LLM cost trends 2025–2026

Updated 12 July 2026 · first published 4 July 2026

The LLM market is moving in two directions at once. Input and output token prices for GPT-4-class models have dropped 80–90% since early 2024. At the same time, reasoning models (o1, o3, DeepSeek-R1, Claude Opus 4) are 5–30x more expensive per task than their non-reasoning equivalents. The net effect: teams that adopt reasoning models see bills go up even as unit prices go down.

This article tracks the key price and quality events from 2025 through mid-2026, explains the forces driving costs, and forecasts what to expect in the second half of 2026.

Key trend 1: GPT-4-class models dropped 90% in 18 months

In March 2024, GPT-4-turbo cost $10 per million input tokens and $30 per million output tokens. By mid-2026, GPT-4.1 costs $2 per million input and $8 per million output - a 75–80% drop. Claude followed a similar trajectory: Claude 3 Opus at $15/$75 dropped to Claude Sonnet 4 at $3/$15. Open-weight models from Mistral, Llama, and DeepSeek pushed prices even lower, with some available for under $0.50 per million input tokens through hosted providers.

This is not just a price war. Inference infrastructure improved: quantization, speculative decoding, KV-cache optimization, and batching all reduced the cost to serve a token. Providers passed some of those savings to customers.

Key trend 2: Reasoning models are 5–30x more expensive

Reasoning models (o1, o3, o4-mini, DeepSeek-R1, Claude Opus 4 with extended thinking) generate internal "thinking" tokens before producing a final answer. These thinking tokens are billed as output tokens. A reasoning model solving a coding task might generate 10K–50K thinking tokens before producing a 2K-token answer. The effective cost per task is therefore much higher than the per-token price suggests.

For teams using reasoning models for every task, this is the single biggest cost driver in 2026. The key question is not "how much does the model cost?" but "does this task need reasoning?" Most tasks do not.

Key trend 3: Open-source closed the quality gap

In early 2024, open-weight models trailed proprietary models by a significant margin on coding and reasoning benchmarks. By mid-2026, models like Llama 4 Maverick, DeepSeek-V3, Qwen 3, and MiMo are within a few points of GPT-4-class performance on most benchmarks. This matters for cost because open-weight models are cheaper to serve - providers competing on price use them to undercut proprietary pricing.

The quality gap has not fully closed for frontier reasoning tasks. But for the majority of production workloads - summarization, extraction, classification, code generation, tool use - open-weight models are good enough and significantly cheaper.

Key trend 4: Caching reduced effective costs 50–90% for repeat workloads

Prompt caching (Anthropic) and automatic context caching (OpenAI, Google) changed the economics of repeated workloads. If you send the same system prompt and context to a model 100 times a day, cached reads cost 50–90% less than uncached reads. For RAG pipelines, agent systems, and batch processing, this is transformative.

Caching benefits are unevenly distributed. Workloads with stable context (same system prompt, same document set) benefit enormously. Workloads with dynamic context (user-specific conversations, real-time data) benefit less. FinOps teams should measure their cache hit rate and optimize context stability to maximize savings.

Key trend 5: Agent workflows multiply token consumption

Agent workflows - where models call tools, read files, execute code, and iterate on results - consume 5–50x more tokens per task than single-turn interactions. A coding agent solving a bug might consume 100K–500K tokens across its tool calls, retries, and context accumulation. An equivalent single-turn prompt might use 2K–5K tokens.

This is the most important cost trend in 2026. Token prices are falling, but token consumption per task is rising faster. Net result: total LLM spend is going up for most teams even as unit prices go down.

Timeline: key events

What to expect in H2 2026

Reasoning model prices will drop 30–50%. Competition from open-source reasoning models (DeepSeek-R2, Qwen 3 reasoning variants) will force proprietary providers to cut prices on their reasoning tiers.

Agent-specific pricing will emerge. Providers are testing per-task and per-session pricing models that bundle tool calls, retries, and context into a single predictable cost. Expect at least one major provider to offer this by Q4 2026.

Caching will become automatic. Manual cache management will disappear. Providers will optimize cache placement transparently, and cache hit rates will improve for most workloads without developer intervention.

Open-weight models will dominate routine work. By end of 2026, expect 70%+ of production inference on routine tasks (classification, extraction, summarization, simple code generation) to run on open-weight models through hosted providers.

Total spend will keep rising. Despite falling unit prices, agent-driven workflows and expanding AI use cases mean total LLM spend will continue to grow. FinOps discipline - not just cheaper models - is what keeps spend under control.

Related

• Agent economics - how coding agents price out in practice.
• Provider arbitrage - switching providers to reduce costs.
• Model routing - sending tasks to the right model.
• FinOps for LLM - the broader framework.

Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →

Back to research