TL;DR — March 2026 was the most eventful month for LLM pricing in over a year. OpenAI launched GPT-5.2 at $1.75/$14.00 per million tokens (cheaper input than GPT-4o), Google slashed Gemini 3 Flash to $0.075/$0.30, DeepSeek V3 dropped to $0.27/$1.10, and Anthropic held Claude pricing steady. The net effect: the price gap between cheapest and most expensive models widened to 250x, making smart routing more valuable than ever. Teams using ClawRouters' automatic routing are saving 60-80% by dynamically switching between these newly repriced models.
March 2026 brought a wave of LLM pricing changes that reshaped the cost landscape for AI-powered applications. If you're running production workloads against any major model provider, at least one of these changes affects your bill. This article catalogs every significant pricing shift from March 2026, explains what drove each change, and shows you how to adapt your cost strategy.
For full model-by-model pricing details, see our complete LLM API pricing guide. For broader trends, read our analysis of AI language model pricing trends in 2026.
OpenAI: GPT-5.2 Launch Reshuffles the Lineup
GPT-5.2 Pricing Breakdown
OpenAI's biggest March move was the general availability release of GPT-5.2, priced at $1.75 per million input tokens and $14.00 per million output tokens. The pricing structure caught many developers off guard: input tokens are actually 30% cheaper than GPT-4o ($1.75 vs. $2.50), while output tokens carry a 40% premium ($14.00 vs. $10.00).
| Model | Input (/1M tokens) | Output (/1M tokens) | Change from Feb 2026 | |-------|--------------------|--------------------|---------------------| | GPT-5.2 | $1.75 | $14.00 | New model | | GPT-4o | $2.50 | $10.00 | No change | | GPT-4o-mini | $0.15 | $0.60 | No change |
What this means for your costs: GPT-5.2 is cheaper than GPT-4o for short-answer tasks (where input tokens dominate) but more expensive for reasoning-heavy workflows that generate long outputs. For a balanced 1:1 input-to-output ratio, GPT-5.2 costs $15.75 per million tokens vs. GPT-4o's $12.50 — a 26% increase for frontier capability.
GPT-4o and GPT-4o-mini Hold Steady
OpenAI kept GPT-4o at $2.50/$10.00 and GPT-4o-mini at $0.15/$0.60. With GPT-5.2 positioned above them, these models remain the workhorses for teams that need reliability without frontier pricing. GPT-4o-mini in particular continues to be one of the best price-performance options for high-volume, lower-complexity tasks.
Google: Gemini 3 Flash Hits Rock-Bottom Pricing
Flash Model Price Cut Details
Google made the most aggressive pricing move of March 2026 by dropping Gemini 3 Flash to $0.075 per million input tokens and $0.30 per million output tokens. This represents a roughly 50% reduction from its January 2026 launch price and positions Flash as the cheapest commercially available model from a major provider.
| Model | Input (/1M tokens) | Output (/1M tokens) | March Change | |-------|--------------------|--------------------|-------------| | Gemini 3 Pro | $1.25 | $5.00 | No change | | Gemini 3 Flash | $0.075 | $0.30 | ~50% cut |
At $0.30 per million output tokens, Gemini 3 Flash is now 250x cheaper than Claude Opus 4 and 47x cheaper than GPT-4o. For simple tasks — classification, extraction, summarization, yes/no routing — the quality difference between Flash and premium models is negligible, but the cost difference is enormous.
Impact on High-Volume Workloads
Teams processing millions of tokens daily see the biggest benefit. A workload generating 50 million output tokens per month now costs just $15 with Gemini 3 Flash, compared to $500 with GPT-4o or $3,750 with Claude Opus 4. This pricing makes it economically viable to use LLMs for tasks that were previously too expensive at scale, such as real-time content moderation, document preprocessing, and automated tagging.
For a guide to routing high-volume requests to the cheapest capable model, see how to reduce LLM API costs.
