Research Note · AI · Cost Comparison

Enterprise LLM cost comparison: per-token pricing.

A procurement-grade reference for 2026 LLM economics: published per-token input and output pricing across Anthropic, OpenAI, Google, Meta, Mistral and Cohere, plus the cached-input, batch, reserved-capacity and volume-discount mechanics that decide realised cost. The cheapest model is rarely the right one — the right one is the cheapest that clears the capability bar for the workload.

By James Hill-WoodUpdated May 20269 min readAI / LLM research cluster
Bottom line

There is no single "cheapest" model. The cheapest enterprise-grade frontier model is Claude Haiku 4 at $0.25 / $1.25 per million input / output tokens; the priciest are Claude Opus 4 ($15 / $75) and OpenAI o1 ($15 / $60) — a roughly 300× spread on output. Cached-input and batch discounts of 50–90% reshape realised cost, and the single largest lever is choosing the cheapest model that still clears the workload's capability threshold.

01 Key findings

  1. The price spread across frontier models is enormous. Claude Haiku 4 bills $0.25 input / $1.25 output per million tokens; Claude Opus 4 and OpenAI o1 bill $15 input and $75 / $60 output. Model choice, not vendor loyalty, is the dominant cost variable.

  2. Output tokens dominate. Output costs 4–5× input on most frontier models. A workload that looks cheap on input counts can be expensive on output; constraining output length and right-sizing the model is the largest single optimisation in 2026.

  3. Cached input and batch discounts are structural, not marginal. Anthropic discounts cached input up to 90%, Google 75%, OpenAI 50%; batch APIs run 50% off across vendors. On the right workload shape these change unit economics more than any negotiated discount.

  4. The right model is the cheapest that meets the capability threshold. Not the cheapest model overall. Matching model tier to workload class beats blanket standardisation on a single frontier model.

  5. Reserved capacity beats per-token at sustained throughput. Provisioned Throughput (Azure PTU, Bedrock, Vertex GSU, Anthropic reservations) typically saves 25–45% above ~100–300 tokens/sec of steady-state load.

  6. Hidden layers understate TCO by 15–35%. Embeddings, multimodal input tokens, and orchestration / observability infrastructure are absent from published per-token rates but real on the invoice.

02 Per-token pricing

Published per-token pricing across the frontier and mid-tier families as of Q2 2026. Rates reflect standard real-time API access on the vendor's direct endpoint or primary cloud channel (Bedrock for Llama and Claude on AWS, Vertex for Gemini and Claude on GCP). All figures are USD per 1M tokens.

ModelVendorInput / 1MOutput / 1MContext window
Claude Opus 4Anthropic$15.00$75.00500K
Claude Sonnet 4Anthropic$3.00$15.00500K
Claude Haiku 4Anthropic$0.25$1.25200K
GPT-4oOpenAI$2.50$10.00128K
GPT-4o miniOpenAI$0.15$0.60128K
GPT-4.1OpenAI$2.00$8.001M
GPT-4.1 miniOpenAI$0.40$1.601M
o1 (reasoning)OpenAI$15.00$60.00200K
o3-mini (reasoning)OpenAI$1.10$4.40200K
Gemini 2.0 ProGoogle$1.25$5.002M
Gemini 2.0 FlashGoogle$0.075$0.301M
Gemini 2.0 Flash-LiteGoogle$0.0375$0.151M
Llama 3.1 405B (Bedrock)Meta / AWS$5.32$16.00128K
Llama 3.1 70B (Bedrock)Meta / AWS$0.99$0.99128K
Mistral Large 2Mistral$3.00$9.00128K
Mistral SmallMistral$0.20$0.60128K
Cohere Command R+Cohere$2.50$10.00128K
Cohere Command RCohere$0.15$0.60128K

03 Relative cost

Output tokens are the dominant cost driver, so relative output price per 1M tokens is the clearest single view of where each model sits. Bars are scaled to the most expensive model (Claude Opus 4, $75).

Claude Opus 4
$75.00
OpenAI o1
$60.00
Llama 3.1 405B
$16.00
Claude Sonnet 4
$15.00
GPT-4o
$10.00
Mistral Large 2
$9.00
Gemini 2.0 Pro
$5.00
Claude Haiku 4
$1.25
Gemini 2.0 Flash
$0.30
The output-token bias

A 2,000-token response costs $0.030 on Claude Sonnet 4 ($15/1M) but $0.0025 on Claude Haiku 4 ($1.25/1M). For high-volume workloads where Haiku 4 quality is sufficient, that 12× reduction is the largest single optimisation available in 2026. Constrain output length explicitly and use the most concise model that clears the quality threshold.

04 Cached & batch discounts

Published rates are starting points. Most workloads at scale qualify for cached-input pricing (stable system prompts) or Batch API pricing (no real-time requirement). Discount structures vary materially by vendor.

VendorCached input discountBatch API discountNotes
Anthropic90% off (1-hour cache)50% offCache write costs 125% of standard for 5-minute cache
OpenAI50% off (5-minute cache)50% offImplicit caching; explicit caching also available
Google Gemini75% off (context caching)50% offPer-token cache storage fee applies
AWS Bedrock (Claude, Llama)Per-model (matches direct)50% off via Batch InferenceBedrock Provisioned Throughput separately priced
MistralNot generally available50% offSelf-hosted gives full caching control
CohereLimited50% off via batch jobsPer-tenant negotiation
Cached input in practice

A RAG chat app with a stable 30,000-token system prompt and 500-token user prompts: without caching, every request bills 30,500 input tokens at $3/1M on Claude Sonnet 4 (~$0.0915/request). With a 90% cache discount on the system prompt, each request bills 3,000 cached plus 500 standard input tokens (~$0.011/request) — an 88% input-side reduction. At 100,000 requests/day that compounds to about $300,000/year saved on one workload.

