Low-cost hosted models and the local Gemma 4 deployment, measured on aligned
material from this corpus rather than vendor claims. Quality is chrF (character
6-gram F-score), which is robust for morphologically rich source languages.
65aligned test segments35Greek (Perseus)30Coptic (Scriptorium)1,468works to translate213 MBsource text
Quality
Model
Kind
chrF Greek
chrF Coptic
chrF mean
BLEU
Length
Errors
Total USD
DeepSeek V4.1 Flash
deepseek/deepseek-v4.1-flash-1017
openrouter
68.1
51.1
61.0
40.2
1.04
0
$41
DeepSeek V4 Flash
deepseek/deepseek-v4-flash
openrouter
65.8
54.5
61.0
40.6
1.00
0
$34
GLM 5.3 Flash
z-ai/glm-5.3-flash
openrouter
64.7
42.7
55.4
36.0
1.00
0
$15
Gemma 4 31B
google/gemma-4-31b-it
openrouter
63.2
29.8
48.9
27.5
1.06
0
$77
Qwen 3.5 Flash
qwen/qwen3.5-flash-02-23
openrouter
65.6
24.9
47.9
25.4
1.06
0
$33
Qwen 3.7 Flash
qwen/qwen3.7-flash
openrouter
65.8
20.3
46.9
29.0
1.00
0
$16
Gemma 4 26B A4B
google/gemma-4-26b-a4b-it
openrouter
61.6
23.8
45.6
25.4
1.00
0
$32
Mistral Small 24B
mistralai/mistral-small-24b-instruct-2501
openrouter
63.3
17.4
41.3
15.6
1.04
0
$18
Llama 3.1 8B
meta-llama/llama-3.1-8b-instruct
openrouter
56.7
14.6
32.6
6.8
1.14
0
$8
Test set: AP, Acts, Barnabas, John, Luke, Mark, Matthew, Romans, abraham, besa-letters, bohairic-habakkuk, bohairic-jonah, bohairic-lausiac. Greek references are
Perseus English (NT and Barnabas); Coptic references are Coptic Scriptorium sentence
translations. Ethiopic has no aligned gold set yet and is not scored. A 65-segment
set is a screening instrument, not a publication benchmark.
Cost for all transcribed untranslated text
1,468 works with a local source text and no complete English
translation: 213 MB of text
(213,256,954 characters), estimated at
258,544,436 input tokens and
62,181,053 output tokens. The “source-only”
column discounts facing Latin apparatus by the measured script ratio. Both include
12 % prompt overhead and 5 % retries.
Model
$/M in · out
Upper USD
Source-only USD
GPU-hours
Wall days
Llama 3.1 8B
meta-llama/llama-3.1-8b-instruct
$0.02 / $0.04
$8
$6
—
—
GLM 5.3 Flash
z-ai/glm-5.3-flash
$0.04 / $0.078
$15
$11
—
—
Qwen 3.7 Flash
qwen/qwen3.7-flash
$0.03 / $0.13
$16
$11
—
—
Mistral Small 24B
mistralai/mistral-small-24b-instruct-2501
$0.05 / $0.08
$18
$13
—
—
Gemma 4 31B (local)
local:gemma4:31b
electricity
$24
$24
215.9
9.0
Gemma 4 26B A4B
google/gemma-4-26b-a4b-it
$0.068 / $0.225
$32
$23
—
—
Qwen 3.5 Flash
qwen/qwen3.5-flash-02-23
$0.065 / $0.26
$33
$24
—
—
DeepSeek V4 Flash
deepseek/deepseek-v4-flash
$0.09 / $0.18
$34
$25
—
—
DeepSeek V4.1 Flash
deepseek/deepseek-v4.1-flash-1017
$0.038 / $0.5
$41
$30
—
—
Gemma 4 31B
google/gemma-4-31b-it
$0.2 / $0.4
$77
$55
—
—
Batch options
Model
$/M in · out
Upper USD
Source-only USD
Note
DeepSeek V4.1 Flash (batch)
deepseek/deepseek-v4.1-flash-1017:batch
$0.05 / $0.15
$22
$16
asynchronous batch endpoint; price from OpenRouter
DeepSeek V4 Flash 0731 (batch)
deepseek/deepseek-v4-flash-0731:batch
$0.05 / $0.15
$22
$16
asynchronous batch endpoint
GLM 5.3 Flash (batch)
z-ai/glm-5.3-flash:batch
$0.05 / $0.18
$24
$17
asynchronous batch endpoint
Gemma 4 31B (free tier)
google/gemma-4-31b-it:free
$0 / $0
$0
$0
free but rate-limited; not dependable for bulk
Best value above the quality floor (chrF ≥ 55):
GLM 5.3 Flash — chrF 55.4,
$15 upper / $11 source-only.
Best measured quality: DeepSeek V4.1 Flash — chrF 61.0,
$41 upper.
Reading the table
DeepSeek V4.1 Flash and V4 Flash tie on mean chrF (61.0). V4 Flash
is cheaper on output and slightly better on Coptic; V4.1 Flash is slightly better
on Greek. Both are the quality choice.
GLM 5.3 Flash is the value choice at chrF 55.4: Greek is close to the
leaders, Coptic is ~12 points behind. Use it if Greek dominates the batch.
Qwen 3.7 Flash and Gemma 4 26B are strong on Greek (~62–66) and
weak on Coptic (~20–25). They are not a mixed-corpus choice.
Llama 3.1 8B is the price floor and scores accordingly (32.6); it is not
usable for publication-grade drafts.
Local Gemma 4 31B has no API bill (~$18–25 of electricity for the corpus),
but on troy-1 today it reloaded 31B weights on every call (30–65 s) because the
GPU was shared; a dedicated card at 80 tok/s would need ~160 GPU-hours. Use it
only when the card is otherwise idle.
The API bill is not the project cost. The same corpus is
roughly 34–35 million English words; review at 1,500 words/hour is ~23,000 hours
(about 11 person-years), two orders of magnitude above every model in the table.
Caveats
Prices are live OpenRouter list prices on 2026-11-14;
providers change them. Batch endpoints are asynchronous and may have their own
limits.
Greek references are New Testament / Barnabas, not patristic prose; a model that
scores well here can still mishandle a catena or a canon-law collection.
Coptic references are sentence-level and sometimes free translations, so the
Coptic column is noisier than the Greek one.
chrF rewards shared characters; a readable but paraphrastic translation can score
lower than a literal one. Read the samples before committing.
The corpus scope is the 606 transcribed non-full works. OCR-only scans are not
included; if the 65 open scans are OCR'd, add their text volume.