As of July 2026, GPT-4.1 nano is the cheapest mainstream LLM API ($0.10/1M input · $0.40/1M output). Among current-generation models, Gemini 3 Flash ($0.50 / $3.00) leads on price, Claude Sonnet 5 ($2 / $10) leads mid-tier value, and Claude Fable 5 ($10 / $50) is the premium frontier tier.
§1 · The full price table
List prices in USD per 1 million tokens, verified against the official OpenAI, Anthropic and Google pricing pages on 2026-07-11. "Blended 3:1" shows the effective per-1M rate for a typical workload with 3 input tokens per output token — a more honest single number for comparing models than input price alone.
| Model | Input $/1M | Output $/1M | Blended 3:1 | Tier |
|---|---|---|---|---|
| GPT-4.1 nano | $0.10 | $0.40 | $0.18 | budget |
| Gemini 3 Flash | $0.50 | $3.00 | $1.13 | budget |
| Claude Haiku 4.5 | $1.00 | $5.00 | $2.00 | budget+ |
| GPT-5.6 Luna | $1.00 | $6.00 | $2.25 | budget+ |
| Gemini 3.5 Flash | $1.50 | $9.00 | $3.38 | mid |
| Claude Sonnet 5 | $2.00 | $10.00 | $4.00 | mid |
| Gemini 3.1 Pro | $2.00 | $12.00 | $4.50 | mid |
| GPT-5.6 Terra | $2.50 | $15.00 | $5.63 | mid |
| Claude Opus 4.8 | $5.00 | $25.00 | $10.00 | frontier |
| GPT-5.6 Sol | $5.00 | $30.00 | $11.25 | frontier |
| Claude Fable 5 | $10.00 | $50.00 | $20.00 | frontier+ |
Blended 3:1 = (3 × input + 1 × output) ÷ 4. List prices only — batch APIs and prompt caching lower real spend (see §4). Tiers are price bands, not quality rankings.
Want these numbers for your prompt? Paste it and compare all eleven models side by side.
Open token cost calculator →§2 · What real workloads cost per month
Per-token prices are abstract; invoices are not. Three common production shapes, computed exactly from the table above.
Workload A — Support chatbot · 100k calls/mo · 800 in / 250 out per call
Monthly volume: 80M input + 25M output tokens.
| Model | Monthly cost |
|---|---|
| GPT-4.1 nano | $18 |
| Gemini 3 Flash | $115 |
| Claude Haiku 4.5 | $205 |
| GPT-5.6 Luna | $230 |
| Claude Sonnet 5 | $410 |
| Gemini 3.1 Pro | $460 |
| GPT-5.6 Terra | $575 |
| Claude Opus 4.8 | $1,025 |
| Claude Fable 5 | $2,050 |
Workload B — RAG answerer · 10k calls/day · 6,000 in / 400 out
Monthly volume (30 days): 1,800M input + 120M output tokens. Input-heavy: the input rate dominates everything.
| Model | Monthly cost |
|---|---|
| GPT-4.1 nano | $228 |
| Gemini 3 Flash | $1,260 |
| Claude Haiku 4.5 | $2,400 |
| GPT-5.6 Luna | $2,520 |
| Gemini 3.5 Flash | $3,780 |
| Claude Sonnet 5 | $4,800 |
Workload C — Extraction pipeline · 1M docs/mo · 2,000 in / 150 out
Monthly volume: 2,000M input + 150M output tokens — the classic "structure everything" batch job.
| Model | Monthly cost |
|---|---|
| GPT-4.1 nano | $260 |
| Gemini 3 Flash | $1,450 |
| Claude Haiku 4.5 | $2,750 |
Extraction at this volume is exactly where structured outputs pay off: shorter, schema-constrained answers cut the expensive output side. See the JSON Schema builder and structured output validator.
§3 · How to read this table (three rules)
- Output costs 4–6× input, everywhere. GPT-5.6 Sol charges $30 out vs $5 in (6×); Claude Fable 5 charges $50 vs $10 (5×). Generation is sequential and compute-bound; prompt reading is not. The cheapest optimization is almost always making answers shorter, not prompts.
- Input-heavy workloads flip the ranking. Sonnet 5 and Gemini 3.1 Pro tie at $2 input, but on Workload B their bills differ by $1,200/mo because of output rates. Know your in:out ratio before choosing — that is why the table shows a blended rate.
- The frontier premium is 5–111×. Fable 5 costs ~111× GPT-4.1 nano per blended token ($20.00 vs $0.18). Routing — cheap model first, escalate hard cases — is the single highest-leverage cost pattern in 2026 production stacks.
§4 · Below list price: caching, batching, routing
Every number above is list price. Three standard discounts change real invoices:
- Prompt caching. All three providers discount repeated prompt prefixes (system prompts, few-shot blocks, RAG boilerplate) heavily — cached input reads cost a fraction of the normal input rate. If your system prompt is 2k tokens across 100k calls, caching is the difference between paying for 200M tokens and paying for a sliver of that.
- Batch APIs. OpenAI, Anthropic and Google all offer asynchronous batch endpoints at roughly half price for non-interactive jobs. Workload C above is a batch job by nature — run it through a batch API and the nano bill drops from $260 toward $130.
- Model routing. Classify first with a budget model; send only ambiguous cases up a tier. Even a crude 90/10 split between Haiku 4.5 and Sonnet 5 on Workload A prices out at $225.50/mo vs $410 all-Sonnet — a 45% cut with negligible quality loss on the easy 90%.
Exact discount rates change more often than base prices, so we keep them qualitative here — check your provider's pricing page the week you ship, and plug the current numbers into the calculator (every price field is editable).
§5 · FAQ
What is the cheapest LLM API right now (July 2026)?
GPT-4.1 nano at $0.10/1M input and $0.40/1M output. It is a previous-generation small model, so quality-sensitive tasks usually step up to Gemini 3 Flash ($0.50/$3.00), Claude Haiku 4.5 or GPT-5.6 Luna ($1.00 input) — still an order of magnitude below frontier pricing.
Why do output tokens cost more than input tokens?
Generation happens one token at a time and monopolizes accelerator time; input processing parallelizes. Providers price accordingly — 4–6× everywhere in the July 2026 table. Practical consequence: response-length discipline (max-token caps, structured outputs, terse formats) beats prompt-trimming for savings.
Which model should I pick for RAG?
For input-heavy RAG, input price dominates: GPT-4.1 nano and Gemini 3 Flash are the volume picks, Haiku 4.5 the accuracy step-up. Reserve mid/frontier models for the final answer-synthesis call, not for every retrieval pass.
How current are these prices, and how often do they change?
Verified 2026-07-11 against the official OpenAI, Anthropic and Google pricing pages; this page is refreshed monthly (the badge at the top shows the verification date). Providers typically reprice at model launches — several times a year — so confirm before committing to volume contracts.