Anthropic recently removed roughly 80% of Claude Code’s system prompt for its newest models.
The reasoning is intuitive: as models get smarter, they need less direction, fewer constraints, and fewer examples. Past a certain capability level, examples stop helping and start boxing the model in.
We believe Anthropic’s decision is backed by data—but that evidence applies only to its latest models.
The obvious follow-up question is: does it work on the run-of-the-mill popular workhorse model like Deep Seek v4?
Small, fast, inexpensive models are exactly the ones you would expect to depend on detailed hand-holding. Cut their prompt in half, and conventional wisdom says they should fall over.
We ran the experiment so you don’t have to.
The setup
Our coding agent, Ante, includes a --short-prompt mode.
It consolidates the system prompt from roughly 5,000 to 2,300 characters and shortens the tool descriptions. Together, those changes reduce the complete per-request prompt from roughly 34,000 to 18,000 characters.
We A/B tested the two versions on:
- Tasks: Terminal-Bench 2.1 complete 89-task suite
- Model: DeepSeek V4 Flash
- Attempts: One per task
- Variable changed: One CLI flag
Everything else remained constant: the same agent build, pinned dataset digest, sandbox pool, effort level, and runtime flags.
The result
Performance: parity. The short prompt actually scored +2.3 points higher, but with one attempt per task that's inside the noise floor.
Input tokens: 32% lower in the same-outcome subset.
Twenty tasks changed pass/fail outcome between runs, which also changed how long those agents worked and made their token counts poor like-for-like comparisons. We exclude those 20 tasks, leaving the 69 whose outcome stayed the same, then compare the two independent medians: 509,498 input tokens with the long prompt versus 346,409 with the short prompt. That's 32% lower. This is a deliberately selected subset, not a whole-suite cost estimate.
The aggregate input-token totals across all 89 tasks are nearly flat, and that's worth being honest about: conversation history dominates input usage, and a few tasks where the short-prompt run took a longer solution path swamp the per-request savings in the sum. Run cost likewise changed only slightly, from 4.25to4.25to4.18.
What we watched for and didn't find: a capability cliff. Twenty tasks flipped outcomes between the runs — 11 newly passed, 9 newly failed — but that churn is characteristic single-attempt variance on this benchmark, not a pattern. No task category collapsWhat this experiment does not prove
This was:
- One model
- One benchmark
- One attempt per task
The 32% token result comes from a selected same-outcome subset. The amount of task churn between runs also demonstrates why multi-attempt validation matters.
With an estimated noise floor of roughly ±5 percentage points, a small regression could still be hiding in these results. We would run a multi-attempt evaluation before changing the default for everyone.
The takeaway
For this model, on this benchmark, we cut the prompt roughly in half and measured nothing worse.
Among the 69 tasks whose outcome stayed the same, the short-prompt run’s median input-token count was approximately one-third lower.
The result matches the direction Anthropic observed one tier up:
When a new model generation arrives, your first move should not be adding to the prompt. It should be deleting.
How much of your system prompt is still there because a model from two generations ago once needed it?
