LLM Pricing Comparison

By the TokenForge team · Last updated July 2026

Every provider quotes two prices — input and output per million tokens — and neither tells you what a model actually costs to run. This comparison lists both prices for 16 production models, adds a blended 4:1 price you can compare directly, and flags the caveats the pricing pages leave out.

LLM API pricing table (July 2026)

Prices last reviewed July 2026 — refreshed daily from the same feed as the calculator.

ProviderModelInput $/1MOutput $/1MBlended 4:1Best forWatch out for
OpenAIGPT-5.6 Sol$5.00$30.00$10.00Frontier reasoning and agentic codingPremium output price adds up fast on long answers
OpenAIGPT-5.6 Terra$2.50$15.00$5.00Strong default for production chat2.5× Luna's price for gains you may not need
OpenAIGPT-5.6 Luna$1.00$6.00$2.00High-volume chat with solid qualityWeaker on multi-step reasoning than Terra and Sol
OpenAIGPT-5.4 Mini$0.75$4.50$1.50Cost-sensitive assistants and summarizationPrevious generation — benchmark before committing
OpenAIGPT-5.4 Nano$0.20$1.25$0.41Classification, extraction, routingA workhorse, not a reasoner — keep tasks narrow
AnthropicClaude Fable 5 †$10.00$50.00$23.40The hardest reasoning and long-horizon agent workHighest list price here, plus ~30% tokenizer overhead
AnthropicClaude Opus 4.8 †$5.00$25.00$11.70Premium coding and analysis below Fable's priceTokenizer overhead applies (~30% more tokens)
AnthropicClaude Sonnet 5 †$2.00$10.00$4.68Balanced quality and cost for production agentsCorrected blended cost sits closer to Terra than list price suggests
AnthropicClaude Haiku 4.5$1.00$5.00$1.80Fast, cheap chat with Claude behaviourQuality gap versus Sonnet shows on complex tasks
GoogleGemini 3.5 Flash$1.50$9.00$3.00Long-context and multimodal work at a mid priceOutput at $9.00 is pricier than the input price suggests
GoogleGemini 3.1 Pro$2.00$12.00$4.00Long-document analysis and grounded answersListed price applies up to 200K-token contexts
GoogleGemini 2.5 Flash-Lite$0.10$0.40$0.16Bulk cheap tasks: tagging, drafts, filteringOlder lightweight model — expect quality limits
DeepSeekDeepSeek V4 Flash$0.14$0.28$0.17The cheapest capable chat model per tokenWeigh data-residency and compliance requirements
DeepSeekDeepSeek V4 Pro$0.44$0.87$0.52Budget reasoning far below Big-3 pricesThroughput and rate limits vary by region and load
Open weightsLlama 3.3 70B (Together AI)$0.88$0.88$0.88Open weights with identical in/out pricingYou can also self-host it — compare the break-even first
Open weightsgpt-oss-120B (Together AI)$0.15$0.60$0.24Cheap open-weights generalist via hosted endpointsNewer model line — benchmark on your own tasks

† Anthropic's newest models tokenize the same text into roughly 30% more tokens, so their blended column includes that correction; the raw input/output columns are list prices as published. All prices are for standard (non-batch) API usage and exclude prompt-caching discounts and volume deals. Providers change prices without notice — treat the live-refreshed table above as the source of truth, not screenshots of it.

Why output tokens cost more than input tokens

Reading is cheap; writing is expensive. When a model ingests your prompt, it processes all input tokens in one parallel pass (the prefill). Generating the answer is sequential: each output token requires a full forward pass through the model, while your entire context sits in GPU memory for the duration. One output token therefore costs the provider many times the compute of one input token — and the price sheet mirrors that, with output typically 2–6× input (6× for GPT-5.6 Sol and Gemini 3.5 Flash, 5× for Claude models, but only 2× for DeepSeek V4 Flash).

The practical consequence: your input:output mix changes which model is cheapest. Input-heavy workloads — RAG pipelines that stuff thousands of context tokens per request, classification, document analysis — should weight the input column. Output-heavy workloads like long-form writing should weight the output column, where the spread between providers is widest. If you're building retrieval, the RAG Cost Calculator does this split for you.

How to compare models fairly: the blended price

Most production traffic is input-heavy — around four input tokens for every output token once you count system prompts, history and context. Weighting the two prices at that 4:1 ratio gives a single comparable number per model: the blended column in the table. If your real mix differs, re-weight accordingly; the LLM Token Cost Calculator has a ratio input that does the arithmetic live, against the same daily price feed.

Worked example. GPT-5.6 Terra at 4:1 blends to (4 × $2.50 + $15.00) ÷ 5 = $5.00 per million tokens. Claude Sonnet 5 blends to (4 × $2.00 + $10.00) ÷ 5 = $3.60 on paper — but Anthropic's tokenizer splits the same text into ~30% more tokens, so the effective blended price is $3.60 × 1.3 = $4.68. List prices make Sonnet look 28% cheaper than Terra; the corrected number narrows the gap to about 6%.

Which model for which job

Frontier tier ($10–$23 blended). GPT-5.6 Sol and Claude Fable 5 are for work where a wrong answer costs more than the tokens: complex agent pipelines, hard code generation, high-stakes analysis. Nobody should run bulk traffic here — the winning pattern routes only the hardest 5–10% of requests up to this tier.

Production tier ($2–$5 blended). GPT-5.6 Terra, Claude Sonnet 5, Gemini 3.1 Pro and GPT-5.6 Luna is where most products should land: strong general quality at single-digit blended prices. Within the tier the differences are workload-specific — run a week of your real traffic through the OpenAI, Claude and Gemini calculators before you decide.

Budget tier (under $2 blended). GPT-5.4 Mini and Nano, Claude Haiku 4.5, Gemini 2.5 Flash-Lite and the DeepSeek models handle classification, extraction, routing, and simple chat at a tenth of production-tier prices. The trap is quality drift on tasks that quietly got harder — keep an evaluation set and re-test.

Open weights. Hosted endpoints like Together AI price Llama 3.3 70B at a flat $0.88 and gpt-oss-120B at $0.24 blended — and the same weights can run on your own GPU. Whether that's worth it is a volume question: the self-hosting break-even guide walks through the math.

Frequently asked questions

Related tools and guides