The Cheap Model Often Ties the Expensive One
On many tasks in the Fronset benchmark, the leading models cannot be separated on quality, so price decides the choice. On engagement opportunity triage, Claude Haiku 5.5 scores 8.52 out of 10 (±0.33) at $0.0003 per task run in sync mode, while Claude Opus 5.5 scores 8.56 (±0.3) at $0.01 [^1]. The exceptions are real, but they are narrower than the price gaps suggest, and they change from task to task.
When the top scores tie, price decides
Every score is out of 10 and carries a ± margin. Where two models’ margins overlap, the measurement cannot separate them, and this article treats them as tied; that means the data cannot tell them apart, not that they are proven equal. Costs are per task run in sync mode, meaning ordinary synchronous calls rather than batch jobs.
Engagement opportunity triage has the broadest comparison. Four more models land in the same band as Haiku and Opus [^1]:
| Model | Score (out of 10) | Cost per task run, sync mode |
|---|---|---|
| Grok 4.7 | 8.51 (±0.32) | $0.0091 |
| Claude Sonnet 5.5 | 8.49 (±0.3) | $0.0047 |
| Qwen 3.7 Plus | 8.7 (±0.15) | $0.0017, repriced at current rates |
| GLM-5.3 | 8.74 (±0.17) | $0.02, repriced at current rates |
The intervals overlap throughout this group, so on quality it is a tie. Claude Haiku 5.5 is the lowest-cost entry in the tie, and the figures give no quality reason to pay more.
Visual theme configuration generation shows the same shape, with a warning attached. Claude Opus 5.5 scores 8.72 (±0.24) at $0.09, Claude Haiku 5.5 scores 8.45 (±0.2) at $0.0022 and Claude Sonnet 5.5 scores 8.36 (±0.29) at $0.04, a tie on stated margins [^2]. Grok 4.7 scores 6.78 (±0.23) at $0.05, clearly below all three [^2]. A tie at the top does not mean every model qualifies.
The same two models can tie on one task and split on the next
Some comparisons do separate, but they separate by task rather than by model. On section prompt generation, Claude Opus 5.5 scores 9.12 (±0.22) at $0.10 against 8.18 (±0.46) at $0.05 for Claude Sonnet 5.5, and the ranges do not overlap [^3]. On voice profile generation the same two models tie: Sonnet 5.5 scores 8.92 (±0.23) at $0.02 and Opus 5.5 scores 8.83 (±0.31) at $0.04, so Sonnet is the cheaper equal [^4].
Gemini 3.8 Flash and Gemini 3.5 Flash Lite split the same way. On profiled document analysis, Gemini 3.8 Flash scores 8.12 (±0.27) at $0.0071 and Gemini 3.5 Flash Lite scores 6.88 (±0.4) at $0.0018, so Gemini 3.8 Flash is higher on quality and higher on cost [^5]. On language identification they score 9.92 (±0.15) and 9.91 (±0.12), a tie, and Gemini 3.5 Flash Lite costs $0.0001 per task against $0.0005 for Gemini 3.8 Flash [^6].
What the pattern means
A quality gap belongs to a task, not to a model. Paying for the larger model is justified where the intervals separate, but elsewhere in the same pairs the cheaper model is not shown to be weaker. A ranking measured on one task should not be carried to another, even for the same two models.
These ties are also readings at one quality bar. A higher quality bar would change the answer, and latency is largely unmeasured in these figures. If either matters for your workload, the tie may not hold for you.
How to choose
Choose per task, starting from the cheapest model. Where leading scores tie, send the work to the cheapest model that clears your bar. Pay for a larger model only where the intervals separate, as on section prompt generation or profiled document analysis. For work that can wait, batch mode lowers cost without changing the reported score, though the size of the saving varies by task and model.
Before switching, confirm the cheap model sits inside the tie rather than just near the top of the list. The visual theme result shows that a model can fall clearly below a tied group on the same task.
The live benchmark at https://fronset.ai/benchmark/ lets you set the quality bar you actually need, since the figures here are one reading of a bar you may want to move. Method is documented at https://fronset.ai/benchmark/methodology/.