You Picked the Wrong AI Model. Here's How Much It's Costing You.

You're starting a new project. Maybe it's a customer support bot. Maybe it's a document analysis pipeline. Maybe it's an internal code assistant. You need an LLM, and you need to pick one.

You pick GPT-4o. It's the most capable, the most talked about, the one your team already has API access to. The decision takes about 30 seconds.

That 30-second decision just set your baseline cost at $8.13 per 1,000 support chats.

Gemini 1.5 Flash handles the same support chat workload for $0.25 per 1,000 requests. That's a 32x difference — on a model that will pass your quality bar for this use case with almost identical output.

If you're processing 100,000 support chats a month, you just signed yourself up for $813/month instead of $25/month. And you haven't shipped yet, so you don't even know if the unit economics work.

This is how it happens. Not from ignorance — from momentum.

---

The Real Pricing Table

Before doing any cost math, you need current numbers. These are approximate as of Q1 2026 and subject to change — always verify against provider pricing pages before committing to an architecture.

| Model | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | GPT-4o | $2.50 | $10.00 | | GPT-4o mini | $0.15 | $0.60 | | Claude 3.5 Sonnet | $3.00 | $15.00 | | Claude 3 Haiku | $0.25 | $1.25 | | Gemini 1.5 Pro | $1.25 | $5.00 | | Gemini 1.5 Flash | $0.075 | $0.30 | | DeepSeek V3 | $0.27 | $1.10 | | DeepSeek R1 | $0.55 | $2.19 |

Pricing approximate as of Q1 2026 and subject to change.

Notice the spread. Between GPT-4o and Gemini 1.5 Flash, the input price ratio is 33x. Between Claude 3.5 Sonnet and Claude 3 Haiku, it's 12x. These aren't marginal differences — they're orders of magnitude that compound with every request your application makes.

---

Real Cost Per 1,000 Requests Across Three Use Cases

Let's make this concrete. Three common workloads, modeled at realistic token counts.

Use Case 1: Customer Support Chat

Token profile: ~1,850 input tokens, ~350 output tokens per turn. This assumes a system prompt (~800 tokens), conversation history (~800 tokens), and a user message (~250 tokens), plus a short-to-medium response.

| Model | Cost per 1,000 requests | |---|---| | GPT-4o | $8.13 | | GPT-4o mini | $0.49 | | Claude 3.5 Sonnet | $9.80 | | Claude 3 Haiku | $0.90 | | Gemini 1.5 Pro | $4.07 | | Gemini 1.5 Flash | $0.25 | | DeepSeek V3 | $0.88 | | DeepSeek R1 | $1.78 |

Winner: Gemini 1.5 Flash, and it's not close. For high-volume support chat where the quality bar is "helpful and coherent" rather than "Nobel Prize reasoning," you are overpaying by 32x if you chose GPT-4o.

Use Case 2: Document Analysis

Token profile: ~6,400 input tokens, ~800 output tokens. A typical pipeline that ingests a contract or report and extracts structured information.

| Model | Cost per 1,000 requests | |---|---| | GPT-4o | $24.00 | | GPT-4o mini | $1.44 | | Claude 3.5 Sonnet | $31.20 | | Claude 3 Haiku | $2.60 | | Gemini 1.5 Pro | $12.00 | | Gemini 1.5 Flash | $1.72 | | DeepSeek V3 | $2.60 | | DeepSeek R1 | $5.27 |

Winner: Gemini 1.5 Flash edges it on price; DeepSeek V3 ties Claude 3 Haiku at $2.60. The cost difference between GPT-4o and DeepSeek V3 is nearly 10x. If document processing is your core product loop, this is your business model.

Use Case 3: Code Generation

Token profile: ~3,000 input tokens, ~1,200 output tokens. A realistic code completion or generation task with context and a substantive output.

| Model | Cost per 1,000 requests | |---|---| | GPT-4o | $19.50 | | GPT-4o mini | $1.17 | | Claude 3.5 Sonnet | $27.00 | | Claude 3 Haiku | $2.25 | | Gemini 1.5 Pro | $9.75 | | Gemini 1.5 Flash | $0.59 | | DeepSeek V3 | $2.13 | | DeepSeek R1 | $4.28 |

Winner on price: Gemini 1.5 Flash. But code generation is one area where quality matters enough to trade some cost for it. Claude 3 Haiku and DeepSeek V3 offer a compelling middle ground — solid output at roughly 10% of GPT-4o cost.

