Cheapest LLM API 2026: 8 Providers Ranked by Blended Cost

The cheapest LLM API in 2026 is Gemini 2.5 Flash-Lite, at $0.175 per million blended tokens on a 3:1 input:output mix. This pillar ranks eight providers (Anthropic, OpenAI, Gemini, DeepSeek, Mistral, Groq, Together, Fireworks) on one consistent cost metric, flags where capability breaks the ranking, shows how caching and batch reorder it, and runs the cheapest tier through the margin calculator so the saving shows up as gross margin.

1. How the blended cost is computed

Comparing input and output rates separately is noisy, so this ranking uses one blended number: cost per million tokens at a 3:1 input:output ratio, computed as (0.75 × input rate) + (0.25 × output rate). A 3:1 ratio is a reasonable default for retrieval-augmented and chat workloads, where the prompt and context outweigh the completion. For each provider, a representative cheap-to-workhorse model is priced; flagship tiers are noted separately because few cost-sensitive products run them by default.

All per-token rates are verified against each provider's official pricing page: DeepSeek on October 2, 2026, the others on May 25, 2026. DeepSeek is priced at its peak rate; off-peak (outside 01:00-04:00 and 06:00-10:00 UTC on weekdays) it costs half. The blended figure is arithmetic on those verified rates, not a benchmark of model quality. A provider can rank first on blended cost and still be the wrong choice if its model fails your task.

2. The 8-provider blended-cost ranking

Representative model per provider, ranked cheapest first by blended 3:1 cost per million tokens.

Rank	Provider / model	Input / Output	Blended (3:1)
1	Gemini 2.5 Flash-Lite[2]	$0.10 / $0.40	$0.175
2	Groq / Together / Fireworks, GPT-OSS 120B[6]	$0.15 / $0.60	$0.263
3	DeepSeek V4.1-Flash, peak (off-peak half)[1]	$0.30 / $1.20	$0.525 ($0.263 off-peak)
4	Groq, Llama 3.3 70B[6]	$0.59 / $0.79	$0.640
5	Mistral Large 3[5]	$0.50 / $1.50	$0.750
6	OpenAI GPT-5.4-mini[3]	$0.75 / $4.50	$1.688
7	Anthropic Haiku 4.5[4]	$1.00 / $5.00	$2.000
8	Gemini 3.5 Flash[2]	$1.50 / $9.00	$3.375

The arithmetic on the leaders: Gemini 2.5 Flash-Lite is 0.75 × $0.10 + 0.25 × $0.40 = $0.075 + $0.10 = $0.175. The open-model hosts cluster at 0.75 × $0.15 + 0.25 × $0.60 = $0.263 because GPT-OSS 120B is priced identically across Groq, Together, and Fireworks. DeepSeek V4.1-Flash is 0.75 × $0.30 + 0.25 × $1.20 = $0.225 + $0.30 = $0.525 at peak, and exactly the GPT-OSS figure off-peak, when every rate halves. For reference, the closed-frontier mid and flagship tiers (not in the cheap ranking) are far higher: OpenAI GPT-5.4 blends to $5.625, Anthropic Sonnet 4.6 to $6.00, on the same 3:1 mix.

3. The capability caveat

This ranking is a price ranking. It does not claim Gemini 2.5 Flash-Lite or any other cheap-tier model produces output equal to a flagship model. Three honest qualifiers:

• Cheap-tier models are smaller. Flash-Lite and V4.1-Flash are built for cost and speed, not maximum reasoning. For complex agentic or coding tasks, a pricier model often does the job in fewer tokens and fewer retries, which can make it cheaper in total despite a higher per-token rate.
• Representative models differ in capability. The eight rows above are not capability-matched. GPT-OSS 120B is a large open model; Gemini 2.5 Flash-Lite is a small fast model. Same blended-cost table, different jobs.
• Total cost includes retries and length. A model that needs two attempts or writes twice as much costs more than its per-token rate implies. Measure dollars per completed task on your workload, not dollars per token.

The correct decision rule: start at the top of the price ranking, run your eval, and stop at the first provider that clears your quality bar. That is usually not the single cheapest row, but it is rarely the most expensive one either.

4. Caching and batch change the order

The headline ranking assumes no caching and no batch. Both reorder the list for the right workload:

• Cache hits. DeepSeek V4.1-Flash cache-hit input is $0.006 per million tokens at peak and $0.003 off-peak^[1], a fiftieth of its cache-miss rate. Anthropic cache reads are 0.1x base (Sonnet 4.6 cached input $0.30 against $3)^[4]. A workload that resends a large stable context collapses input cost toward the cache rate, which can pull a nominally pricier provider below a cheaper one.
• Batch. Anthropic and OpenAI both offer a 50% batch discount^[3][4]. For non-time-sensitive bulk jobs (classification, document processing), the batch rate is the number to rank on, not the standard rate.

