These agentic ai vs generative ai examples show that one-shot generative tasks are the cheapest way to produce a given amount of text, while an agent loop re-reads its context every step and can cost many times more for the same answer. Verdict: generate with one-shot calls, decide with agents, and run agent loops on the cheap-input models (GLM-5.3 FlashX, Claude Haiku 5.5). In the loop example below, Claude Opus 5.5 costs 12.3x as much as GLM-5.3 FlashX for an identical run (Anthropic pricing, OpenRouter catalog).
For the market-level view of when each approach fits, see the market-level verdict; this page is the arithmetic behind it.
TL;DR
- Generative work bills each token once; agentic work re-bills the growing context at every step, so input price dominates.
- Five product descriptions cost $0.31 on Claude Opus 5.5 and $0.008 on Claude Haiku 5.5 (Anthropic pricing); a 30-step research agent costs $0.96 per run on Opus 5.5 against $0.078 on GLM-5.3 FlashX (OpenRouter catalog), a 12.3x gap.
- Reading a 1M-token corpus is decided by the per-million input price alone.
- A mixed two-model pipeline cost about 6x less than the same pipeline on Opus 5.5 alone.
- All prices were read on 4 November 2026 and will drift. Recompute with your own token counts.
How do you calculate the cost of an AI task?
Cost = input tokens x input price, plus output tokens x output price, per million. Every example uses only that formula; no caching or batch discount unless stated.
Rupee figures use 94.83 INR per USD, the rate for 4 November 2026 from open.er-api.com. The ECB-based Frankfurter feed read Rs 94.75 for 3 November; the choice of rate moves totals by under 0.1 per cent.
What do current API prices look like?
Prices per 1M tokens, input and output, with INR at the rate above (rounded). We pulled the ten rows below live on 2026-11-04 from OpenRouter's public catalog (n=10 listed models) and cross-checked the Anthropic rows against Anthropic's own pricing page.
| Model | USD in / out | INR in / out | Source |
|---|---|---|---|
| Claude Fable 5.1 | $10 / $50 | Rs 948 / Rs 4,742 | Anthropic |
| Claude Opus 5.5 | $4 / $20 | Rs 379 / Rs 1,897 | Anthropic |
| Claude Sonnet 5.5 | $2 / $10 | Rs 190 / Rs 948 | OpenRouter |
| Claude Haiku 5.5 (prompts up to 100K) | $0.10 / $0.50 | Rs 9 / Rs 47 | Anthropic |
| GPT-6.1 Terra Pro | $1 / $5 | Rs 95 / Rs 474 | OpenRouter |
| Gemini 4 Argon | $2 / $10 | Rs 190 / Rs 948 | OpenRouter |
| GLM-5.3 Prime | $2.80 / $8.80 | Rs 266 / Rs 835 | OpenRouter |
| GLM-5.3 FlashX | $0.37 / $1.25 | Rs 35 / Rs 119 | OpenRouter |
| GLM-5.3 Flash | $0.03 / $0.09 | Rs 3 / Rs 9 | OpenRouter |
| DeepSeek V4.1 Flash (17 Oct refresh) | $0.15 / $0.60 | Rs 14 / Rs 57 | OpenRouter |
The earlier DeepSeek V4.1 Flash listing was $0.04 / $0.50; Z.ai publishes the GLM-5.3 family prices on its model-api page. Context windows: 1,000,000 tokens for Opus 5.5 (Anthropic pricing), 1,050,000 for GPT-6.1 Terra Pro, and 1,048,576 each for GLM-5.3 FlashX and Gemini 4 Argon (OpenRouter catalog).
What does a generative AI example cost?
Take five product-description drafts: 5 x (a 3,000-token brief in, a 2,500-token draft out) = 15,000 input and 12,500 output tokens, billed once each at the prices in the table above (Anthropic pricing, OpenRouter catalog). There are no tool calls and nothing is re-read.
| Model | Cost per batch | In rupees |
|---|---|---|
| Claude Opus 5.5 | $0.310 | Rs 29.40 |
| GPT-6.1 Terra Pro | $0.078 | Rs 7.35 |
| Claude Haiku 5.5 | $0.008 | Rs 0.73 |
The Opus 5.5 working: 15,000 x $4 / 1M = $0.06 input plus 12,500 x $20 / 1M = $0.25 output (Anthropic pricing, OpenRouter catalog); on cheaper models, output dominates.
What does an agentic AI example cost?
Now give an agent the same brief: 30 steps, each re-reading a 5,000-token context and writing a 600-token step, billed at the same table prices (Anthropic pricing, OpenRouter catalog). That is 150,000 input tokens and 18,000 output tokens, with no caching.
| Model | Cost per run | In rupees |
|---|---|---|
| Claude Opus 5.5 | $0.96 | Rs 91.04 |
| GPT-6.1 Terra Pro | $0.24 | Rs 22.76 |
| GLM-5.3 FlashX | $0.078 | Rs 7.40 |
| Claude Haiku 5.5 | $0.024 | Rs 2.28 |
Input is now ten times the generative volume, so input price decides the ranking: GLM-5.3 FlashX reads at $0.37 per million against $4.00 for Opus 5.5 (OpenRouter catalog), the 12.3x gap. Real contexts usually grow with the task, so treat these totals as a floor.
