Reasoning models need four traits: skills, calibration, strategy, and abstraction.
2
Planning must become a model-native capability so agents can choose directions, break down tasks, manage effort, and use other models.
3
Scaling reinforcement learning may bring post-training compute closer to pretraining compute than many expected.
Summary
Nathan Lambert argues that reasoning models have moved beyond a stage where higher benchmark scores are the main goal. Math, code, and tool use are improving, but models still struggle with medium and long time-horizon tasks. He proposes four traits for the next generation: skills, calibration, strategy, and abstraction. The first two are already receiving substantial attention, while strategy and abstraction describe the planning abilities agents need to choose a direction, recover from mistakes, divide work into tractable subtasks, and delegate difficult parts. Lambert also discusses the cost of overthinking, the need for model-native planning, and the role of parallel model calls. He expects progress to depend on data and stable reinforcement-learning infrastructure, then suggests that post-training compute could approach parity with pretraining compute as long-horizon reinforcement learning scales.
Reasoning models are opening applications beyond math and code
Lambert describes reasoning as a source of new language-model applications. He uses OpenAI's o3 finding a forgotten search term and returning a direct download link as an example. Deep Research can inspect a website for typos or restrict its research to material on that site. Claude Code helps him build small projects, while Codex points toward more autonomous agents. He expects these tools to become part of everyday work as their capabilities improve. The common cause is the step change from reasoning-model training, which has extended the time horizon of tasks models can complete.
The next frontier requires planning beyond learned skills
Lambert says gains in AI do not arrive automatically as benchmark scores rise. Human effort is still needed to decide what models must learn next. His research question is how to train reasoning models that can work autonomously and have meaningful planning behavior. He separates skills such as math and code from planning abilities. A model can be highly capable at a local task and still fail to decide what to do across a longer task. This distinction matters for the language-model products and agent systems that are beginning to appear.
The taxonomy has four traits, with strategy and abstraction handling planning
Lambert proposes skills, calibration, strategy, and abstraction. Skills cover abilities such as math and code. Calibration means matching output-token effort to problem difficulty, since reasoning models can spend hundreds or thousands of tokens on a simple question. Strategy means moving in the right direction and knowing which approaches to try, including the ability to change course. Abstraction means deciding how to divide a hard problem into tractable subtasks, or when to call a larger model. Lambert treats strategy and abstraction as parts of planning that current models do not reliably produce on their own.
Planning must become native to the model instead of a prompt template
Current agent applications often prompt models to make a plan at the beginning, as Deep Research and Claude Code do. Lambert argues that this behavior should become model-native. A model needs to decide how to manage memory, when to compress context, which parts of a task are tractable, and when to offload difficult work. It should also be able to call multiple models in parallel. Today, people approximate this by opening several terminal sessions or Claude Code windows. Lambert sees no fundamental reason that a language model could not learn to coordinate those calls itself.
Poor strategy can make reasoning too slow and too expensive
Lambert shows why calibration and strategy are connected. A reasoning model may begin a hard math problem by constructing a polynomial without first sketching the problem or checking whether the approach is promising. If it then spends 10,000 to 40,000 tokens going in the wrong direction, another factor of ten would create high cost and latency. He says many applications already expect responses within roughly one to 30 minutes, so wasted reasoning can cause users to leave. Simple questions create the opposite problem: models may use hundreds to a thousand tokens for a task whose answer could be one token.
Planning data can be easier to create than long reasoning traces
Lambert compares planning with the earlier development of reasoning models. OpenAI reportedly spent 12 to 18 months building initial reasoning traces for the system that became o1, including data for backtracking and verification. Planning may offer a more accessible training target. An expert can write a good five- to ten-step plan or review a plan produced by another model, even though writing a 10,000-token reasoning trace with backtracking is difficult. Lambert proposes starting with supervised fine-tuning on planning data, then testing whether reinforcement learning can improve planning on larger tasks. He also suggests structuring models so they plan before producing their full answer.
Parallel compute can improve reliability without solving every problem
Lambert distinguishes ordinary inference-time scaling from parallel compute. Inference-time scaling can make previously failed attempts succeed. Parallel compute can make answers more robust by encouraging exploration and then selecting or combining a well-developed result. He describes o1 Pro as a strong and robust model, while saying that parallel computation produces a different benefit from simply extending one chain of thought. It does not solve every problem, and the right choice depends on whether the task needs exploration, refinement, or another capability.
Stable data and infrastructure matter more than small recipe changes
Lambert's rough training plan starts with many questions that have verified answers across different domains, although available data is concentrated in math and code. Questions should be filtered by difficulty relative to the base model. Items solved zero out of 100 times or 100 out of 100 times waste compute and make reinforcement-learning updates noisier. The remaining questions need a stable reinforcement-learning run where performance keeps improving. Lambert says methods such as filtering, clipping, and resetting the reference model can add a few percentage points, but the larger requirements are data and stable infrastructure.
Post-training compute may approach pretraining compute
Lambert suggests renaming post-training as training because reinforcement learning is becoming a much larger part of model development. He cites OpenAI's claim that o3 increased post-training compute by ten times relative to o1, after o1 reportedly used about one percent of the compute used for pretraining. He also discusses estimates around DeepSeek, where the original DeepSeek-R1 paper reported 0.18 percent of compute on post-training in GPU hours, while a later account suggested the R1 training took a few weeks. Lambert labels the assumptions behind his estimate as strong and probably imperfect, but uses it to argue that scaling reinforcement learning is a real direction.
"And then as tasks become very hard, we need to do abstraction which is like the model has to choose on its own how to break down a problem into different things that it can do on its own."05:12
Who should watch
You are building agents that need to work across tasks lasting longer than a single model response, and you need a vocabulary for their failures.
You are deciding whether to spend effort on prompting, model training, memory management, or parallel calls in an agent system.
You work on post-training and want Lambert's view of the data, infrastructure, and compute needed to scale reinforcement learning.
