The Feature: A Support Assistant With Retrieval
A typical B2B support feature: system prompt + retrieved docs as input, a grounded answer as output. 4,000 input / 1,000 output per request, 50% cacheable input, 1M requests a month. Priced at verified 2026-08-28 rates:
| Model | Rate ($/1M in, out) | Per request | Monthly (1M req) | Composite (int/cod/agt) |
|---|---|---|---|---|
| GPT-5.6 Terra | $2 / $12 | $0.0200 | $20,000 | 56.6 / 76.7 / 50.2 |
| Claude Sonnet 5 | $2 / $10 | $0.0180 | $18,000 | 55.3 / 71.5 / 49.7 |
| Gemini 3.7 Flash | $1.50 / $6 | $0.0120 | $12,000 | 56.0 / 76.1 / 45.1 |
| GLM 5.3 | $1.40 / $4.40 | $0.0100 | $10,000 | 59.5 / 74.8 / 59.1 |
| DeepSeek V4 Pro | $1.32 / $3.96 | $0.0092 | $9,240 | 53.2 / 68.8 / 49.6 |
| DeepSeek V4 Pro (off-peak) | $0.66 / $1.98 | $0.0046 | $4,620 | 53.2 / 68.8 / 49.6 |
The headline: the same feature spans a 4.3× cost range while the quality spread is ~6 points on intelligence and ~8 on coding. Terra is the trap — identical input rate to Sonnet 5 but a $12 output rate that makes it the most expensive option here. Output price is the whole game.
With Prompt Caching: The Ranking Narrows
Support features repeat the same system prompt and docs constantly — exactly what caching is for. At 90% cached input, per-request costs drop across the board:
The shape is consistent: cached input shrinks the input line 90%, but output tokens are never discounted — Terra still pays $12 per 1K outputs while DeepSeek off-peak pays $1.98. The cheapest per-request model stays cheapest after caching, which is why output rate is the first number to check on any rate card.
Quality Verdicts, Use Case by Use Case
Weighted fit scores (composite blend per workload) rank the five differently depending on what the feature does:
| Use case | Best fit | Runner-up | Why |
|---|---|---|---|
| Support chatbot | GLM 5.3 | Claude Sonnet 5 | Highest intelligence + agentic blend; strong instruction following |
| Code assistant | GPT-5.6 Terra | Gemini 3.7 Flash | Coding 76.7 edges out Flash's 76.1 — but Flash is 40% cheaper |
| RAG / grounded answers | GLM 5.3 | GPT-5.6 Terra | Agentic 59.1 vs 50.2 decides retrieval-heavy loops |
| Extraction pipeline | DeepSeek V4 Pro | GLM 5.3 | Fit is close; off-peak V4 Pro wins on value (fit ÷ cost) |
Full fit methodology and weights are documented on the methodology page, and every head-to-head is on the comparison arena.
The Build Verdict
Ship DeepSeek V4 Pro for the volume
At off-peak, $4,620/month for the same feature that costs $20,000 on Terra. Extraction and support answers don't need the extra 6 intelligence points.
Keep Gemini 3.7 Flash for interactive paths
Instant latency tier + coding 76.1 — the best balance for user-facing turns where speed shows.
Use GLM 5.3 where quality is the product
Highest composites of the five, $10,000/month. If the feature's output is the differentiator, this is the defensible premium.
Skip Terra at this workload
A $12 output rate with no composite advantage over Flash/GLM makes it the worst value in the five — the model-choice mistake most teams make by inertia.
Frequently Asked Questions
Which model is cheapest for a production SaaS feature?
At 1M requests/month with a 4K-in/1K-out workload, DeepSeek V4 Pro off-peak is cheapest at $4,620/month, then Gemini 3.7 Flash at $12,000. GPT-5.6 Terra is the most expensive at $20,000 — its $12 output rate makes it cost more than Sonnet 5 despite matching input pricing.
Does the cheapest model mean lower quality?
Third-party composites say the gap is modest: GLM 5.3 leads the five on weighted fit (59.5 intelligence / 74.8 coding / 59.1 agentic), Terra and 3.7 Flash are within a few points, and DeepSeek V4 Pro trails slightly on coding (68.8). For most chatbot/extraction features the difference is invisible.
What makes output price the deciding factor?
At 4K in / 1K out, output is 20% of tokens but 60-80% of the cost. Terra charges $12/1M output vs Sonnet 5's $10 and Flash's $6 — identical-looking models differ 2x purely on the output line.
How does prompt caching change this comparison?
With 90% cached input, Terra's per-request cost drops from $0.020 to $0.0128 (-36%) and the monthly bill from $20,000 to $12,800. Caching narrows but does not reverse the ranking — output remains the dominant line.
Which of the five would you ship in production?
DeepSeek V4 Pro off-peak for price-sensitive traffic with a fallback to Gemini 3.7 Flash for latency-critical paths, and GLM 5.3 or Claude Sonnet 5 for the tasks where quality visibly matters. Terra at $20,000/month is the one to justify — its output rate makes it hard to defend.