Why Cutting DeepSeek's Reasoning Tokens Crashes Accuracy by 68% (And Which Models Survive)

CFOs are demanding cuts to inference costs, but before you throttle your reasoning model's token budget, you need to understand the brutal trade-off: some models collapse in accuracy while others survive. A new systematic study of four frontier models—including DeepSeek-V3.2—reveals that truncated chain-of-thought (CoT) reasoning can slash compute expenses but may also destroy performance, and the impact varies wildly across model families.

The business question is immediate: Can we safely reduce the length of reasoning traces without sacrificing the intelligence we pay for? The answer, for DeepSeek-V3.2, is a resounding no.

What "Reasoning Tokens" Actually Are—And Why They Cost Money

Chain-of-thought is the technique where large language models generate intermediate reasoning steps before producing a final answer. Instead of jumping straight to a solution, the model writes out its logic: "First, calculate the interest. Then, add the principal. Finally, subtract fees." This extended trace improves accuracy on complex tasks like mathematics, coding, and multi-step analysis. However, each reasoning token consumes GPU time and memory, translating directly into higher inference costs.

Enterprises deploying reasoning-specialized models such as DeepSeek-V3.2, OpenAI's GPT-5.1, or Grok 4.1 face a tempting lever: limit the number of reasoning tokens allowed per query. Shorter chains mean cheaper inference. The assumption has been that as long as the model produces some reasoning, performance will degrade gracefully as budgets shrink. The new study turns that assumption on its head.

The researchers—from independent academic teams—subjected four frontier models to a rigorous token budget ablation. They forced each model to reason exclusively through natural language, code, comments, or both, then systematically reduced the token allocation to 10%, 30%, 50%, and 70% of what the model would ideally use. Performance was measured on AIME, GSM8K, and HMMT mathematical benchmarks.

The Catastrophic Drop: DeepSeek-V3.2 at 50% Budget

The headline finding: DeepSeek-V3.2's accuracy plummeted from 53% with full reasoning to just 17% when its CoT tokens were cut to half the optimal length. That is a 68% relative drop in correctness—effectively rendering the model useless for high-stakes decisions.

Equally striking: the model's performance with no reasoning at all was 53%, identical to the full-reasoning scenario. In other words, giving DeepSeek-V3.2 an incomplete chain of thought is worse than giving it none. Truncated reasoning actively misleads the model, leading it to confidently wrong conclusions.

This pattern flips entirely for Grok 4.1. At 30% of optimal token budget, Grok maintained 80–90% of its peak accuracy. While all models suffered some degradation, Grok demonstrated exceptional robustness to budget cuts, while OpenAI's GPT-5.1 and DeepSeek-V3.2 collapsed to 7–27% in the same regime.

The study uncovered another nuance: reasoning in code degrades more gracefully than reasoning in natural language comments. Gemini's comment-based reasoning collapsed to 0% under budget pressure, while code-based reasoning preserved 43–47% accuracy. Hybrid reasoning—mixing code and text—underperformed both single-modality approaches.

What Your Competitors Are Doing With This

The instinct to trim token budgets is widespread. In Q1 earnings calls, dozens of CIOs mentioned "optimizing inference spend" as a top priority. But few have validated whether their specific model tolerates truncation. That gap creates a two-tier landscape:

• Reactive players are applying uniform token limits across all models. They will see cost savings but may unknowingly sacrifice accuracy, especially when DeepSeek-V3.2 is in production. The first major failure—a mispriced financial calculation, a flawed code generation, a wrong medical inference—will trigger an expensive rollback and reputational hit.

• Strategic players are either keeping full reasoning budgets for DeepSeek-V3.2 and experimenting with Grok for budget-sensitive tasks, or they are abandoning reasoning-specialized models altogether for distilled variants like DeepSeek-R1-Distill, which likely behave differently under truncation (though that must be tested). A few advanced teams are running their own ablation studies before tightening budgets, using this very paper as a benchmark.

• Smart procurement teams are now demanding token robustness data from vendors as part of RFPs. They ask: "How does your model perform when reasoning tokens are reduced to 50% or 30% of optimal?" A vendor that cannot answer is a vendor that cannot be trusted with production workloads.

What This Means for Your AI Procurement Decision

The decision framework is now clear. If DeepSeek-V3.2 is in your fleet—or if you are evaluating it—take these steps:

1. Do not truncate. Keep the model's token budget at or near its preferred length until you have domain-specific evidence that shorter reasoning works. The default assumption must be that DeepSeek-V3.2 will break.

2. If you must reduce costs, consider switching to a model that degrades more gracefully under token constraints, such as Grok 4.1, or evaluate whether a non-reasoning configuration (no CoT) meets your needs without the misleading intermediate steps.

3. Test, don't guess. Run your own ablation on a representative sample of your actual tasks before rolling out token limits. Use the methodology from this study as a template: vary the budget, measure accuracy on your key metrics, and find the knee of the curve.

4. Include robustness in vendor scoring. When comparing models, factor in how each behaves under token pressure. A model with higher baseline accuracy but catastrophic truncation may be a worse long-term bet than a slightly less accurate but more robust model.

The window to act is narrow. Many enterprises are already tightening token budgets in response to CFO压力; the failures will surface within weeks. Those who understand the asymmetric risks—and adapt their deployments accordingly—will avoid costly outages and stay ahead of competitors who learn the hard way.

Your move: audit your reasoning model configurations today. If DeepSeek-V3.2 is in use, verify that its token budget has not been cut. If it has, measure the accuracy impact immediately. The cost of a broken chain is far higher than the savings.

Source: "Broken Chains: The Cost of Incomplete Reasoning in LLMs" (arXiv:2602.14444v1)