← All articles

The cheapest LLM API in 2026: GPT-5.4-nano vs Claude Haiku vs DeepSeek Flash

By LW Forge — maintainer of LLM Scout · Updated July 20, 2026

Searching for the cheapest LLM API returns a wall of tables that all stop at the same place: price per million tokens, sorted ascending, buy the top row. That ranking answers the wrong question. The cheapest API is the cheapest model that still does your job — and at the bottom of the price list, capability differences between models are at their largest. Here's the budget tier of every major provider as of July 2026, with verified list prices, and a way to choose between them that won't backfire in production.

The contenders

Every provider now fields a budget model, and the list prices cluster into two groups.

The ultra-budget group. OpenAI's GPT-5.4-nano (around $0.20 in / $1.25 out per million tokens), DeepSeek V4 Flash (about $0.14 / $0.28) and Meta's hosted Llama 4 Maverick (roughly $0.20 / $0.60, varying by host). These are priced for volume: classification, extraction, routing, short structured answers.

The premium-budget group. GPT-5.6 Luna ($1 / $6), Claude Haiku 4.5 ($1 / $5) and Gemini 2.5 Flash ($0.30 / $2.50). Several times pricier than the ultra-budget group — though still a fraction of any flagship tier — and noticeably more capable — these handle multi-step instructions and light reasoning that the cheaper group fumbles.

Gemini 2.5 Flash deserves a special mention as the odd one out: its input rate sits near the ultra-budget group while its capability leans premium, which makes it a frequent winner on input-heavy work like summarization — one more reason not to pick by brand loyalty.

Two observations jump out of the table. First, output-rate spread is wider than input-rate spread: DeepSeek Flash's $0.28 output rate versus nano's $1.25 makes output-heavy tasks a very different ranking than input-heavy ones. Second, the gap between the two groups (roughly 4-7×) is larger than the gaps within them — the real decision is which group your task belongs in, not which logo. Compare the current budget-tier rates for your own token profile in the GPT vs Claude comparison.

What "cheap per token" hides

Three effects routinely make the nominally cheapest model the most expensive choice.

Failure retries. A model that misclassifies 8% of inputs against another's 3% doesn't cost you 5 points of accuracy — it costs a retry pipeline, human review, or silent errors downstream. Cost per successful task is the honest metric, and it often ranks the premium-budget group above the ultra-budget group on tasks near the edge of their ability.

Verbosity and token counting. The same job takes different token counts on different models — tokenizers differ, default verbosity differs, and a chattier model bills more output at whatever its rate is. Measure your real prompts in the token calculator rather than assuming a fixed token budget transfers between providers.

Instruction brittleness. Cheap models follow simple instructions well and complex ones erratically. If your prompt needs six rules followed simultaneously, the ultra-budget group may need prompt surgery and output validation that costs engineering time — real money the price table never shows.

An honest shootout, step by step

The good news: at these prices, testing is nearly free. A 500-example evaluation across all four ultra-budget and premium-budget contenders costs pennies to a few dollars total.

  1. Pull 100-500 real examples of your task, with known correct answers.
  2. Freeze one prompt, one output format, one max-token cap; run it unchanged on every candidate.
  3. Score correctness automatically where possible (exact match, schema validation) — subjective grading at this volume invites bias.
  4. Compute cost per successful task, charging retries to the model that caused them.
  5. Pick the cheapest model that clears your quality bar with margin — a model that barely passes today fails quietly after your traffic shifts.

Teams that run this usually land in one of two places: the task is genuinely easy (sentiment, tagging, simple extraction) and DeepSeek Flash or nano wins at a price that ends the conversation — or the task has hidden sharp edges and the premium-budget tier earns its multiple. Both outcomes are wins; the expensive outcome is not knowing which you're in.

When to climb a tier — and when to split

If your chosen budget model's error rate forces retries or review on more than a few percent of traffic, do the arithmetic on the next tier up: at, say, 5× the token price but half the failure handling, the "expensive" model is frequently cheaper end to end — the full framework for that trade is in cost per task, not cost per token. And remember the answer can be two models: route the easy majority to the ultra-budget tier and escalate flagged cases. Routing is how you get a blended rate no single model can match — estimate each side's contribution in the OpenAI cost calculator and let your own traffic pick the winner.