Since the 2020 OpenAI scaling laws paper and the 2022 Chinchilla update, the field has treated the relationship between parameters, data, and loss as a near-physical law. In 2025 several labs began reporting that pretraining gains were flattening, prompting a wave of essays declaring the end of scaling. The picture on the ground is more nuanced.
What actually flattened
The plateau that researchers describe is real, but narrow. Pretraining loss on generic web text has been improving predictably, yet the downstream benchmarks that matter — coding, mathematical reasoning, long-horizon planning — no longer track that loss curve cleanly. Adding tokens of ordinary CommonCrawl to a base model past a certain point yields improvements that are measurable in perplexity and invisible in user-facing quality.
This is not a failure of scaling; it is a failure of a specific input. The frontier has moved from raw token count to the composition of the training mix. Labs now invest heavily in synthetic data pipelines, verifier-graded reasoning traces, and curated code corpora, all of which continue to produce clean log-linear improvements when scaled.
The new axes
Three axes are absorbing most of the compute budget that used to fund larger pretraining runs. The first is post-training: reinforcement learning from verifiable rewards, distillation from stronger teacher models, and preference optimization on domain-specific tasks. The second is test-time compute — letting the model think longer, sample many candidates, and self-verify before returning an answer. The third is orchestration: composing multiple specialized models into agentic systems that plan, call tools, and critique intermediate outputs.
- Post-training now accounts for 20–40% of total compute at leading labs, up from roughly 5% two years ago.
- Test-time compute scaling shows clean power-law improvements on reasoning benchmarks well beyond the point where pretraining alone stalls.
- Multi-model orchestration is producing better accuracy per dollar than a single larger model on many enterprise workloads.
“The bitter lesson has not been repealed. It has been generalized. Compute still wins — we are just learning where to spend it.”
Why the narrative went wrong
Public discourse latched onto a leaked chart from one lab and generalized it. But labs are not running the same experiment. Some invest their next order of magnitude in larger dense models; others spend it on mixtures of experts, longer context, or reasoning-focused post-training. Averaging across those choices produces a misleading picture of the frontier.
The framing also confuses research plateaus with product plateaus. A model that adds three points on a saturated benchmark can still transform a downstream workflow if latency drops by half or tool use becomes reliable. Those wins do not show up in the loss curve at all.
What to watch next
The most interesting frontier right now is not the size of the next base model but the interaction between reasoning and retrieval. Systems that can plan a research task, retrieve fresh evidence, and revise their plan produce qualitative improvements that no benchmark cleanly captures. Expect the next round of eye-catching results to look more like scientific assistants than chatbots.
For engineering leaders, the takeaway is not to bet on a single scaling story. Budgets that assume linear model-size progress will overpay; budgets that assume progress has stopped will be blindsided by the next agentic release.
Key Topics
Extended Knowledge
- Chinchilla (2022) reset the compute-optimal ratio between parameters and training tokens, and much of what came after has been variations on that theme.
- Reinforcement learning from verifiable rewards (RLVR) is emerging as the dominant post-training paradigm for reasoning-heavy tasks such as math and code.
- OpenAI's o-series and comparable reasoning-tuned models from other labs are the clearest public evidence that test-time compute scales cleanly.
Frequently Asked
Pretraining on generic web text has diminishing returns, but overall system capability is not. Reasoning-focused post-training and test-time compute continue to produce clean improvements at additional compute.
Smaller models are catching up on specific benchmarks, especially when distilled from a stronger teacher, but the frontier of open-ended reasoning still favors labs willing to spend heavily on post-training.
Assume that model families will refresh every 3–6 months and design product architecture so the underlying model can be swapped. Invest in evaluations that measure your workflow, not generic leaderboard scores.
