Cheapest LLM API (July 2026): The Floor-Tier Table, Cost Per 1,000 Requests, and When Cheap Gets Expensive

TL;DR: There is no single cheapest LLM API, and any page naming one is hiding the split. On our July 10, 2026 snapshot, GPT-5-nano owns the input rate ($0.05 per 1M tokens), DeepSeek V4 Flash owns output ($0.18), and Gemini 2.5 Flash-Lite owns context per dollar ($0.10/$0.40, 1M window). On one fixed chatbot job the whole floor tier lands between $0.35 and $0.50 per 1,000 requests. The sticker is the easy part. The hard part is the five ways a cheap model quietly re-inflates the bill.

Search "cheapest llm api" and you get two kinds of pages: raw price grids with no verdict, and listicles quoting rates that went stale two model releases ago. This page does neither. Every price below comes from the same July 10, 2026 snapshot that powers our live LLM price tracker , which is the canonical table this page inherits from and the place to check when these numbers age out. Then it does what a price grid cannot: it prices one identical job across the floor tier, names the quality cliff below the cheap models, and walks through the multipliers that make "cheap" the most expensive line item on some invoices.

Disclosure: we have no affiliate or paid relationship with OpenAI, Google, DeepSeek , Meta, Anthropic, xAI, or any hosting provider named on this page as of publication. If that changes, this paragraph will say so. We read prices from OpenRouter 's public model feed, cross-checked against vendor pricing pages; how that feed works is documented in our OpenRouter review .

The cheapest LLM APIs at a glance

The cheapest LLM API depends on your token mix, not one sticker. As of the canonical July 10, 2026 snapshot, GPT-5-nano has the lowest input rate at $0.05 per million tokens, DeepSeek V4 Flash the lowest output at $0.18, and Gemini 2.5 Flash-Lite the best context per dollar at $0.10/$0.40 with a 1M window.

Here is the floor tier, with one budget-tier row per provider for scale.

Prices per 1M tokens, verified July 10, 2026, from the OpenRouter models snapshot behind our live tracker , cross-checked against vendor pages. The tracker is re-pulled and re-dated when providers move; treat it, not this page, as the freshest copy of any number.

Two footnotes the grid needs, because rates move:

• DeepSeek V4 Flash is the one row where the sticker depends on where you buy. Our table carries $0.09/$0.18 from the OpenRouter routing snapshot. DeepSeek 's own first-party pricing page lists $0.14 input (cache-miss) / $0.28 output, plus a deep cache-hit discount on repeated input — DeepSeek 's caching discounts are the steepest of any major provider (on V4 Pro, cached reads run under 1% of the input rate). Rates move, and this row moves more than most; either figure keeps it the cheapest output per token on the page, but budget off the first-party rate if you call DeepSeek directly. One dated housekeeping fact from the same page: the legacy deepseek-chat and deepseek-reasoner model names deprecate on July 24, 2026 (they map to V4 Flash's non-thinking and thinking modes).
• Llama 4 Scout has no single price. It is an open-weight model, so the rate is whatever a host charges: roughly $0.08/$0.30 at the low end to $0.11/$0.34 elsewhere as of July 2026, a spread that can reach several-x across providers. Our $0.10/$0.30 is one point in that range, not Meta's rate; Meta does not sell tokens.

Cost per 1,000 requests: one job, whole floor tier

A per-million-token rate means nothing until it meets a workload. Here is the same fixed job our live tracker uses, a support chatbot at 3,000 input and 500 output tokens per request, priced per 1,000 requests at the July 10, 2026 snapshot rates. The formula is open: (input tokens ÷ 1,000,000 × input rate + output tokens ÷ 1,000,000 × output rate) × 1,000.

All figures computed from July 10, 2026 snapshot rates on the fixed 3,000-in / 500-out mix. Your mix will differ; redo the multiplication before you budget.

Read the top four rows honestly and the "cheapest" crown wobbles. GPT-5-nano and DeepSeek V4 Flash land a penny apart on this mix, and the DeepSeek row swings on which sticker you pay: at the first-party $0.14/$0.28 rate the same job computes to about $0.56 per 1,000 requests, which would slot it behind Gemini 2.5 Flash-Lite, not ahead of GPT-5-nano. That is not a rounding quibble. It is the whole answer changing on a routing detail, which is why this page links a live tracker instead of pretending a table dated July 2026 is permanent.