The practical instruction: compute the blended cost with your real cache-hit ratio and batch share. A product that caches 80% of input on Sonnet 4.6 has a far lower effective input cost than the $3 standard rate suggests, which narrows the gap to the cheap-tier leaders considerably.

5. Margin on a $19 SaaS at the cheap tier

The margin calculator prices a $19/month SaaS where each user triggers 40 API calls a day at 800 input and 600 output tokens, on the cheap-tier leader Gemini 2.5 Flash-Lite ($0.10 / $0.40). The engine renders gross margin from the verified rate:

At the cheapest tier, API cost almost disappears from the per-user economics: the engine returns a high gross margin because well under a dollar of model cost per user a month leaves the bulk of the $19 as contribution. The strategic read for a cost-sensitive product is that the cheap tier turns the model from the dominant cost line into a rounding error, which frees the budget for acquisition or for a stronger model on the calls that need it.

The mirror scenario is on the pricier closed-frontier tiers, where the same call volume can consume a large share of a low subscription price. The Anthropic vs OpenAI comparison, the DeepSeek vs Gemini comparison, and the open-model host comparison drill into each pairing. The AI Product Margin Calculator lets you run your own rates.

6. Decision guidance

• Cheapest raw cost: Gemini 2.5 Flash-Lite at $0.175 blended, with Google Cloud and multimodal headroom. DeepSeek V4.1-Flash matches the open-model hosts if your batch work can run off-peak, and its cache hits cost a fiftieth of a miss.
• Open weights, want speed or self-host option: GPT-OSS 120B on Groq (fastest), Together, or Fireworks at $0.263 blended.
• EU data residency at low cost: Mistral Large 3 at $0.75 blended, the cheapest EU-based option here.
• Need flagship capability: step up to GPT-5.5, Claude Opus 4.8, or Gemini 3.1 Pro (the current Pro flagship; Gemini 3.5 Pro is its announced successor, rolling out mid-2026) and accept the higher blended cost. Rank those on the matched-tier comparisons, not this cheap-tier list.

Re-verify every provider's pricing page before committing; per-token rates at this layer move with each model release, and DeepSeek has repriced twice in 2026 and now charges by time of day. Rank on your real cache-hit and batch ratios, and stop at the cheapest provider that passes your eval.

Per-token figures verified against official pricing pages: DeepSeek on 2026-10-02, the other providers on 2026-05-25.

Frequently asked questions

What is the cheapest LLM API in 2026?

On a blended 3-to-1 input-to-output mix, Gemini 2.5 Flash-Lite ($0.10 input / $0.40 output) is the cheapest of the eight providers checked, at $0.175 per million blended tokens. The three open-model hosts (Groq, Together, Fireworks) running GPT-OSS 120B at $0.15 / $0.60 follow at about $0.26, level with DeepSeek V4.1-Flash off-peak ($0.15 / $0.60). At peak hours DeepSeek's rates double to $0.30 / $1.20, about $0.53 blended. All rates verified on each provider's official pricing page.

Is the cheapest LLM API also the best choice?

Not always. Blended cost ranks raw price, not capability. Gemini Flash-Lite and the GPT-OSS hosts win on dollars per token, but a harder task may need a stronger model where the cost is justified by output quality and retention. The right method is to find the cheapest provider that clears your quality bar on your actual prompts, not the cheapest provider overall.

How much do caching and batch change the ranking?

A lot for the right workload. Anthropic and OpenAI both cut cached input by roughly 90 percent and offer a 50 percent batch discount, and DeepSeek V4.1-Flash bills a cache hit at $0.006 per million input tokens at peak ($0.003 off-peak). A heavily cached or batched workload can reorder the ranking, pulling a nominally pricier provider below a cheaper one. Compute the blended cost with your real cache-hit and batch ratios, not the headline rates.

What is the cheapest LLM API for classification and extraction on a solo SaaS under $1 per user per month?

For classification, extraction, and routing on a solo SaaS, the cheapest option is Gemini 2.5 Flash-Lite ($0.10 input / $0.40 output), at $0.175 per million blended tokens on a 3-to-1 input-to-output mix, followed by GPT-OSS 120B on Groq, Together or Fireworks and DeepSeek V4.1-Flash off-peak at about $0.26. On the $19 SaaS scenario modeled here at 40 API calls per user per day, the AI Product Margin Calculator puts Flash-Lite's API cost at well under $1 per user per month. These small fast models are built for exactly this kind of high-volume, low-reasoning task; reserve a pricier tier for the minority of calls that genuinely need stronger reasoning.

Which cheap LLM API should I pick if I need EU data residency under a tight budget?

Mistral Large 3 at $0.50 input / $1.50 output per million tokens, which blends to $0.75 per million on a 3-to-1 mix, is the cheapest EU-based option in this eight-provider ranking. It costs more than the global cheap-tier leader (Gemini 2.5 Flash-Lite at $0.175 blended) but is the lowest-cost choice when EU data residency is a hard requirement. Verify Mistral's current rates and residency terms on their docs before committing, since per-token pricing at this layer moves with each model release.