Danielle Hoopes7 pieces8 min
The AI Price War Is Getting Ridiculous (And Developers Win)
A year ago, I was paying $15 per million input tokens for GPT-4. Last week, Google dropped Gemini 3.1 Flash-Lite at $0.25 per million.
That's a 98% price drop. In twelve months.
If your grocery bill did that, you'd be eating wagyu every night.
The AI API price war officially went nuclear in March 2026, and if you're a developer building anything with LLMs, you need to understand what just happened. Because the model you're paying for right now? It's probably already overpriced.
1 min read
What Happened This Month
March 2026 has been absolute chaos. Here's the highlight reel:
OpenAI dropped GPT-5.4 with a 1M token context window at $2.50/M input tokens. 33% fewer hallucinations than GPT-5.2. They're clearly feeling the heat.
Google countered with Gemini 3.1 Flash-Lite at $0.25/M input tokens. That's 10x cheaper than GPT-5.4 for input. And it's 2.5x faster.
Anthropic is sitting at $3/M input tokens for Claude Sonnet 4.6, but just shipped multi-agent code review and an agent SDK that makes the per-token cost less relevant when you're measuring by task completed.
DeepSeek released V4 with 1 trillion parameters (32B active) and their pricing continues to undercut everyone from China.
Meanwhile, 12+ models launched in a single week. I lost count.
1 min read
The Price Collapse Timeline
Here's how fast this happened:
| Date | Model | Input Cost (per 1M tokens) | Context Window |
|---|---|---|---|
| Mar 2024 | GPT-4 Turbo | $10.00 | 128K |
| Jun 2024 | Claude 3.5 Sonnet | $3.00 | 200K |
| Dec 2024 | Gemini 2.0 Flash | $0.10 | 1M |
| Jul 2025 | GPT-5 | $5.00 | 256K |
| Oct 2025 | Claude 3.5 Opus | $15.00 | 200K |
| Jan 2026 | DeepSeek V3 | $0.27 | 128K |
| Mar 2026 | GPT-5.4 | $2.50 | 1M |
| Mar 2026 | Gemini 3.1 Flash-Lite | $0.25 | 1M |
| Mar 2026 | Claude Sonnet 4.6 | $3.00 | 1M |
Two years ago, a million tokens cost you $10+. Today you can get it for a quarter. Literally 25 cents.
2 min read
Head-to-Head: Who Costs What (March 2026)
Here's the current landscape for the models developers actually use:
Flagship Models
| Model | Input (/1M) | Output (/1M) | Context | Best For |
|---|---|---|---|---|
| Claude Opus 4.6 | $15.00 | $75.00 | 1M | Complex reasoning, code architecture |
| GPT-5.4 | $2.50 | $10.00 | 1M | General purpose, large context tasks |
| Gemini 3.1 Pro | $1.25 | $5.00 | 2M | Multimodal, massive context |
Mid-Tier (Best Bang for Buck)
| Model | Input (/1M) | Output (/1M) | Context | Best For |
|---|---|---|---|---|
| Claude Sonnet 4.6 | $3.00 | $15.00 | 1M | Coding, analysis, daily driver |
| GPT-5.4 Mini | $0.40 | $1.60 | 1M | Chat, summarization |
| Gemini 3.1 Flash | $0.075 | $0.30 | 1M | High-volume, cost-sensitive |
Budget Tier
| Model | Input (/1M) | Output (/1M) | Context | Best For |
|---|---|---|---|---|
| Claude Haiku 4.5 | $0.80 | $4.00 | 200K | Fast responses, classification |
| Gemini 3.1 Flash-Lite | $0.25 | $1.00 | 1M | Bulk processing, prototyping |
| DeepSeek V4 | $0.27 | $1.10 | 128K | Budget builds, experimentation |
The spread is wild. You could pay $75/M output tokens for Opus, or $0.30/M for Gemini Flash. That's a 250x difference.
1 min read
Which Model Should You Actually Use?
Stop overthinking it. Here's the decision framework:
Building a prototype or side project? Use Gemini 3.1 Flash-Lite or DeepSeek V4. You'll spend pennies. Seriously. A weekend hackathon might cost you $0.50 in API calls.
Production app with moderate traffic? Claude Sonnet 4.6 or GPT-5.4. Both are excellent. Claude is better at coding tasks. GPT-5.4 is more general purpose. Pick based on your use case, not the price.
Need the absolute best output quality? Claude Opus 4.6 for reasoning and code. GPT-5.4 for broad knowledge tasks. Yes, Opus is expensive. But if you're using it for architecture decisions or complex analysis, the quality gap is real.
High-volume, cost-sensitive production? Gemini Flash. The $0.075/M input cost is absurdly cheap for a model this capable. If your app processes thousands of requests daily, the cost difference adds up fast.
My actual setup: I use Claude Sonnet 4.6 as my daily driver for coding, Opus when I need to think hard about architecture, and Gemini Flash for bulk tasks like processing logs or summarizing large documents. Monthly API bill went from ~$200 to ~$60 after switching to this tiered approach.
1 min read
Why Prices Are Crashing
Three things are happening at once:
1. The DeepSeek Effect
When DeepSeek V3 dropped at $0.27/M input tokens in January 2026 with genuinely competitive quality, it proved that frontier-level performance doesn't require frontier-level pricing. Every other provider had to respond or lose developers.
2. Inference Optimization
The models are getting cheaper to run. Mixture-of-experts architectures (like DeepSeek's 1T total / 32B active setup) mean you only pay for the compute you actually use. NVIDIA's new inference chips from GTC are pushing this further.
3. Market Share Land Grab
OpenAI, Google, and Anthropic are all fighting for developer lock-in. If you build your app on their API, switching costs are real. So they're racing to the bottom on price to win you now. Classic platform strategy.
1 min read
What This Means for Indie Developers
This is the part that excites me.
A year ago, building an AI-powered app meant either eating a massive API bill or gatekeeping features behind expensive subscription tiers.
Now? The math has completely changed.
Real example: An app that processes 10,000 user requests per day, each using ~2,000 tokens input and ~1,000 tokens output:
| Model | Daily Cost | Monthly Cost |
|---|---|---|
| GPT-4 (Mar 2024) | $350 | $10,500 |
| Claude Sonnet 4.6 | $90 | $2,700 |
| GPT-5.4 | $60 | $1,800 |
| Gemini 3.1 Flash | $3 | $90 |
From $10,500/month to $90/month for the same task. That's not an optimization. That's a different business model entirely.
Indie developers can now build AI features that would have required VC funding a year ago. You can run a profitable AI app on revenue from a few hundred users.
1 min read
The Catch (There's Always a Catch)
Before you go rewrite everything on the cheapest model:
Cheap models have real quality gaps. Flash-Lite is great for bulk work but will fumble complex reasoning. Don't use a budget model for tasks where accuracy matters. Test on your actual use case, not benchmarks.
Rate limits vary wildly. The cheapest tiers often come with aggressive rate limiting. Google's free tier is generous but throttles hard under load. Check the limits before you architect around a model.
Pricing can change. These companies are subsidizing prices to win market share. When the war ends (or when your app is locked in), prices could go up. Build your app to be model-agnostic if possible. Use abstraction layers.
Context window size isn't everything. A 1M token context window sounds amazing until you realize that stuffing 800K tokens of context into every request at $2.50/M adds up fast. Be smart about what you send.
Output quality per dollar is the real metric. Don't compare input prices alone. A model that costs 2x more but needs half the retries is actually cheaper. Track your actual cost-per-successful-task, not cost-per-token.