OpenAI's o-series and DeepSeek's R1 showed that training models to reason with RL against checkable outcomes (math, code, tool results) sharply improves multi-step reliability. Anthropic pursues the same direction for agentic coding, optimizing for tasks where success can be automatically verified rather than judged by preference alone.

Cognition AI's Devin, Manus, and Genspark wrap frontier models in explicit planning loops, scratchpads, and persistent memory so the agent can decompose goals and resume after failures. The bet is that orchestration and state management around the model recover much of the reliability the base model lacks.

Adding a verification pass — the model or a second model checking each step before proceeding — reduces error accumulation. This is increasingly built into both lab APIs (Anthropic, OpenAI) and agent platforms, trading extra inference cost for fewer cascading mistakes.

Contextual AI and Glean ground answers in a customer's own documents with permissions-aware retrieval, so responses cite sources the user can check. The thesis is that RAG turns an unreliable generator into a defensible system of record when the retrieval layer is good enough.

Perplexity built its product around inline citations, forcing every claim back to a retrieved source and making hallucinations visible to the user. The approach trades some fluency for auditability, which is exactly what knowledge-work buyers want.

Anthropic and Google DeepMind invest in training models to express calibrated uncertainty and decline when evidence is thin. Getting a model to reliably say 'I don't know' is as hard as getting it to be right, and it is essential for high-stakes vertical agents like Harvey and Nabla.

Mistral AI and DeepSeek use sparse mixture-of-experts so only a fraction of parameters activate per token, delivering large-model quality at a fraction of the inference cost. This is the dominant lever behind the recent collapse in cost-per-capability.

Liquid AI builds compact, non-transformer foundation models designed to run efficiently on edge and constrained hardware, while most labs now ship distilled 'mini' variants of their frontier models. The bet is that most production tasks don't need the biggest model.

Lower-precision serving, speculative decoding, and hardware-specific inference kernels squeeze more throughput from each GPU. The labs and a wave of inference-infrastructure companies treat the serving stack itself as a source of durable cost advantage.

Ndea, founded by ARC-AGI creator François Chollet, pursues benchmarks that test fluid reasoning on novel problems rather than memorized knowledge — explicitly designed to be hard to game by scaling alone. The ARC family has become a reference point for whether scaling buys genuine generalization.

Evaluations that score an agent on completing real software, research, or computer-use tasks end-to-end (SWE-bench-style suites) are the metric coding agents like Cognition AI optimize against. They measure whether the whole system works, not just whether the model answers a question.

Large-scale human preference arenas and adversarial evaluation remain the backstop for capabilities benchmarks can't capture. Anthropic and OpenAI invest heavily in red-team pipelines to surface failure modes that static benchmarks miss entirely.

Anthropic pioneered Constitutional AI — training models against an explicit set of principles with AI-generated feedback — to scale oversight beyond human labeling. It is the leading production technique for steering model behavior, though its limits at higher capability are openly debated.

Anthropic and Google DeepMind invest in reverse-engineering the internal computations of models to detect deception or unsafe reasoning before deployment. The wager is that you cannot control what you cannot inspect, making interpretability a prerequisite for safe frontier systems.

Safe Superintelligence was founded on the premise that safety and capability must be advanced together from the start rather than bolted on, raising on that thesis with no commercial product. It is the most concentrated bet that alignment is the binding constraint on the whole field.

Anthropic and OpenAI ship models trained to perceive a screen and control a mouse and keyboard directly, generalizing across applications without per-app integration. Reliability on unfamiliar interfaces and safe handling of consequential actions remain the limiting factors.

Genspark, Manus, and Perplexity's Comet browser build structured environments that constrain and assist the agent — exposing tools, capturing state, and recovering from errors — rather than relying on raw pixel control. The harness compensates for the model's gaps.

Where APIs exist, agents call typed tools rather than clicking UIs, which is far more reliable. Standardized tool interfaces (function calling and protocols like MCP, championed by Anthropic) are turning ad-hoc integrations into a composable layer agents can reason over.

World Labs, founded by Fei-Fei Li, builds foundation models that generate explorable, geometrically consistent 3D worlds from minimal input — treating spatial intelligence as a foundational layer for downstream embodied and spatial applications.

Decart's Oasis generates interactive worlds frame by frame in real time, pushing toward world models you can act inside rather than just render. The challenge is maintaining coherence and physics as the user interacts over long horizons.

Sakana AI explores nature-inspired and simulation-heavy methods to build understanding of dynamics more sample-efficiently than brute-force video scaling. The bet is that smarter model design, not only larger datasets, is the path to usable world models.