Vals reports Mercury 3 costs $1.25 per coding task, versus $0.38 for its predecessor

• Mercury 3 reached 60.9% on Vibe Code Bench versus 3.7% for Mercury 2.5 in ValsAI's comparison, while reported spending rose from $0.38 to $1.25 per app in its cost breakdown.
• Input accounted for 88% of Mercury 3's Vals Index spending, according to the cache breakdown; only 84% of input tokens came from cache.
• Terminal-Bench 4 accuracy was 3.0%, and ValsAI reported repeated agent crashes on a task that exhausted the output budget on reasoning.

Inception advertises 1,068 tokens per second, and its API counts reasoning tokens as output. The live Vals model page uses launch-discount prices and reports a different cost per test from the thread.

Mercury 3 placed #23 on the Vals Index, just ahead of Beam 1.0, according to ValsAI. Its reported improvements over Mercury 2.5 span three benchmarks:

• Vibe Code Bench: 3.7% → 60.9%, +57.2 percentage points.
• Legal Research: +24.5 points.
• Tax Agent: +15.7 points.

Six of the 50 apps passed every test; Mercury 2.5 produced no perfect apps. The headline 60.9% averages per-app test scores: under Vals' scoring methodology, a UI test passes when at least 90% of its substeps succeed.

The reported app-building cost rose from $0.38 to $1.25, about 3.3 times as much, despite slightly cheaper output tokens. Vals quoted $0.70 per million output tokens for Mercury 3 versus $0.75 for Mercury 2.5, with input unchanged at $0.20 per million.

Vals attributed much of the performance gain to longer reasoning. Mercury 3 generated substantially more output:

• Legal Research: about 12 times Mercury 2.5's output-token volume.
• Tax Agent: about 9 times its predecessor's output-token volume.

At the prices quoted in ValsAI's thread, Mercury 3 averaged $0.97 per Vals Index test. Three higher-scoring models cost less:

| Model | Reported cost per Index test |
| --- | ---: |
| Mercury 3 | $0.97 |
| GPT-6 Luna | $0.45 |
| DeepSeek V4.1 Flash | $0.66 |
| MiMo V2.6 Flash | $0.22 |

Uncached input carried the sting: each token cost 10 times the cached rate. Vals attributed the cost gap to these input-side numbers:

• Cached share of input: 84% for Mercury 3, versus 95–98% for those competitors.
• Input's share of total token spending: about 88% for Mercury 3.
• Code Migration spending: $4.41 per task.

Inception's official pricing table lists a 50% launch discount: $0.10 per million input tokens and $0.35 per million output tokens, versus the $0.20/$0.70 rates quoted in Vals' cost thread.

Vals now lists those discounted token rates on its model profile, alongside $0.478 per Index test, compared with the thread's $0.97. The $1.25-per-app figure accompanies the undiscounted rates in the thread.

Mercury 3 scored 3.0% on Terminal-Bench 4. On one task, it consumed its entire output budget on reasoning and the agent crashed; Vals observed that failure in two separate runs.

Vals' published default settings list high reasoning effort and a maximum output of 131,072 tokens. Individual benchmarks can use different providers and parameters.