When does agentic vs generative stop mattering?
When the job is almost entirely input: reading a roughly 1M-token contract corpus in one prompt, about 1,000 tokens of output, input cost only (OpenRouter catalog; Anthropic pricing).
| Model | Input cost for 1M tokens | In rupees |
|---|---|---|
| Claude Opus 5.5 | $4.00 | Rs 379.34 |
| GPT-6.1 Terra Pro | $1.00 | Rs 94.83 |
| GLM-5.3 FlashX | $0.37 | Rs 35.09 |
FlashX does the read for about a tenth of Opus 5.5's price. Haiku 5.5's $0.10 rate covers only the first 100,000 tokens of a prompt, so it stays out rather than guess the higher tier (Anthropic pricing). Call it agent or single call: the per-million input price is what you pay.
What does a mixed pipeline cost?
Most production systems chain the two. A generative layer summarises 100 support tickets (60,000 tokens in, 15,000 out); an agentic layer then triages refunds, re-reading about 100,000 tokens and writing 8,000, all at the table prices above (OpenRouter catalog).
- Generative layer on GPT-6.1 Terra Pro: 60,000 x $1 / 1M plus 15,000 x $5 / 1M gives $0.135 (Rs 12.80) (OpenRouter catalog).
- Agentic layer on GLM-5.3 FlashX: 100,000 x $0.37 / 1M plus 8,000 x $1.25 / 1M gives $0.047 (Rs 4.46) (OpenRouter catalog).
- Mixed total: about $0.18 (Rs 17.26) across the two layers (OpenRouter catalog; Anthropic pricing).
- Everything on Opus 5.5: $0.54 for the generative layer plus $0.56 for the agentic layer, $1.10 in total (Rs 104) (Anthropic pricing).
All-Opus costs about 6x more (6.1x on the rounded figures): the gap between models within a layer beats any single-model discount. For a deeper two-family fight, see Claude Sonnet price vs GLM-5.3.
How do you run these numbers yourself?
- Count tokens for one real task: prompt, average step context, output.
- Multiply by the table above, or the live OpenRouter catalog, which returns the full price list as JSON without an API key.
- For agents, price context x steps before committing.
- Pilot the cheaper model on your own eval set; a low token price only helps if the output is good.
Direct access goes through console.anthropic.com, platform.openai.com, z.ai/model-api and ai.google.dev. OpenRouter serves them under one key as anthropic/claude-opus-5.5, openai/gpt-6.1-terra-pro, google/gemini-4-argon and z-ai/glm-5.3-flashx; confirm a gateway's price matches the first-party list.
Two levers change the verdict: Anthropic's batch API runs at 50 per cent of listed rates (Opus 5.5: $2 / $10), and prompt caching charges Opus 5.5 cache reads at $0.20 to $0.50 per million depending on the write TTL (5-minute and 1-hour writes $5 and $8) (Anthropic pricing). Caching turns a long agent loop from the worst case into a manageable one; measure with it on.
For the architecture behind these loops, read how agentic AI differs from earlier AI and MCP in agentic AI; to pick a provider, see Gemini vs Claude API pricing.
What are the limits of this comparison?
These examples count tokens only: no retries, tool-call latency, failed runs, engineering time or quality - a cheap model that needs three attempts is not cheap. Prices were read on 4 November 2026; providers change them often.
FAQ
Q: Is agentic AI always more expensive than generative AI?
A: Usually, for the same output, because each step re-reads the context. In the research example, one Opus 5.5 agent run cost $0.96 against $0.31 for a five-draft generative batch, though the tasks differ in shape (Anthropic pricing).
Q: Which model is cheapest for an agent loop?
A: Of the models compared here, Claude Haiku 5.5 at $0.024 per run and GLM-5.3 FlashX at $0.078 were lowest in the 30-step example (Anthropic pricing, OpenRouter catalog).
Q: Does prompt caching change the verdict?
A: Yes, for agents: cached reads on Opus 5.5 cost $0.20 to $0.50 per million instead of the $4 input rate (Anthropic pricing), shrinking the re-read cost that makes loops expensive.
Q: Should I use one model or several?
A: Yes, where the workload allows: the split pipeline ran about $0.18 against $1.10 for all-Opus (OpenRouter catalog).
Q: What exchange rate do the rupee figures use?
A: 94.83 INR per USD, the 4 November 2026 rate from open.er-api.com. Rupee totals will shift slightly as the rate moves.
Corrections log
Prices and the USD-INR rate drift; re-check the table against the live OpenRouter catalog before you budget. Last verified 2026-11-04. Drafted with AI assistance and edited by the aitecharchive.com editorial process; see /pages/how-we-work for how we verify pricing.
Top comments (0)