Cheapest LLM API per million tokens

The cheapest LLM API per million tokens is not a single model — it depends on how your traffic splits between input and output — but the budget field clusters tightly at the bottom, and since October 2026 a current-generation model, Qwen3.7-Flash, undercuts all of it — including the legacy model that used to hold the floor. This is that tier, ranked, with the fine print that decides the real winner.

To rank models that price input and output differently on one axis, we use a blended price: (3 × input + 1 × output) ÷ 4, a rough 3-to-1 input-to-output mix typical of RAG and chat. Your real ratio changes the order — an output-heavy workload punishes high output rates far more — so treat the ranking as a starting point and price your own mix.

The budget tier, ranked by blended price

Blended prices are rounded to the cent. Qwen3.7-Flash enters at the top of this table at $0.030 / $0.130 — about 2.3× below the previous leader. That rate is the ≤32K context band on Alibaba's Singapore endpoint; above 32K it steps to $0.100 / $0.400, and above 256K to $0.200 / $0.800. Re-verified 2026-11-02. The previous leaders were a tie: Qwen-Flash and legacy GPT-5 nano, both $0.05 / $0.40, or $0.14 blended. Earlier versions of this post said Alibaba halved Qwen-Flash's input rate from $0.10 to $0.05 in early September 2026; that change does not appear in Alibaba's release notes or price notices, so we have removed it (checked 2026-11-02). Google has closed Gemini 2.5 Flash-Lite to new customers; existing users keep access. Our 2026-10-04 update said Google no longer published a rate for it — that was wrong: Google's pricing page still lists it at $0.10 / $0.40 (re-checked 2026-11-02). Llama 4 Scout has left the table: Together AI removed its serverless endpoint on 2026-02-06, and it is now offered only on dedicated endpoints and for fine-tuning. Together's serverless catalog no longer lists any Llama model. DeepSeek moved twice in under a month: it split its pricing into peak and off-peak bands on 2026-08-16, with peak hours billed at twice the off-peak rate, then cut the budget tier on 2026-09-10, when DeepSeek-V4.1-Flash replaced V4-Flash at $0.15 / $0.60 off-peak, down from $0.22 / $0.66. That takes it from $0.33 blended to $0.26, level with Mistral Small 4 — and off-peak it has the cheapest cached input in the tier, covered below.

The cheapest overall: Qwen3.7-Flash, under 32K tokens

On raw price, Qwen3.7-Flash now leads at $0.030 / $0.130 — a blended $0.06 — but only for prompts up to 32K tokens; above that it steps to $0.100 / $0.400. Behind it and Ministral 3 3B ($0.10 blended), legacy GPT-5 nano and Qwen-Flash tie at $0.05 / $0.40 — a blended $0.14 each. The asterisk on GPT-5 nano is "legacy": OpenAI has deprecated it and scheduled its API shutdown for December 2026, so anything built on it must migrate within months. Qwen-Flash carries no such clock, which makes the tie lopsided in practice — the same sticker price, without the migration deadline. Among current-generation, generally-available budget models, the floor is now Qwen3.7-Flash at $0.06 for short prompts.

What that means in dollars

Take a high-volume workload of 10M input and 2M output tokens per month. Qwen3.7-Flash runs about $0.56; GPT-5 nano and Qwen-Flash both about $1.30; Qwen3.5-Flash about $1.80; DeepSeek V4.1 Flash off-peak about $2.70. Gemini 2.5 Flash-Lite would also land near $1.80, but Google has closed it to new customers. Claude Haiku 5.5, launched 2026-10-07, runs it for about $2.00, a tenth of Haiku 4.5. The same workload on GPT-5.4 mini is about $16.50, and on Claude Haiku 4.5 about $20.00 — a 35× spread inside what everyone calls the "budget" tier. Picking the wrong budget model costs more than picking a flagship for a task that needed one.

DeepSeek still has the cheapest automatic cached input

If your workload reuses a large, stable prompt prefix, DeepSeek V4.1 Flash is the cheapest hands-off cache: cache hits list at $0.003 per 1M tokens off-peak ($0.006 peak) — a 98% discount off its $0.15 input rate, applied automatically with nothing to manage. Off-peak, that now undercuts the two models that used to read cheaper on paper. Legacy GPT-5 nano is $0.005, and it retires in December 2026. Qwen-Flash is also $0.005, but only on its explicit cache, which you pay $0.063 per 1M tokens to create and then have to manage; its automatic implicit cache reads at $0.01. The catch on DeepSeek is the same off-peak caveat: during peak hours the rate doubles to $0.006, just above both. For a cache-heavy pipeline scheduled off-peak, it is the one to beat.