The other honest read: the floor-to-flagship spread on this one job is roughly 85x ($0.35 to $30.00). Nothing else you tune, not caching, not batch discounts, not prompt trimming, moves a bill like the model picker does.

The quality cliff: what a $0.05 input rate actually buys

The floor tier is cheap because it is worse. The useful question is where it is worse, and whether your job lives there.

The best-documented data point is one tier up. NIST's CAISI evaluation (May 2026) put DeepSeek V4 Pro, the $0.44/$0.87 budget model, about eight months behind the US frontier, with the widest gaps on agentic work: 32% vs 71% for GPT-5.5 on CTF-Archive-Diamond and 44% vs 78% on PortBench. The same evaluation found V4 Pro more cost-efficient than GPT-5.4-mini on five of seven benchmarks. Both findings are true at once, and that is the whole floor-tier story in miniature: per task completed, cheap models win on routine work and lose exactly where tasks chain into multi-step, tool-using loops.

For the floor models themselves, GPT-5-nano, Flash-Lite, V4 Flash, we will not quote benchmark scores, because no evaluation we trust publishes a rigorous nano-class comparison as of July 2026, and vendor-run numbers are marketing. What holds up in practice reports is the shape of the cliff, not its exact height: floor models handle classification, extraction, tagging, routing, summarization of provided text, and template-following reliably. They degrade on multi-step reasoning, ambiguous instructions, and long agentic chains, and the degradation is not graceful. A workhorse model that is 95% right and a floor model that is 85% right are not 10% apart in production; the second one generates a review queue.

So the rule: match the model to the job's failure cost. If a wrong answer costs a retry, buy the floor. If a wrong answer costs a customer or an engineer-hour, the $1.75-per-thousand budget tier is the cheap option. Our best LLM for coding page runs this same logic against coding benchmarks specifically.

When cheap gets expensive: the five multipliers

The sticker price assumes one clean call per task. Production does not work that way, and every deviation multiplies the bill. Five mechanisms, each documented in cost post-mortems from 2025 to 2026:

1. Reasoning tokens bill as output. If a floor or budget model runs in a thinking mode, the invisible reasoning tokens are charged at the output rate. A real per-request cost of 3x to 9x the headline output price for reasoning-heavy calls is commonly reported in practitioner cost writeups — treat it as an order-of-magnitude planning number, not a measured constant. A $0.18-output model that thinks for 4,000 tokens before writing 300 is not a $0.18-class expense.
2. Verbosity is a price multiplier. A model that needs twice the output tokens to say the same thing doubles your effective output rate. Structured output (JSON, tables) runs 2x to 3x the tokens of plain prose. Cheap models tend to be the chatty ones, so a per-token win can lose per-answer.
3. Retries rebill everything. Every 429, 500, or malformed-JSON response you retry pays full input price again. Weaker models fail validation more often, so the retry multiplier correlates inversely with model quality, the exact wrong direction for the floor tier.
4. Agent fan-out is 3x to 30x, not 1x. One user action in an agent workflow triggers tool calls, retries, and reasoning steps, our long-standing guidance on the tracker is to multiply per-call cost by 3 to 30, and it keeps getting reconfirmed. The compounding version: a weaker model needs more steps to finish the same task, so downgrading the model can raise the step count faster than it cuts the step price. This is the single most common way "we switched to the cheap model" produces a bigger invoice.
5. The system prompt is a standing tax. A 2,000-token system prompt or RAG preamble resent on every call is 2,000 tokens of input, per call, forever. At floor-tier volumes this fixed overhead routinely dominates the bill. Prompt caching is the fix where it exists, roughly 10% of the input rate on cached reads across the US big three, under 1% on DeepSeek V4 Pro ($0.004 cached read against a $0.44 input rate), but caching only pays when a stable prefix actually repeats, and some providers meter cache storage separately, so check the vendor's caching terms before assuming the discount.

None of this argues against the floor tier. It argues for doing the multiplication: sticker × verbosity × retries × fan-out × prompt overhead. A model that is 4x cheaper per token and 2x worse on the other four factors is not 4x cheaper.

The jurisdiction asterisk on the DeepSeek discount

The cheapest output tokens on this page come with a paragraph the price grid cannot show. DeepSeek 's direct API (api.deepseek.com) routes prompts to servers in Hangzhou, China; its terms permit training on API data; and as of July 2026 it offers no SOC 2 Type II report and no HIPAA BAA. For a hobby project or public-data batch job, that may cost you nothing. For anything touching client data, it is a compliance conversation, not a pricing one.

