LLM trends

May made it clear that frontier competition is moving from chat quality toward models that can plan, use tools, write code, manage long-running tasks, and operate inside real workflows. Google introduced Gemini 3.5 Flash on May 19 as the first model in a new family built around “frontier intelligence with action,” with emphasis on agents, coding, and complex long-horizon tasks.

Google also launched Managed Agents in the Gemini API, letting developers spin up agents that reason, use tools, and execute code inside isolated cloud environments. Anthropic followed on May 28 with Claude Opus 4.8, highlighting stronger coding, agentic work, professional tasks, effort control, faster modes, and dynamic workflows in Claude Code. Microsoft also pushed this direction by making computer-using agents generally available in Copilot Studio: Google on Gemini 3.5, Google on Managed Agents, Anthropic on Claude Opus 4.8, Microsoft on computer-using agents.

OpenAI’s May 7 release of GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper showed that voice models are becoming more than speech interfaces. GPT-Realtime-2 is designed for live conversations where the model can reason, call tools, handle interruptions, recover from problems, and keep longer context for agentic sessions.

Google’s I/O updates pushed the same broader direction from a multimodal angle. Gemini Omni was introduced as a model family that can create from mixed inputs, starting with video, while Gemini 3.5 Flash was rolled into Search, Gemini apps, developer tools, and agent platforms. Together, the May releases point toward LLMs becoming realtime, multimodal work layers rather than only text-based assistants: OpenAI on realtime voice models, Google on Gemini Omni and I/O updates, Google on AI Search updates.

On May 20, OpenAI reported that an internal general-purpose reasoning model disproved a longstanding conjecture in the planar unit distance problem, a discrete geometry problem first posed by Paul Erdős in 1946. OpenAI said the result produced an infinite family of examples with a polynomial improvement over the long-assumed square-grid construction.

The significance was not only the mathematical result, but how it was found. OpenAI said the proof came from a general-purpose reasoning model, not a model specifically trained or scaffolded for that problem, and that the proof was checked by external mathematicians. For the LLM field, this made May a milestone month for evaluating whether advanced models can contribute directly to frontier research: OpenAI on the unit distance problem.

May also showed that model progress is increasingly tied to trust infrastructure. On May 19, OpenAI announced new provenance work using C2PA conformance, Google SynthID watermarking for images, and a public verification tool for checking whether images came from OpenAI systems. The update reflected a broader industry shift toward making AI-generated media easier to identify and verify.

Later in the month, OpenAI published its Frontier Governance Framework, mapping its safety and security practices to emerging legal requirements such as California’s Transparency in Frontier AI Act and the EU AI Act’s Code of Practice for General Purpose AI. This made governance less of a side issue and more of a deployment requirement for frontier systems: OpenAI on content provenance, OpenAI on the Frontier Governance Framework.

The business side of LLMs was just as important as the model releases. Reuters reported on May 28 that Anthropic raised $65 billion at a $965 billion post-money valuation, with demand for Claude creating compute and usage-limit pressure. That made the funding story inseparable from the infrastructure story: frontier AI companies need capital not only for research, but also for chips, cloud capacity, and reliable serving.

At the same time, Reuters reported that DeepSeek would make a 75% price cut on its flagship V4-Pro model permanent. This reinforced the other side of the economics story: as capable models become cheaper to run, competitive pressure shifts toward cost, latency, routing, and how efficiently products can turn model calls into useful work: Reuters on Anthropic’s funding and valuation, Reuters on DeepSeek’s price cut.

April showed that leading labs are pushing further into domain-specific models rather than relying only on general-purpose assistants. OpenAI introduced GPT-Rosalind on April 16 as a purpose-built reasoning model for life sciences research, focused on biology, drug discovery, and translational medicine.

This matters because it signals a broader shift in LLM product strategy: frontier capability is increasingly being packaged into models tuned for high-value workflows rather than presented only as a universal chat layer: OpenAI on GPT-Rosalind.

Anthropic’s April 16 release of Claude Opus 4.7 reinforced that the market is rewarding models that perform better in long-running, tool-using, multi-step workflows. The company positioned the model around stronger instruction following, lower tool error rates, and better performance in coding and agentic work.

That framing is important because it shows how frontier competition is shifting away from abstract benchmark headlines and toward whether a model can sustain real work over time with fewer failures: Anthropic on Claude Opus 4.7.

