Reasoning effort is becoming a dial, not a default
Several new studies and one new model converge on the same point: running every request at maximum reasoning effort wastes tokens, and lower effort often costs much less with little lost.
- In a preprint study of seven frontier models across eight benchmarks, 22 of 56 model-benchmark combinations matched their best performance at a lower reasoning effort and lower cost. Self-reported m:…8216
- In a preprint, a method called LoCo trains reasoning models to stop deliberating once they have enough information, cutting token counts while holding performance in mathematics and code generation. Self-reported m:…6306
- Aleph Alpha's Kolibri model lets users select reasoning effort, adjusting its derivation length to prompt difficulty; lower settings retain most performance while using far fewer tokens. Self-reported m:…6316
- In a preprint, researchers argue long context is often unnecessary because frequent context compaction preserves accuracy, and note the cost disparity between later and earlier tokens can rise to as much as 7 times on frontier open-weight models. Self-reported m:…9460
- In a preprint, the Route Anything framework learned routing from an omni-modal trajectory matrix of 374,401 runs and outperformed fixed models in video generation and robotics while improving efficiency. Self-reported m:…2002
The expensive setting is the default
Vals AI tested seven frontier models across eight benchmarks to measure what reasoning effort actually buys. In 22 of the 56 model-benchmark combinations, a lower effort setting matched the best score at a lower cost. At their lowest setting, four of the seven models kept at least 80% of their score for half the cost or less, which means many teams are paying for reasoning they do not need. m:…8216
The practical conclusion is that effort should be set per task rather than left at maximum. The analysis suggests that tuning effort to the request can improve efficiency without giving up score, a shift from treating the strongest setting as the safe choice to treating it as one option among several. m:…8216
A model that lets you choose
Aleph Alpha's new Kolibri model builds that choice into the model itself. It offers user-selectable reasoning effort, so the same model can spend more time on a hard prompt and less on an easy one. It was trained on post-hoc annotated reasoning traces, which taught it to adjust the length of its internal derivation to the difficulty of the request. m:…6316
At low effort, the model retains 96% of its high-effort score while using 35% of the reasoning length, and the released model retains 98% while using 76% of completion tokens. That trade matters twice over: fewer tokens mean lower cost and lower latency, so the dial controls both the bill and the wait. m:…6316
Learning to stop sooner
A second line of work, described in a preprint, attacks the same waste from inside the reasoning process. LoCo uses local preference learning: the model trains on self-generated mathematical traces and learns to detect overthinking states from its own traces, stopping once further deliberation adds little. Completion length fell by 41.7% to 45.4% on mathematics benchmarks and 23.4% to 36.3% on code benchmarks, with performance maintained. m:…6306
The method also transferred across domains, showing efficiency gains in both mathematics and code generation tasks. That transferability matters because it suggests the stopping skill is general, not a trick tuned to one benchmark. m:…6306
Another preprint makes a related case about context rather than reasoning steps. Its authors argue that long-context models are often unnecessary, because compacting context frequently preserves accuracy on long-horizon and long-input tasks. They note that generating later tokens is far more expensive than earlier ones, with the cost disparity between later and earlier tokens rising to as much as 7 times on frontier open-weight models, and they propose semantic state space models as a more efficient alternative for long-horizon software engineering work. m:…9460
Routing replaces the single setting
The third piece is routing: instead of one model at one setting, a system picks the right model and policy for each task. In a preprint, the Route Anything framework does this with learned routing trained on an omni-modal trajectory matrix of 374,401 runs. The researchers report that learned routing can improve performance while increasing efficiency, and that it outperformed fixed models in domains including video generation and robotics while reducing policy time. m:…2002
Taken together, the three lines of work point in one direction. Effort is becoming a control surface: a dial on the model, a stopping rule inside reasoning, and a router above it all. The teams that treat effort as something to tune, rather than something to max out, are the ones that will keep both quality and cost under control as agent workloads grow. m:…8216 m:…6316 m:…2002
What is still unresolved
- Which task types genuinely need maximum reasoning effort, and which never do?
- Does frequent context compaction hold up on very long software engineering tasks in production, not just in experiments?
- How well does learned routing generalize to task types outside the 374,401-run trajectory matrix?
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.