05 Cost drivers

Three levers move realised cost more than headline per-token rates: context (how many input tokens each request carries, and whether they are cacheable), caching / batching (workload shape), and output length (the 4–5× multiplier). Beyond those, three hidden layers understate three-year TCO by 15–35% on most workloads.

Hidden layerTypical magnitudeDetail
Embedding generation<5% of total spendOpenAI text-embedding-3-large $0.13/1M; Voyage voyage-3 $0.18/1M. Small but a discrete line item that surprises buyers who modelled only generation.
Image / multimodal input tokensCan dominate on OCR-heavy workA standard image to GPT-4o is ~765–1,105 tokens by resolution. Document OCR, chart and screenshot workloads can dominate token spend.
Orchestration & observability+15–30% of model spendLangChain / LlamaIndex orchestration, vector database, observability, prompt management and evaluation tooling around the model.

06 Reserved capacity & volume discounts

For sustained high-throughput workloads, per-token billing is not the cheapest option. Vendor-side reserved capacity typically delivers 25–45% cost reduction above 100–300 tokens/sec of sustained throughput.

Reserved capacity SKUPricing modelBreakeven vs per-token
Azure OpenAI Provisioned Throughput UnitPer-PTU per month, reserved capacity~150 tokens/sec sustained
AWS Bedrock Provisioned ThroughputPer-model-unit per hour~200 tokens/sec sustained
Anthropic Capacity ReservationPer-tenant monthly commitNegotiated per use case
Google Vertex Provisioned ThroughputPer-GSU per month~100 tokens/sec sustained

Above roughly $50,000/month in API spend, all the frontier vendors entertain custom commercial terms. Discount realisation tracks committed spend.

Monthly committed spendTypical discount bandNotes
$50K to $100K0 to 8 percentPay-as-you-go, no commit
$100K to $500K8 to 18 percent12-month commit with quarterly true-up
$500K to $2M15 to 25 percentMulti-year commit, custom SLAs
$2M+20 to 35 percentCapacity reservation + bespoke terms

Token economics are usually identical across cloud channels, but the commercial accounting is not. Claude via AWS Bedrock burns an AWS EDP balance; Azure OpenAI burns Microsoft MACC; Vertex burns a GCP enterprise agreement commit. Channel arbitrage is a commercial-structure decision; capability is unaffected.

07 Cost by workload class

The right model for a workload is the cheapest that meets its capability threshold. Representative monthly cost across the primary workload classes at enterprise scale, with the typical model choice for each.

WorkloadVolume assumptionRecommended modelMonthly cost
Enterprise chat assistant (RAG)5M requests / month, 8K tokens avgClaude Sonnet 4 with caching$48K to $72K
Document analysis (legal review)250K documents / month, 50K tokens avgClaude Opus 4 or Gemini 2.0 Pro$94K to $187K
Code generation (developer assistant)800K completions / month, 4K tokens avgClaude Sonnet 4 or GPT-4o$28K to $35K
Classification / extraction at scale200M items / month, 1K tokens avgClaude Haiku 4 or Gemini Flash$8K to $32K
Multi-step reasoning (analysis)50K tasks / month, 20K tokens avgo1 or Claude Opus 4 extended thinking$40K to $115K

08 Model-selection framework

Four factors drive the model decision. Weight them to your workload before committing spend.

Factor 01

Capability threshold

Identify the cheapest model that clears the quality bar for the task, not the cheapest model outright. Over-provisioning capability is the most common source of avoidable LLM spend.

Factor 02

Workload shape

Steady-state, high-throughput workloads favour reserved capacity; bursty or low-throughput ones favour per-token. Mixed workloads reserve the baseline and burst on per-token.

Factor 03

Prompt structure

Stable system prompts unlock cached-input discounts of 50–90%; latency-tolerant work unlocks 50% batch pricing. Design prompts and pipelines to qualify for both.

Factor 04

Commercial channel

Existing cloud commitments change net cost even when token rates match. Route Claude, Llama or GPT through the channel that burns down an EDP, MACC or GCP commit.

09 Which model when

Use Haiku 4 / Flash
High-volume & extraction

Classification, extraction and high-volume routing where Claude Haiku 4 or Gemini 2.0 Flash-Lite quality suffices. This is where the 12× cost gap versus frontier models pays off — keep output short.

Use Sonnet 4 / GPT-4o
Broad chat, RAG & code

Broad-population chat with RAG, developer assistants and multimodal work. Claude Sonnet 4 or GPT-4o with cached input is the cost-efficient default for the majority of enterprise workloads.

Use Opus 4 / o1
Long-doc & reasoning

Long-document analysis (Claude Opus 4 or Gemini 2.0 Pro) and multi-step reasoning (OpenAI o1 or Opus 4 extended thinking). Reserve the premium tier for tasks that genuinely need it. For sovereign deployment, self-host Llama 3.1 or Mistral.

Deeper per-vendor detail lives in our enterprise AI vendor selection framework, ChatGPT Enterprise pricing 2026, Claude Enterprise pricing 2026, OpenAI Enterprise pricing, Gemini Enterprise and Microsoft 365 Copilot pricing 2026. For counsel, see AI procurement advisory, cloud contract negotiation and software licensing advisory.

Cut LLM spend with the right model decision

Independent LLM spend reviews identify a median 32% of consumption as recoverable. We run the review on a fixed fee and produce a 90-day optimisation plan.

AI procurement advisory →

The Licensing Edge

Weekly AI, cloud and licensing intelligence for enterprise IT leaders. 3,000+ subscribers.