---

The Switching Cost Nobody Mentions

Say you built six weeks of your project on Claude 3.5 Sonnet and you've just done the math above. You want to move to Gemini 1.5 Flash. The API cost will drop 55x. You change the model string.

Here's what the budget doesn't show: switching cost.

• Prompt engineering rework. Prompts are not portable. A system prompt tuned for Claude's instruction-following will behave differently on Gemini. Tone, formatting defaults, tool-use behavior, response length — all shift. Count on 1–2 sprints of re-tuning and regression testing.
• Context window differences. Gemini 1.5 Pro has a 1M token context window. GPT-4o has 128K. Claude 3.5 Sonnet has 200K. If a feature is designed around a specific ceiling, switching breaks it.
• Response formatting variations. Claude defaults to structured markdown. GPT-4o varies more. Gemini can be terse. If your application parses or displays LLM output in any structured way, expect inconsistencies.
• Tool call syntax. Function calling is conceptually similar across providers but syntactically different. Rewriting tool use from Anthropic's format to OpenAI's — especially with complex schemas — is non-trivial.
• Rate limit structures. RPM, TPM, and daily limits differ per provider and tier. Your retry logic and concurrency model may need adjustment.

A real migration from Claude to GPT mid-project at a team of 3–5 engineers costs 2–3 sprints of engineering time. That's the operational reality reported by teams who've done it.

The decision you make in week 1 is cheap. Undoing it in week 8 is not.

---

The Tier Heuristic: When to Use What

Flagship tier (GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro): Complex multi-step reasoning, nuanced judgment, quality-critical user-facing responses, agentic tasks where errors cascade.

Mid-tier (DeepSeek V3, GPT-4o mini for structured tasks): Well-defined structured tasks — classification, information extraction, translation, formatting. The instruction is clear, the output schema is fixed.

Cheap tier (Gemini 1.5 Flash, Claude 3 Haiku, GPT-4o mini): High-volume, low-stakes operations — support triage, content moderation pre-screening, bulk summarization for internal tools, pipeline steps followed by human review.

The mistake is applying the flagship tier uniformly because it feels "safer." It's not safer — it's just more expensive, and the cost structure eventually becomes a product constraint.

---

The DeepSeek Caveat

DeepSeek V3 and R1 offer remarkable price-performance. For pure cost-per-output-quality, they are competitive with models 5–10x more expensive. However: DeepSeek is a Chinese company, and their models run on infrastructure subject to Chinese jurisdiction.

For teams in regulated industries, or handling data with GDPR, HIPAA, or contractual data residency requirements, this is a real compliance consideration — not a vague concern. For internal tools with no PII? The calculus is different than for a product processing customer documents.

---

Batch API: The 50% Discount Almost Nobody Uses

Both OpenAI and Anthropic offer batch processing APIs at approximately 50% of standard pricing. You submit requests; results come back within 24 hours (often much faster).

For any workload that doesn't require a real-time response — nightly document processing, weekly report generation, bulk classification, async enrichment pipelines — batch mode cuts your API bill in half.

On a $2,000/month API bill, that's $1,000/month sitting on the table. Most developers don't use it because they build everything as synchronous request-response loops and never revisit it. Check whether any of your scheduled workloads are eligible.

---

The Call You Make in Week 1

The model decision doesn't feel consequential at the start. You're writing a prototype. You just want something that works.

But the model you pick sets your token budget, your context window assumptions, your prompt engineering investment, and your integration patterns. By the time those are load-bearing, switching has real cost.

Pick based on task requirements and cost math, not name recognition. Run the numbers for your specific token profile. Check whether batch mode applies to any of your workloads.

If you want to see what this looks like across your actual keys and projects in one place, API Lens tracks spend per model, per provider, and per project — so the math is always current, not hypothetical.