There is an escape hatch, because the model weights are open: US-hosted DeepSeek . Fireworks serves V4 Flash at $0.14/$0.28 and V4 Pro at $1.74/$3.48 (about 4x DeepSeek 's first-party Pro rate), Together AI serves the line with SOC 2 certification, and Bedrock and Azure AI Foundry run it with no vendor data access, all July 2026 provider-listed rates. The framing that survives scrutiny: part of the DeepSeek discount is a jurisdiction discount. US-hosted V4 Pro at $1.74/$3.48 is still 65 to 78 percent cheaper than the GPT-5.6 Sol flagship at $5.00/$30.00 depending on the input/output mix, but it is 4x the direct-API sticker that headlines the comparison posts. Decide which number applies to you before it decides for you.

The actual $0 routes, and their catches

"Free LLM API" is mostly a search query, not a product, but three routes genuinely bill $0 per token:

• Self-host open weights. Llama , Qwen, and DeepSeek weights run on your own hardware with no API meter. The bill moves to GPUs and ops time; at low volume that is usually more expensive than the floor tier, not less.
• Gemini's free tier. Google offers free API usage on 2.5 Pro, 2.5 Flash, and 2.5 Flash-Lite via an AI Studio key ( Gemini 3.1 Pro has no free tier as of July 2026). We deliberately print no request-per-day numbers: Google stopped publishing a per-model limits table, the caps are per-project, shown live in the AI Studio dashboard, and revised without notice. One line worth knowing: free-tier data may be used for training; paid-tier data is not.
• $0-prompt models on OpenRouter. Our July 10, 2026 snapshot lists 26 models at a $0 prompt price, rate-limited and rotated without warning. Fine for prototyping, not a production plan.

All three share a ceiling: the moment traffic is sustained and availability matters, you are back on the metered floor tier, which is the point of this page.

Where this comparison falls short

Honest limits of this page, so you can weight it properly:

• It is a snapshot pretending as hard as it can not to be. Every figure is dated July 10-12, 2026. Floor-tier prices move more often than flagship prices, and the DeepSeek row already shows a live divergence between routed and first-party rates. The tracker is re-pulled; this page is only re-dated when it is revised.
• The floor-tier quality section has no scores in it, on purpose. No benchmark we trust covers the nano class rigorously as of this writing, so we described the cliff's shape from one-tier-up data (NIST CAISI on V4 Pro) and practice reports. If a vendor quotes you a floor-model benchmark, ask who ran it.
• The per-request math uses one fixed token mix. 3,000-in/500-out flatters input-cheap models. An output-heavy mix (drafting, generation) reorders the top of the table in DeepSeek 's favor; redo the arithmetic on your own mix.
• We priced tokens, not throughput. Rate limits, concurrency caps, latency, and regional availability gate real deployments and appear nowhere in this table. Provider concurrency caps apply and change without notice — check each provider's current rate-limits page; floor models commonly gate by spend tier.
• Llama 4 Scout's row is one host's price. The open-weight market spreads several-x across providers; our single row compresses that dishonestly, and we flagged it rather than fixed it.

All guides in this topic

• LLM API Pricing Compared (July 2026): Live Cross-Provider Cost Per Token — the canonical live tracker every number on this page inherits from.
• OpenAI API Pricing (July 2026): Live Per-Token Rates, Real Workload Costs, and Cheaper Swaps — the full OpenAI table, including GPT-5-nano's ultra-budget lane.
• Claude API Pricing (2026): Live Cost Per Token for Opus, Sonnet, and Haiku — Anthropic's lineup; its floor starts at Haiku, above this page's tier.
• Gemini API Pricing (July 2026) — the full Google table: Flash-Lite's flat 1M window, the Pro long-context surcharge, and cache storage billed by the hour.
• Claude vs ChatGPT (2026): The Task-by-Task Verdict on Verified Live Pricing — the flagship matchup, task by task.
• Gemini vs ChatGPT (July 2026) — Google vs OpenAI across consumer plans and API.
• Grok vs ChatGPT (July 2026) — xAI's cheap-output flagship against the default.
• DeepSeek vs ChatGPT (July 2026) — the price-gap matchup this page's jurisdiction section summarizes.
• Best LLM for Coding (July 2026): Verified Prices, Benchmark Reality, and Which Model Wins Each Job — price joined to coding benchmarks, including the budget lane.

Frequently asked questions

Частые вопросы