Anthropic: Claude Pricing Unchanged, But the Context Matters
Why Anthropic Held Prices
Anthropic made no pricing changes to Claude models in March 2026. Claude Opus 4 remains at $15.00/$75.00, Sonnet 4 at $3.00/$15.00, and Haiku 3.5 at $0.25/$1.25. This stability is deliberate: Anthropic is positioning its models on capability and safety rather than competing on price.
| Model | Input (/1M tokens) | Output (/1M tokens) | March Change | |-------|--------------------|--------------------|-------------| | Claude Opus 4 | $15.00 | $75.00 | No change | | Claude Sonnet 4 | $3.00 | $15.00 | No change | | Claude Haiku 3.5 | $0.25 | $1.25 | No change |
What Unchanged Claude Pricing Means for Routing
With competitors cutting prices around them, Anthropic's static pricing has two effects. First, Claude Opus 4 is now more of a premium outlier — justified only for tasks where its reasoning quality measurably exceeds alternatives. Second, Claude Sonnet 4 at $3.00/$15.00 faces increasing pressure from GPT-5.2 ($1.75/$14.00) and Gemini 3 Pro ($1.25/$5.00) for general-purpose work.
This is exactly the scenario where smart LLM routing delivers maximum value: routing only the hardest 10% of requests to Opus while sending the other 90% to cheaper models that perform equally well for simpler tasks.
DeepSeek and Open-Source: The Budget Tier Gets Cheaper
DeepSeek V3 and R1 Pricing Updates
DeepSeek continued its aggressive pricing strategy in March 2026. DeepSeek V3 settled at $0.27/$1.10 per million tokens, while DeepSeek R1 (their reasoning-focused model) sits at $0.55/$2.19. Both models deliver quality that rivals models 5-10x their price on coding and mathematical tasks.
| Model | Input (/1M tokens) | Output (/1M tokens) | March Change | |-------|--------------------|--------------------|-------------| | DeepSeek V3 | $0.27 | $1.10 | Slight decrease | | DeepSeek R1 | $0.55 | $2.19 | No change | | Llama 3.3 70B (hosted) | $0.18 | $0.40 | No change | | Mistral Small 3 | $0.10 | $0.30 | No change |
Open-Source Model Hosting Costs
Meta's Llama 3.3 70B and Mistral Small 3 held at $0.18/$0.40 and $0.10/$0.30 respectively through hosted providers. These models' existence continues to put downward pressure on commercial API pricing across the entire market. For teams with privacy requirements or regulatory constraints, self-hosted open-source models offer a fixed-cost alternative to per-token API pricing.
The March 2026 Pricing Landscape at a Glance
Complete Price Comparison After March Changes
Here's the full picture after all March 2026 pricing changes:
| Provider | Model | Input (/1M) | Output (/1M) | March Status | |----------|-------|-------------|-------------|-------------| | Anthropic | Claude Opus 4 | $15.00 | $75.00 | Unchanged | | OpenAI | GPT-5.2 | $1.75 | $14.00 | New | | Anthropic | Claude Sonnet 4 | $3.00 | $15.00 | Unchanged | | OpenAI | GPT-4o | $2.50 | $10.00 | Unchanged | | Google | Gemini 3 Pro | $1.25 | $5.00 | Unchanged | | DeepSeek | DeepSeek R1 | $0.55 | $2.19 | Unchanged | | DeepSeek | DeepSeek V3 | $0.27 | $1.10 | Decreased | | Google | Gemini 3 Flash | $0.075 | $0.30 | ~50% cut | | OpenAI | GPT-4o-mini | $0.15 | $0.60 | Unchanged | | Mistral | Mistral Small 3 | $0.10 | $0.30 | Unchanged |
The output price range now spans from $0.30 (Gemini 3 Flash) to $75.00 (Claude Opus 4) — a 250x spread. This is the widest the gap has ever been, up from approximately 100x just twelve months ago.
Cost Impact by Workload Type
How these changes affect real-world monthly costs (assuming 10 million output tokens/month):
| Workload Type | Best Model Pre-March | Best Model Post-March | Monthly Savings | |---------------|---------------------|----------------------|----------------| | Simple Q&A, classification | GPT-4o-mini ($6) | Gemini 3 Flash ($3) | 50% | | General coding, writing | GPT-4o ($100) | GPT-5.2 or Sonnet 4 ($140-150) | +40% (pay more for frontier) | | Complex reasoning | Claude Opus 4 ($750) | Claude Opus 4 ($750) | 0% | | Mixed workload (smart routed) | ~$45 | ~$30 | 33% |
The biggest winner from March's changes? Mixed workloads using smart routing. With the cheapest models getting even cheaper, a router like ClawRouters that sends simple requests to Flash and reserves premium models for complex tasks now delivers even more dramatic savings.
How to Adapt Your Cost Strategy After March 2026
Audit Your Model Distribution
The first step is understanding where your tokens are going. If more than 20% of your requests hit a frontier model (Opus 4, GPT-5.2), you're likely overspending. According to ClawRouters usage data, approximately 60% of production API requests are simple enough for budget-tier models.
