Federal Communications Commission has added advanced robotic equipment to its Covered List, which identifies communications products and services considered to present national-security risks.
According to CNBC, the decision comes as the Trump administration prepares for further engagement with China and seeks to reduce US dependence on Chinese technology supply chains.
The restrictions also cover certain foreign-manufactured connected power inverters used in solar installations, data centres and other energy infrastructure.
Humanoid robots present a distinctive security challenge because they combine multiple technologies in one mobile physical system:
- Cameras and microphones capable of collecting environmental data.
- Network connections for cloud services and remote management.
- AI systems interpreting people, objects and physical spaces.
- Mechanical components capable of interacting with equipment.
- Software-update channels that may alter behaviour after deployment.
- Data flows that may cross borders or depend on foreign infrastructure.
A compromised connected robot could create consequences beyond an ordinary data breach. Depending on its environment and capabilities, it might collect sensitive information, interfere with industrial processes or cause physical disruption.
Supporters of the US measures argue that early restrictions can prevent strategically sensitive infrastructure from becoming dependent on equipment that cannot be fully inspected or controlled.
Critics warn that the policy may hurt US researchers and startups that rely on relatively affordable Chinese robotic hardware. China possesses extensive manufacturing capacity and has become a leading supplier of humanoid and quadruped systems. Restricting access may raise development costs before a comparable US or allied supply chain exists.
The measure could therefore protect domestic manufacturers while slowing adopters that need physical platforms for testing AI software.
Existing authorised products are not automatically removed by the decision, but the FCC retains substantial power over future approvals. China has described the restrictions as protectionist and has pledged to defend the interests of its companies.
The dispute illustrates how embodied AI is becoming part of the same geopolitical contest already affecting semiconductors, telecommunications equipment, drones and cloud infrastructure.
The central policy question is not simply where a robot was manufactured. Regulators need technical standards covering software transparency, secure updates, data residency, remote access, component provenance and vulnerability disclosure.
Without such standards, country-based bans may reduce one category of exposure while leaving similar risks in domestically produced equipment.
HIPTHER’s earlier coverage of the global AI robotics market and competition between the United States and China shows how a once-commercial technology race has evolved into a strategic-security issue.
Robo.ai appoints Ahmed Naser Al-Raisi to chair Alif Holding
Robo.ai has appointed former INTERPOL president H.E. Dr Ahmed Naser Al-Raisi as chairman of its proposed subsidiary Alif Holding.
Alif Holding is being established in Abu Dhabi as an intelligent industrial-technology group serving government, public-sector and enterprise customers.
According to the company announcement, the business will operate through two integrated platforms covering intelligent software and intelligent equipment.
Its proposed activities span energy, oil and gas, ports, utilities, mining, transportation, public safety, smart cities and critical infrastructure. Alif Holding says its systems will be developed, integrated and manufactured in the UAE under a technology-neutral and open-architecture approach.
The company intends to serve the UAE and wider Gulf Cooperation Council region before expanding into global markets. Its strategy is aligned with the UAE National Strategy for Artificial Intelligence 2031.
Al-Raisi served as president of INTERPOL between 2021 and 2025 and currently chairs Neurovia AI, another Robo.ai subsidiary. He brings experience in public-sector digital transformation, data compliance, security governance and cross-border digital risk.
At Alif Holding, he will chair the board and provide oversight of corporate governance, long-term strategy, security, risk and compliance.
The appointment reflects a broader shift in the AI market. Companies selling systems to governments and critical-infrastructure operators cannot rely solely on technical performance. Buyers require assurances covering data sovereignty, supply-chain security, operational continuity and accountability.
“Made in UAE” manufacturing may also help Alif Holding respond to governments seeking greater control over strategically important technology. Sovereign AI is increasingly understood to include not only locally hosted models, but also control over data, hardware, software integration and industrial production.
That ambition creates demanding responsibilities. AI deployed in public safety, transportation or utilities must meet higher standards than a typical commercial application. Systems need independent testing, access controls, incident-response procedures, human oversight and clear limits on surveillance or automated decision-making.
Open architecture can reduce dependence on one technology provider, but it must be combined with strict integration standards. A system assembled from multiple models and devices can create new security gaps if responsibility is divided between vendors.
Alif Holding’s credibility will therefore depend on how its governance promises are translated into verifiable controls, procurement requirements and transparent operational standards.
illumend wins Stevie Award for insurance-compliance AI
Illumend, the AI-native insurance-compliance platform developed by myCOI, has received a Bronze Stevie Award in the Domain Specific AI Platform of the Year category.
The platform addresses a deceptively complex problem: determining whether contractors, tenants, vendors and other third parties carry insurance that genuinely meets an organisation’s contractual requirements.
Collecting a Certificate of Insurance does not necessarily confirm adequate protection. A document may be expired, incomplete, inconsistent with an underlying policy or missing coverage required by a contract.
According to the award announcement, illumend analyses insurance records, compares coverage with organisational requirements and flags deficiencies before they become claims, delays or disputes.
Its capabilities include:
- Converting unstructured insurance documents into structured information.
- Comparing coverage with contractual requirements.
- Detecting potential underinsurance and coverage gaps.
- Monitoring document expiration and renewal.
- Managing communication and issue resolution.
- Maintaining audit-ready records of compliance activity.
The platform’s conversational AI guide, Lumie, explains insurance requirements and identified problems in plain language. It also recommends steps users can take to address a deficiency.
Illumend is built on myCOI’s 16 years of insurance-compliance experience. The company says that foundation includes more than 45 million processed documents, 1.2 million managed agreements, 750,000 cleared third-party partners and two million coverage gaps identified before claims.
This historical information can give a specialised AI system more useful context than a general-purpose model. Domain-specific terminology, document structures and established review practices can be embedded directly into its workflows.
The platform also demonstrates an important form of explainable AI. A compliance tool should not simply label coverage as acceptable or deficient. Users need to understand which requirement was tested, which document supplied the evidence and why the system reached its conclusion.
Human review remains necessary for ambiguous wording, unusual contractual conditions and high-value risks. Insurance policies are legal documents, and apparently small differences in definitions, exclusions or endorsements can materially change coverage.
Illumend’s strongest role is therefore to automate document-heavy work while directing professional attention towards the cases that require judgement.
The bigger picture: the system around the model is becoming the advantage
These five stories challenge the assumption that progress in AI is mainly a contest to build a more capable model.
OpenAI’s ARC-AGI-3 experiment shows that memory and context management can transform performance without changing the model. Its infrastructure report demonstrates how routing, kernels, caching and agent orchestration determine whether frontier intelligence can be delivered economically.
The humanoid-robot dispute moves the argument into hardware and geopolitics. Robo.ai’s Alif Holding initiative places governance, industrial production and data sovereignty at the centre of public-sector AI. Illumend shows how specialised data and explainable workflows can make AI useful in an established compliance function.
Models remain important, but they are becoming one component in a larger system.
The organisations that produce durable value from AI will be those that combine capable models with persistent context, efficient infrastructure, secure hardware, reliable governance and domain-specific knowledge.
The next competitive frontier is not simply who has the smartest model. It is who can build the most trustworthy and effective system around it.
