Social Copy Is a Tie at the Top, So Buy Latency, Not Rank

On short promotional posts, reply review and post selection, the leading models score so close that the benchmark’s judges usually cannot separate them. That makes the model choice for most social and promotional work a question of latency and price, not quality. The one exception is picking which community to post in, where the leaders pull apart and the head-to-head winner is worth paying for.

How close “close” is

The benchmark has models answer the same live input, scores each answer on a 10-point scale, and reports the gap between the top two. For writing a short promotional message, the top two answers scored exactly the same in 41.9% of 31 judged occasions, and finished within half a point in 96.8% — median gap 0.1 on a 10-point scale [^1][^2].

Reviewing replies for public engagement is tighter still, with the top two answers scoring exactly the same in 50.0% of 34 judged occasions and finishing within half a point in 97.1%, median gap 0.1 [^3]. Choosing a post for X scored exactly the same in 64.3% of 84 judged occasions, and finished within half a point in 90.5% — median gap 0.0 [^4][^5].

Individual scores tell the same story. On short promotional post work, GLM-5.3 Flash gained 0.08 points since 2026-09-05, to 8.54 on a 10-point scale [^2], while GLM-5.3 stands at 8.53 out of 10, ±0.22 after 30 independent judgments [^2]. A hundredth of a point inside a ±0.22 interval is a tie, and a leaderboard that ranks one above the other is not telling you anything you can act on.

What to optimize instead

When quality is interchangeable, the operational numbers are the whole decision, and those are not close. For writing a short promotional social post, GPT-5.6 Terra answers in 1.5 seconds at the median, measured over 104 synchronous calls under a 600-second per-call deadline [^6][^7][^8]. DeepSeek V4 Flash takes 9.2 seconds at the median on the same work, measured over 2,371 synchronous calls under the same deadline [^6][^7][^8][^9][^10].

That is a large gap in wait time between two models a judge would struggle to tell apart on output. If your product shows the post to a user while they wait, latency is the number that decides which model you ship.

Length is a second lever, and it points the same way. When writing a promotional campaign, the answers judges picked were shorter than the ones they passed over — 3,338 against 4,089 output tokens [^1][^2]. Paying for extra output tokens on promo copy is paying for something the judges marked down, so a tight output budget saves money without costing quality.

The exception: community selection

Choosing which community to post in was the most differentiated task in this set. The top two answers scored exactly the same in only 6.9% of 72 judged occasions, and finished within half a point in 48.6% — median gap 0.6 on a 10-point scale [^11][^12]. That 0.6 median is three to six times the 0.0–0.2 medians the same judges produced on short messages, teasers, translation and voice.

Here the head-to-head record is worth reading. GPT-5.6 Terra was judged the best answer in 17 of the 41 occasions it competed in — 41.5% — every one of them against other models answering the same live input [^11][^12]. Community routing is the one social task where you should pick the winner rather than the cheapest model that clears your bar.

Cheap on one task does not mean cheap on the next

Paired cost and quality figures are thin in this cycle, but the one that exists comes with a warning attached. GPT-5.4 Nano costs $0.0015 per task run in batch mode on publication title packages, measured on this model’s own usage and repriced at current rates [^13][^14]. On engagement triage, the same model answered the same inputs as its rivals in 50 judged occasions and was chosen best in none of them [^15][^16].

An average quality score never surfaces a zero-win record like that. So route by task rather than by model: keep Nano on the high-volume title work where it is priced, and keep it out of triage, where it never won a comparison.

Where to look next

These are snapshot numbers, and the ties move a little every cycle. Before you lock a route, check the live figures at https://llm-bench.kapualabs.com/ and the per-task pages such as https://fronset.ai/benchmark/task/promotional-message-generation/ and https://fronset.ai/benchmark/task/subreddit_selection/. Set the quality bar your workflow needs, then compare latency and batch-versus-sync pricing yourself among the models that clear it.