Run a cost analysis by task type:
- Classify your requests — What percentage are simple (classification, extraction, Q&A) vs. complex (multi-step reasoning, long-form generation)?
- Map to optimal models — Simple tasks → Gemini Flash or GPT-4o-mini. Moderate tasks → Sonnet 4 or GPT-4o. Complex → Opus 4 or GPT-5.2.
- Calculate potential savings — The gap between your current single-model cost and an optimally routed cost is your savings opportunity.
Implement Smart Routing to Capture the Wider Price Gap
With the price gap at 250x and growing, manual model selection can't keep up. LLM routing automates the classification-and-routing process, ensuring every request hits the cheapest model capable of handling it.
ClawRouters makes this a one-line change — just swap your base_url:
import openai
client = openai.OpenAI(
base_url="https://api.clawrouters.com/v1",
api_key="your-clawrouters-key"
)
# model="auto" routes each request to the optimal model
response = client.chat.completions.create(
model="auto",
messages=[{"role": "user", "content": "Classify this support ticket..."}]
)
# → Routed to Gemini 3 Flash ($0.30/M) instead of Sonnet 4 ($15/M)
No SDK changes. Works with any OpenAI-compatible client. The free tier lets you test with your own API keys — see the setup guide for a 5-minute integration.
Re-Evaluate GPT-5.2 for Your Use Cases
GPT-5.2's unusual pricing structure (cheap input, expensive output) means it's not a universal upgrade over GPT-4o. Test it on your actual workloads:
- Good fit for GPT-5.2: Tasks with long prompts and short answers (classification, entity extraction, scoring)
- Better to stay on GPT-4o: Tasks that generate long outputs (content writing, code generation, detailed analysis)
- Best approach: Let a model router dynamically choose between GPT-4o and GPT-5.2 based on each request's characteristics
Frequently Asked Questions
What were the biggest LLM pricing changes in March 2026?
The three biggest changes were: (1) OpenAI launched GPT-5.2 at $1.75/$14.00 per million input/output tokens, (2) Google cut Gemini 3 Flash pricing by approximately 50% to $0.075/$0.30, and (3) DeepSeek V3 decreased slightly to $0.27/$1.10. Anthropic's Claude models and most other models remained unchanged.
Is GPT-5.2 cheaper than GPT-4o?
GPT-5.2 has cheaper input tokens ($1.75 vs. $2.50 per million) but more expensive output tokens ($14.00 vs. $10.00 per million). For tasks with short outputs, GPT-5.2 can be cheaper. For output-heavy workloads, GPT-4o remains more cost-effective. The optimal choice depends on your specific input-to-output ratio.
What is the cheapest LLM API available after March 2026 pricing changes?
Gemini 3 Flash is now the cheapest major-provider LLM API at $0.075 per million input tokens and $0.30 per million output tokens. Mistral Small 3 is comparably priced at $0.10/$0.30. Both deliver strong quality for simple to moderate tasks like classification, extraction, and summarization.
How much can smart routing save after March 2026 price changes?
Smart routing through platforms like ClawRouters can save 60-80% on LLM API costs by automatically sending each request to the cheapest capable model. With the price gap between budget and frontier models now at 250x, the savings from proper routing are larger than ever. A mixed workload that would cost $13,500/month on a single mid-range model can drop to under $3,000 with smart routing.
Did Anthropic change Claude pricing in March 2026?
No. Anthropic kept all Claude model pricing unchanged in March 2026: Opus 4 at $15/$75, Sonnet 4 at $3/$15, and Haiku 3.5 at $0.25/$1.25 per million tokens. Anthropic appears to be competing on model capability and safety rather than price, making routing to Claude selectively (only for tasks that need its strengths) the most cost-efficient strategy.
Should I switch from GPT-4o to GPT-5.2?
Not automatically. GPT-5.2 offers improved reasoning capabilities but costs more for output-heavy tasks. The best approach is to benchmark both models on your actual workloads and use an LLM router to dynamically select the better option per request. ClawRouters' auto-routing handles this automatically, choosing between GPT-4o and GPT-5.2 based on task complexity.
How often do LLM API prices change?
Major pricing changes occur every 6-8 weeks on average across providers. March 2026 was particularly eventful due to multiple simultaneous updates. To stay current, bookmark our LLM API pricing guide, which we update as changes happen, or use ClawRouters' model dashboard for real-time pricing data.