The 2026 LLM Pricing Landscape
Last updated: July 2026 — pricing changes weekly; treat this as a snapshot, not gospel.
Frontier model prices have fallen roughly 10× in 18 months while raw capability keeps climbing. This is the "cost-of-intelligence curve" — and it changes TokenOps math every quarter.
Per-1M-token pricing snapshot
| Model | Input | Cached Input | Output | Notes |
|---|---|---|---|---|
| GPT-5 | $1.25 | $0.125 (90% off) | $10 | 400K context; ~55% cheaper than GPT-4o at launch |
| GPT-5 Mini | $0.25 | proportional | $2.00 | Default for routine tasks |
| GPT-5 Nano | $0.05 | proportional | $0.40 | Classification, extraction, routing |
| Claude Opus 4.5 | $5 | $0.50 hit / $6.25 5-min write / $10 1-hr write | $25 | 67% price cut vs Opus 4.1 |
| Claude Sonnet 5 (intro) | $2 | $0.20 hit | $10 | Reverts to $3/$15 after Aug 31 2026 |
| Claude Haiku 4.5 | $1 | $0.10 hit | $5 | Cheapest current-gen Claude |
| Gemini 3 Pro | $2–3.50 (context-tiered) | implicit, automatic | $12–14 | 1M context; implicit caching on 2.5+ |
| DeepSeek V3.2 | $0.28 | $0.028 (90% off) | $0.42 | Open-weights; aggressive cache pricing |
| Llama 4 (hosted) | $0.10–0.40 | host-dependent | $0.40–1.20 | Together / Groq / Fireworks / Bedrock |
The tokenizer inflation trap
Different provider tokenizers split the same text into different numbers of tokens — sometimes materially so. A model whose per-token price is 40% lower can still be more expensive per task if its tokenizer emits meaningfully more tokens on your workload, or if it produces longer outputs by default.
Rule: benchmark cost per task, not cost per token. Measure token counts empirically on a representative sample of your traffic before locking in a "cheaper" model.
What the price war means for architecture
- Assume prices halve every 6–12 months. Do not lock multi-year budgets to today's per-token rate; lock to value per useful output.
- Nano/mini tiers are now serious infrastructure. GPT-5 Nano at $0.05/M input makes previously uneconomic ideas (classify every log line, embed every event) viable.
- Cached input is the new base rate. With 90% cache discounts on GPT-5 and DeepSeek, un-cached prompts are the anomaly to justify — not the default.
- Reasoning tiers are the new premium SKU. Frontier non-reasoning pricing is commoditising; reasoning/thinking modes are where vendors preserve margin.
Sources
- OpenAI GPT-5 pricing announcements
- Anthropic Claude Opus 4.5 launch (InfoWorld coverage, Nov 2025)
- Google Gemini 3 Pro documentation
- DeepSeek V3.2 launch (VentureBeat, Sept 2025)
- llm-stats.com pricing tracker