The pricing landscape changed dramatically in 2026. OpenAI's GPT-4.1 undercuts Claude Sonnet on input tokens. Anthropic's Haiku 4.5 remains the cheapest capable model. Here's the real math.

Current Pricing (June 2026)

OpenAI

Model	Input (per 1M tokens)	Output (per 1M tokens)	Context	Best For
GPT-5.4	$2.50	$10.00	128K	Flagship reasoning
GPT-4.1	$2.00	$8.00	128K	Code, instruction-following
GPT-4.1 mini	$0.40	$1.60	128K	Balanced performance
GPT-4.1 Nano	$0.10	$0.40	128K	High-volume simple tasks
GPT-4o	$2.50	$10.00	128K	Multimodal

Discounts: 50% prompt caching on cache hits. 50% batch API discount.

Anthropic

Model	Input (per 1M tokens)	Output (per 1M tokens)	Context	Best For
Claude Opus 4.7	$5.00	$25.00	200K	Complex analysis, research
Claude Sonnet 4.6	$3.00	$15.00	200K	Balanced, coding
Claude Haiku 4.5	$1.00	$5.00	200K	Fast, cheap, capable

Discounts: 90% prompt caching on read hits. 25% batch API discount.

Head-to-Head: Same Tier

Tier	OpenAI	Anthropic	Cheaper
Flagship	GPT-5.4: $2.50/$10	Opus 4.7: $5/$25	OpenAI (2x cheaper)
Mid-tier	GPT-4.1: $2/$8	Sonnet 4.6: $3/$15	OpenAI (input 33% cheaper, output 47% cheaper)
Budget	GPT-4.1 Nano: $0.10/$0.40	Haiku 4.5: $1/$5	OpenAI (10x cheaper input)

On raw pricing, OpenAI wins every tier in 2026. But pricing alone doesn't determine your real cost.

Why Raw Pricing Is Misleading

1. Anthropic's Caching Is Better

Anthropic gives 90% discount on cached prompt reads. OpenAI gives 50%.

If 70% of your tokens hit cache (common in production):

Sonnet 4.6 effective input: $0.30/1M (90% off on 70% of tokens)
GPT-4.1 effective input: $1.30/1M (50% off on 70% of tokens)

With heavy caching, Anthropic becomes cheaper on input.

2. Opus 4.7's New Tokenizer

Opus 4.7 introduced a new tokenizer that produces ~35% more tokens for the same text compared to older models. Your effective cost per word is higher than the rate card suggests.

3. Output Tokens Dominate Your Bill

Most applications generate 3-5x more output tokens than input. At the output tier:

GPT-4.1: $8/1M output
Sonnet 4.6: $15/1M output

OpenAI is nearly 2x cheaper on output. If your app is output-heavy (code generation, long-form writing), this matters more than input pricing.

Real-World Cost Examples

Customer Support Bot (50K conversations/month)

Average: 400 input tokens, 200 output tokens per conversation.

Provider	Model	Monthly Cost
OpenAI	GPT-4.1 Nano	$4.00
OpenAI	GPT-4.1 mini	$24.00
Anthropic	Haiku 4.5	$70.00
Anthropic	Sonnet 4.6	$210.00

Winner: GPT-4.1 Nano — if quality is sufficient.

Code Generation (5K requests/month)

Average: 500 input tokens, 1,500 output tokens.

Provider	Model	Monthly Cost
OpenAI	GPT-4.1	$65.00
Anthropic	Sonnet 4.6	$120.00
OpenAI	GPT-5.4	$81.25
Anthropic	Opus 4.7	$200.00

Winner: GPT-4.1 — best code model per dollar.

RAG Pipeline (100K queries/month, heavy caching)

Average: 2,000 input tokens (80% cached), 500 output tokens.

Provider	Model	Monthly Cost (with caching)
Anthropic	Sonnet 4.6 (90% cache discount)	$98.00
OpenAI	GPT-4.1 (50% cache discount)	$120.00
Anthropic	Haiku 4.5 (90% cache discount)	$35.00
OpenAI	GPT-4.1 mini (50% cache discount)	$56.00

Winner: Anthropic — caching discount dominates for RAG.

The Real Strategy: Don't Pick One Provider

The best cost setup uses both. Route by task complexity:

import { CostLens } from 'costlens';
import OpenAI from 'openai';

const costlens = new CostLens({ apiKey: 'cl_...' });
const openai = costlens.wrapOpenAI(new OpenAI());

// CostLens routes simple prompts to cheaper models automatically
// Complex prompts stay on your default model
// You see exactly what each request costs in real-time
const resp openai.chat.completions.create({
  model: 'gpt-4.1',
  messages: [{ role: 'user', content: 'Summarize this document...' }],
});

// Terminal output: ✓ Routed to gpt-4.1-nano (saved $0.003)

Install in 30 seconds:

npm install costlens

No config needed — it tracks costs immediately and suggests cheaper routing opportunities.

Cost Optimization Checklist

Audit your model usage — most teams use GPT-4 for tasks GPT-4.1 Nano handles fine
Enable prompt caching — Anthropic's 90% discount is massive for repetitive prefixes
Use batch API for non-realtime work — 50% off on OpenAI, 25% off on Anthropic
Route by complexity — not every request needs your most expensive model
Track actual spend — pricing pages show rates, not your real bill with retries and long contexts

Bottom Line

Raw pricing winner: OpenAI (cheaper at every tier in 2026)

With heavy caching: Anthropic (90% cache discount beats OpenAI's 50%)

For code generation: GPT-4.1 ($2/$8) — best instruction-following per dollar

For high-volume simple tasks: GPT-4.1 Nano ($0.10/$0.40) — unbeatable

For long-context RAG: Haiku 4.5 with caching — cheapest capable option

Best overall strategy: Use both. Route intelligently. Track everything.

Prices sourced from OpenAI API pricing and Anthropic pricing. Last updated June 2026.

OpenAI vs Anthropic Pricing 2026: Real Cost Per Task Compared