Google’s April rollout of Gemma 4 kept open-model pressure firmly in the conversation. Google described Gemma 4 as its most capable open model family to date and said it was built for advanced reasoning and agentic workflows.

That makes April another month where open models are not just a side story. They remain part of the main distribution and deployment contest, especially for developers who want lower-friction experimentation and customization: Google on Gemma 4.

April also made it clearer that model progress alone is not enough. Microsoft introduced its Agent Governance Toolkit on April 2, framing runtime security and policy enforcement as necessary parts of autonomous agent deployment.

At the same time, Reuters reporting in April highlighted the scale of infrastructure competition and pressure around LLM economics, including reported OpenAI spending on Cerebras-powered servers, Google’s work on new AI chips, and broader concerns about whether current AI business assumptions can hold under reliability and cost constraints: Microsoft on agent governance, Reuters on OpenAI and Cerebras, Reuters on Google and Marvell, Reuters on AI reliability and business-model pressure.

A major March development was the U.S. government’s move against Anthropic after a dispute over whether Claude’s safeguards should be relaxed for military use. Reuters reported that Anthropic sued on March 9 to block Pentagon blacklisting, after the government moved to designate the company as a national-security or supply-chain risk tied to restrictions on how its models could be used.

Reuters later reported that the Pentagon ordered a six-month phase-out of Anthropic tools, even as internal users resisted because Claude was already deeply integrated into military systems. This turned AI model policy into a procurement and national-security issue, with usage guardrails directly affecting whether a frontier model provider could remain inside sensitive government workflows: Reuters on the lawsuit, Reuters on the Pentagon phase-out.

OpenAI’s most relevant March move was not the late-February funding round, but its deeper entry into U.S. government work. Reuters reported on March 17 that OpenAI secured a deal to provide its models to U.S. defense and government agencies through Amazon Web Services, including both classified and unclassified work.

The report tied this expansion directly to the Pentagon’s break with Anthropic, making March an important distribution and government-access month for OpenAI. Reuters also reported on March 21 that OpenAI planned to nearly double its workforce to 8,000 by the end of 2026, underscoring how quickly the company is scaling product, engineering, and sales capacity during this period: Reuters on the AWS government deal, Reuters on OpenAI workforce expansion.

March reinforced a broader competitive trend coming out of China: fast iteration, lower-cost models, and stronger open-source pressure. Reuters reported on March 2 that MiniMax was pushing to become a global AI platform company after strong revenue growth, following the earlier release of its M2.5 open-source model.

Reuters also reported on March 23 that a U.S. advisory body warned China’s open-source dominance could threaten the U.S. AI lead. In parallel, Reuters had already noted in February that Alibaba’s Qwen 3.5 was being marketed as cheaper and more efficient for the “agentic AI era,” helping frame the cost and deployment pressure now carrying into March: Reuters on MiniMax, Reuters on U.S. warning over China’s open-source dominance, Reuters on Alibaba Qwen 3.5.

Microsoft introduced the AI Performance report in Bing Webmaster Tools as a public preview. It shows when your pages are cited as sources inside AI-generated answers across Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations.

The report includes metrics such as total citations, average cited pages, and “grounding queries,” making AI inclusion measurable inside a major webmaster platform. Coverage and breakdowns: Search Engine Land, Search Engine Journal.

As agents expand beyond chat, “connectivity” shifts from one-off integrations to shared conventions for how models discover tools, request actions, and operate safely across systems. This reduces duplicated connector work and makes agent deployments easier to standardize.

Model Context Protocol (MCP) moved further into mainstream enterprise discussion as connective infrastructure for agentic systems, including CIO-focused coverage explaining why it is on executive agendas and what it unlocks: CIO on MCP. Broader industry framing: The Verge on MCP.

Multi-step agents fail in ways that single-answer benchmarks miss. Evaluation starts to look more like production: tool calls, state changes, and end-to-end task success, not just “did the model say the right thing.”

February surfaced concrete examples. Agent-Diff (posted Feb 11, 2026) benchmarks agentic LLMs on enterprise API tasks using state-diff outcome evaluation. In healthcare, an npj Digital Medicine study on benchmarking LLM-based agent systems for clinical decision tasks was published Feb 18, 2026: npj Digital Medicine.

