LLM Token Pricing Guide: Feb 2026 Edition

LLM pricing in 2026 is a moving target. New models drop monthly. Provider rate cards change quietly. Two months ago, GPT-5.2 was the cheapest frontier model — today, Gemini 3 Flash is half its price for many use cases.

We maintain a live LLM pricing index that updates weekly. This post is a snapshot as of Feb 14, 2026 — for the always-current numbers, use the index.

How LLM pricing works (the 4 things to know)

1. Input vs output pricing. Almost every provider charges 2–5× more for output tokens than input tokens. A workload that summarizes 10K-token docs into 500-token TL;DRs is roughly 75% input cost. A workload that generates code is 60–80% output cost. The math matters.

2. Per-million-token rates. Don't compare $/1K — providers have moved to $/1M to avoid 4-digit pricing. Always normalize to $/1M for cross-vendor math.

3. Context window vs price. A 1M-token context window sounds nice but you pay for every token sent. Use RAG before paying for big contexts — it's almost always cheaper.

4. Caching credits. Anthropic, OpenAI, and Google all now offer prompt caching — repeated prompts at 75–90% discount. Critical for any production RAG or agent loop.

The frontier tier (Feb 2026)

Model	Input $/1M	Output $/1M	Best for
GPT-5.2	$2.50	$10.00	General-purpose, function calling, code
Claude Sonnet 4.5	$3.00	$15.00	Reasoning, long context, tool use
Claude Opus 4.5	$15.00	$75.00	Highest accuracy, complex agents
Gemini 3 Pro	$1.25	$5.00	Multi-modal, very long context

The cheap tier (still good)

Model	Input $/1M	Output $/1M	Best for
Gemini 3 Flash	$0.075	$0.30	Bulk classification, summarization
Claude Haiku 4.5	$0.25	$1.25	Customer support, latency-critical
GPT-4o-mini	$0.15	$0.60	High-volume, function calling

The open tier (host yourself or rent cheap inference)

Model	Cheap-inference $/1M	Best for
Llama 3.3 70B	~$0.20 / $0.40 (Groq, Together)	OSS-equivalent of GPT-4 quality
Llama 3.3 405B	~$2.50 / $4.00	OSS-equivalent of Claude Sonnet
Qwen 2.5 72B	~$0.30 / $0.60	Strong code + Chinese / multilingual
Mixtral 8x22B	~$0.80 / $1.20	MoE, fast, function-calling

The math: when does it matter?

For a 2M-token-per-day workload (a typical mid-stage SaaS chatbot):

Model	Daily cost	Monthly
GPT-5.2	$25	$750
Claude Sonnet 4.5	$36	$1,080
Gemini 3 Pro	$12.50	$375
Gemini 3 Flash	$0.75	$22

Yes, you read that right. Gemini 3 Flash is ~50× cheaper than Claude Sonnet 4.5. For workloads that don't need frontier reasoning (most workloads), this is real money.

The 5 LLM cost levers that always work

1. Cascade routing. Send 80% of traffic to the cheap tier, route only complex requests to the frontier tier. Median savings: 60–80% with no quality drop on user-facing metrics.
2. Prompt caching. Stable system prompts cached at 90% discount means a $1,000/month bill drops to $200/month.
3. Context compression. Drop irrelevant chunks before they hit the LLM. Most RAG pipelines send 4× more context than needed.
4. Aggressive max-tokens. Set max_output_tokens to the actual upper bound, not infinity. Stops runaway loops.
5. Batch processing. Most providers offer 50% discount on async batches with 24-hour SLA. Use it for any non-interactive workload.

What we publish at CARTIE

The live LLM pricing index — every model, every provider, updated weekly, with a free cost calculator. No signup, no auth. Bookmark it.

If you want CARTIE to continuously route your LLM traffic across providers and minimize $/query in real time, that's our paid product (the LLM cost router). Typical first-month savings: 40–70%.

---

Related: Pinecone Cost Optimization: 7 Patterns That Cut Vector-DB Bills 50%.