The caveats that decide the real winner

• ◦Legacy status: GPT-5 nano is a prior-generation model — cheap on sticker, but subject to deprecation. On 2026-10-01 OpenAI also deprecated GPT-5.4 nano, with shutdown on 2027-04-01 and GPT-6 Luna as the replacement. Weigh migration risk for anything long-lived.
• ◦Off-peak vs peak: DeepSeek's $0.15 / $0.60 (and its $0.003 cached rate) are off-peak; the peak rate is 2× and applies only 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays — weekends and those holidays are off-peak all day. These are the DeepSeek-V4.1-Flash rates in effect since 2026-09-10.
• ◦Context bands: Qwen3.7-Flash's $0.030 / $0.130 holds per request only under 32K prompt tokens; above 32K it bills $0.100 / $0.400, and above 256K $0.200 / $0.800. With 100K-token prompts, the 10M + 2M workload above costs about $1.80 a month, not $0.56. Claude Haiku 5.5 has the same shape: $0.10 / $0.50 holds only for prompts up to 100K tokens, and above that the whole request bills $0.50 / $2.50. It also uses Anthropic's newer tokenizer, which counts roughly 30% more tokens than Haiku 4.5 for the same text.
• ◦Tiered and promotional pricing: Qwen-Flash's $0.05 / $0.40 is the Singapore-region rate for prompts up to 256K input tokens; above that it bills $0.25 / $2.00, a 5× step. Qwen's budget rates can also include limited-time promos. Verify both the tier and the current price before you rely on them.
• ◦Audio surcharge: Gemini Flash-Lite models price audio input higher than text — Gemini 2.5 Flash-Lite text input is $0.10, but audio runs higher. Check the modality you actually send.
• ◦Cost per task, not per token: the cheapest per-token model is not automatically the cheapest per task — output length and tokenizer differences can flip the ranking. Model your real prompts.

Which should you pick

For most teams the honest answer is one of three:

• ◦Absolute lowest cost: Qwen3.7-Flash ($0.030 / $0.130), as long as your prompts stay under 32K tokens. Legacy GPT-5 nano ($0.05 / $0.40) no longer holds the floor, and Qwen-Flash matches it without the deprecation risk.
• ◦Lowest blended cost after Qwen3.7-Flash: Ministral 3 3B ($0.10 / $0.10), then Qwen-Flash ($0.05 / $0.40) — Qwen-Flash is the cheaper of the two only when input outweighs output by more than 6 to 1; DeepSeek V4.1 Flash off-peak wins for cache-heavy pipelines you would rather not hand-manage.
• ◦Lowest cost with no off-peak clock and a US-hosted API: GPT-6 Luna ($0.10 / $0.50) or Claude Haiku 5.5 ($0.10 / $0.50, for prompts up to 100K tokens).

What is the cheapest LLM API right now?

By blended price, Qwen3.7-Flash is the cheapest at about $0.06 ($0.030 / $0.130), though that rate applies only to prompts under 32K tokens. Behind it come Ministral 3 3B at about $0.10 ($0.10 / $0.10), then legacy GPT-5 nano and Qwen-Flash tied at about $0.14 ($0.05 / $0.40) — but GPT-5 nano retires in December 2026, so Qwen-Flash is the one to build on. Qwen3.5-Flash follows at about $0.18 ($0.10 / $0.40), the price point Gemini 2.5 Flash-Lite held before Google closed it to new customers; DeepSeek V4.1 Flash, after DeepSeek's 2026-09-10 cut, sits level with Mistral Small 4 at about $0.26 ($0.15 / $0.60 off-peak).

Is a cheaper model always the right choice?

No. A budget model that needs three attempts or a longer prompt to match a workhorse's single-shot answer can cost more in total, and it may miss on quality, latency, or tool use. Cheapest per token is a starting filter, not the decision.

How do I compute the cost for my own workload?

Take your monthly input and output token volumes, multiply each by the model's per-1M rate, and add them — then layer in cache-hit rate and any off-peak or long-context surcharge. The pricing calculator does this for every model it covers and lets you compare two side by side.

Do these prices change?

Frequently. These are list prices as of 2026-11-02; budget pricing is the most volatile tier — DeepSeek split its pricing into peak and off-peak bands on 2026-08-16, then cut its budget tier on 2026-09-10 — and promos and off-peak windows come and go. Re-verify with each provider before committing.