Tool-enabled agents create a bigger attack surface: indirect prompt injection, poisoned context, malicious “skills,” and compromised connectors. Security expectations shift toward practical guardrails like scoped permissions, tool confirmations, and monitoring of tool execution.

Recent reporting highlighted prompt injection risks in real agent workflows: The Verge on the Cline prompt injection incident. Supply chain risk in agent skills also surfaced in vendor research: Snyk on malicious agent skills. Microsoft security research also documented AI manipulation techniques such as recommendation poisoning: Microsoft Security Blog.

LLM usage is increasingly shifting away from isolated prompts toward agent-based workflows that combine planning, tool execution, memory, and validation. Instead of asking a model to complete a task in one step, systems now orchestrate multiple controlled interactions to reach a result.

This shift aligns with the growing emphasis on tool use and structured workflows across major research hubs such as OpenAI Research, Microsoft Research, and the broader academic record on arXiv.

LLMs are increasingly used as the first point of discovery for information-heavy tasks, including explanations, comparisons, and evaluations. Instead of navigating multiple sources, users rely on models to synthesize answers directly.

This direction is reinforced by ongoing AI integration into consumer and enterprise experiences, reflected in updates from Google AI and broader research agendas published via Microsoft Research.

As LLM usage scales beyond experimentation, cost awareness has become a core architectural constraint. Teams increasingly design systems that minimize unnecessary generation, selectively load context, and route tasks to models based on complexity.

Research and platform guidance increasingly foreground efficiency and disciplined deployment patterns, with relevant work visible through Anthropic Research and OpenAI Research.

Although models continue to support larger context windows, real-world usage shows a clear pattern: long context is invoked only when justified, rather than applied universally. Excessive context is increasingly seen as costly and sometimes counterproductive.

This view is reinforced by academic work accessible through arXiv and by enterprise deployment perspectives published via Microsoft Research.

By late 2025, open-ended chat interfaces were increasingly treated as secondary or fallback experiences, rather than the main way users interact with LLMs. Products began prioritizing embedded, task-specific interactions over generic chat boxes.

This shift was visible in how LLMs were integrated into productivity tools, developer environments, and internal systems. Research and product updates from major AI labs consistently emphasized contextual actions, structured outputs, and workflow-driven interactions rather than conversational depth alone. The trend reflected a growing understanding that chat is useful for exploration, but inefficient for repeatable work.

Context: ongoing work published across OpenAI Research, Microsoft Research, and arXiv.

A clear architectural preference emerged for retrieval-first systems, where LLMs operate on curated or indexed information rather than generating responses from parametric knowledge alone. This approach improved factual consistency and reduced cost and hallucination risk.

The trend was reinforced by research publications and platform guidance highlighting retrieval-augmented generation as a baseline pattern rather than an advanced optimization. By December 2025, treating retrieval as optional was increasingly seen as a design flaw in production systems.

Context: research and guidance surfaced via Anthropic Research, OpenAI Research, and arXiv.

Late 2025 saw increasing experimentation with model specialization, where different models were used for reasoning, planning, and execution instead of relying on a single general-purpose model. Systems began separating high-cost reasoning steps from lower-cost execution and formatting tasks.

This pattern appeared across research discussions and implementation guides, signaling a move toward modular AI systems. The trend suggested that future LLM stacks would resemble pipelines rather than monolithic model calls.

Context: examples and discussion across arXiv and lab research hubs such as OpenAI Research.

By December 2025, there were clear early signals that LLM-mediated visibility was beginning to diverge from traditional search ranking logic. Content that was frequently summarized, referenced, or cited by LLMs did not always align with top-ranking pages in search engines.

Research discussions and industry analysis increasingly focused on how models select sources, compress information, and attribute authority. While still emerging, the trend indicated that discoverability was becoming partially detached from classic link and ranking signals.

Context: see broader research threads via Google AI, Microsoft Research, and arXiv.

Prompt engineering began losing prominence as a standalone skill, replaced by system-level design involving orchestration, validation, memory, and error handling. Teams increasingly treated prompts as configuration details rather than core intellectual assets.

This shift was visible in both research framing and real-world tooling, where emphasis moved toward architecture patterns instead of handcrafted prompts. By the end of 2025, reliance on complex prompt chains without system safeguards was increasingly viewed as fragile.

Context: ongoing work reflected in OpenAI Research, Anthropic Research, and arXiv.