:q :q!— type: table purpose: research-brief title: “AI Model Pricing Comparison: Direct API vs OpenRouter vs Subscription” subtitle: “” description: “Per-token and subscription pricing for frontier AI models across direct API, OpenRouter, and flat-rate plans, with cost-per-task estimates for common workflows.” date: ‘2026-07-27’ section: tech tags: [ai, pricing, comparison, research-brief] ai_percentage: 90 published_at: [] —

AI Model Pricing Comparison: Direct API vs OpenRouter vs Subscription

This brief compares the cost of frontier AI models across three payment channels: direct API from the provider, OpenRouter (aggregated API), and flat-rate subscription plans. The goal is to identify the most cost-effective path for each of the three primary creative outputs: coding software, writing essays, and producing comic books.

Models and Channels

The following is accurate as of July 2026. Pricing is per million tokens input/output unless otherwise noted. Context length varies by model.

OpenRouter Pricing (Per-Token)

All models on OpenRouter are pay-per-use with no monthly commitment. OpenRouter adds response caching (identical requests at zero cost) and sticky routing via session_id for prompt cache persistence.

Table.1.OpenRouterPricing

Note on free models: OpenRouter :free endpoints have rate limits (50 requests/day for free accounts, 1,000/day after adding $10 in credits). When exhausted, OpenRouter silently falls back to the paid endpoint. Always use the :free suffix explicitly.

Direct API Pricing (DeepInfra)

DeepInfra provides direct access to DeepSeek and Qwen models without OpenRouter markup. Pricing below is from DeepInfra’s API.

Table.2.DeepInfraPricing

Note: DeepInfra’s DeepSeek V4 Flash is the same price as OpenRouter ($0.09/$0.18), while DeepSeek V4 Pro is significantly more expensive on DeepInfra ($1.30/$2.60 vs $0.43/$0.87 on OpenRouter). Always compare before routing.

Direct API Pricing (x.ai, Moonshot, Z.AI)

These providers require separate API keys and are not included in the current Hermes config, but are available via OpenRouter.

Table.3.DirectAPIPricing

Subscription Plans

Several providers offer flat-rate subscriptions. These cover web/chat access and, in some cases, CLI tools.

Table.4.SubscriptionPlans

Anthropic’s Pro subscription is the standout – it covers Claude Code CLI usage at a flat $20/mo with no per-token billing. This is dramatically cheaper than API rates for heavy users.

Cost Scenarios: Three Use Cases

The following scenarios estimate monthly costs based on realistic usage patterns for each of the user’s three primary creative outputs.

Scenario A: Coding (Heavy, 100k+ tokens/day)

Assumes 20 coding sessions/month, each generating ~10k input + ~5k output tokens. Includes the multi-turn context overhead typical of debugging and refactoring sessions.

Table.5.CodingCosts

Scenario B: Essay Writing (Moderate, 50k tokens/day)

Assumes 10 essay editing sessions/month, each generating ~15k input + ~8k output tokens including revisions.

Table.6.EssayCosts

Scenario C: Comic Panel Generation (Image API)

Assumes 200 panels/month at ~$0.002/image via Runware MCP.

Table.7.ComicCosts

Key Insights

1. The optimal split is hybrid. Claude Code via Pro subscription ($20/mo) covers the expensive reasoning work (complex coding, deep essay editing). DeepSeek V4 Flash via DeepInfra ($0.09/M) covers everything else (quick queries, research, boilerplate, drafts). Runware MCP ($0.002/image) covers comic panels.

2. OpenRouter is not always more expensive. DeepSeek V4 Flash is the same $0.09/M on both OpenRouter and DeepInfra. DeepSeek V4 Pro is actually cheaper on OpenRouter ($0.43/$0.87 vs DeepInfra’s $1.30/$2.60). Always check both.

3. GLM 5.2 and Kimi K3 are competitive on reasoning but not on price. GLM 5.2 ($0.73/$2.28) costs 8x more than DeepSeek V4 Flash and achieves comparable coding scores. Kimi K3 ($3/$15) costs 30x more. Their value is in hard reasoning (HLE: 0.547 and 0.560 respectively) where DeepSeek V4 Flash scores only 0.081.

4. Free tier pitfall. The :free OpenRouter suffix has strict rate limits (50-1,000 requests/day). Exhaustion triggers silent fallback to the paid endpoint. Always monitor your OpenRouter dashboard for unexpected usage on paid model endpoints.

5. Subscription for the expensive model, pay-per-use for the rest. The best cost optimization is to route expensive tasks (deep reasoning, complex coding) through the flat-rate Claude Pro subscription, and cheap tasks (drafts, queries, images) through per-token services.

6. Direct API for DeepSeek, OpenRouter for everything else. DeepSeek is configured as a direct provider at api.deepseek.com/v1 because OpenRouter adds no price advantage ($0.09/$0.19 same on both) and direct access gives native prompt caching without intermediary latency. For x.ai (Grok) and Moonshot (Kimi), OpenRouter provides unified billing and response caching with adequate prompt caching via session_id sticky routing – the ~5-15% markup is worth the convenience of a single API key.

7. MoA (Mixture of Agents) is not cost-effective for daily use. MoA with 2-3 reference models plus an aggregator multiplies per-turn cost by 3-4x. The caching benefits (shared prompt prefixes) only partially offset this. Reserve MoA for deep research synthesis tasks where multiple model perspectives genuinely improve output quality. For daily coding and editing, a single frontier model is cheaper and just as effective.

Compiled 2026-07-27 from OpenRouter API, DeepInfra API, provider pricing pages, and independent benchmark aggregators.