The cheapest AI APIs in 2026, and what a task actually costs

The cheapest AI APIs right now

TokenMonopoly's live tracker ranks coding-capable models by lowest input price across every host that serves them. The bottom of the market in early November:

• Qwen3.8 27B from under $0.01 per million input tokens and $0.32 output via Primeintellect.
• GLM 5.2 from under $0.01 input and $1.40 output via Morph.
• GLM-5.3 Flash from $0.03 input and $0.07 output via Relace.
• DeepSeek V4.1-Flash from $0.04 input and $0.15 output via Sail Research.
• GLM-5.3 flagship from $0.07 input and $0.88 output via Baidu.
• GPT-6 Luna at $0.10 input and $0.50 output direct from OpenAI.
• Kimi K3 from $0.15 input via Sail Research, though output runs $7.15 per million.

The pattern: third-party hosts undercut first-party APIs by 50 to 95 percent on the same weights, because open models are a commodity and hosts compete on margin. If you are spending real money, never pay list price for an open model.

The cheapest AI API for coding

For coding agents specifically, three options dominate the value conversation. DeepSeek V4.1-Flash at $0.30 per million input and $1.20 output first-party (half that off-peak) is the default for high-volume agent loops. GLM-5.3 Flash at $0.15 input and $0.50 output matches it on coding benchmarks at lower cost. And LongCat-2.5 at $0.30 input and $1.20 output on an Anthropic-compatible endpoint drops straight into Claude Code as a frontier-adjacent coding model at a budget price.

At the frontier end, the cheapest serious option is OpenAI's GPT-6.1 Terra at $1 per million input tokens and $5 output, half the price of the Sol tier, with computer-use capability the open models do not match yet. Everything above that (Claude Opus 5.5, GPT-6 Astra) is premium spend for the hardest tasks.

First-party pricing for comparison

List prices per million tokens, input and output, as of early November 2026:

• DeepSeek V4.1-Flash: $0.30 / $1.20 peak, $0.15 / $0.60 off-peak.
• GLM-5.3 Flash: $0.15 / $0.50, cached input $0.03.
• Ultramarine 1 Flash: $0.15 / $0.50, cached $0.03.
• LongCat-2.5: $0.30 / $1.20.
• Mistral Large 4: $1.36 / $4.18.
• GPT-6.1 Terra: $1 / $5, cached input $0.05.
• Qwen3.8-Max: $2 / $6 international.
• Kimi K3: $3 / $15, cached input $0.30.

Peak and off-peak rules matter more than most teams expect. DeepSeek halves prices outside peak hours, and Z.ai's coding plan does the same on weekdays outside 14:00 to 18:00 China time. Batch jobs, nightly evaluations and bulk document processing should never run at peak rates.

Why per-token price is the wrong unit

Cost per completed task is the honest metric, and it flips the table. Ultramarine 1 Flash finishes professional knowledge-work tasks at $0.13 each while Claude Fable 5 at maximum reasoning costs $8.19 for a lower score, a 60x gap that survives scrutiny because both the prices and the token counts are public. On Terminal-Bench 4.0, Codex with GPT-6.1 Sol matches Codex with GPT-6 Astra at 58.2 percent while costing $1.92 per task against $9.90. The cheapest API is the one that finishes your task in the fewest tokens at an acceptable success rate, so measure cost per accepted PR or per resolved ticket, not per million tokens.

Pricing traps that erase the savings

Three gotchas show up repeatedly when teams chase the cheapest AI API. First, output tokens cost 3 to 8 times more than input almost everywhere, and reasoning models generate long hidden chains of output tokens, so a model with a cheap input price can still bill heavily on agentic tasks. Second, cached-input pricing only helps if your prompts share a long stable prefix; agent loops that rewrite context every turn never hit the cache rate. Third, the very cheapest hosts achieve their prices with aggressive quantization or shared capacity, which can quietly change model behavior. Pin the exact checkpoint, run your eval suite against the hosted version before migrating, and treat any saving above 90 percent off list as a prompt to test harder, not a reason to skip testing.

How to pick the cheapest AI API for your workload

1. 1Define the task unit first: a resolved support ticket, an accepted PR, a finished report. Estimate tokens per unit from a week of logs.
2. 2Shortlist by benchmark fit, not price: a model that fails half your tasks is infinitely expensive.
3. 3Check third-party hosts for the same weights before paying first-party prices.
4. 4Move schedulable work to off-peak windows for an automatic 50 percent cut on DeepSeek and Z.ai.
5. 5Turn on prompt caching: GLM's cached input at $0.03 and Terra's at $0.05 make repeated context nearly free.
6. 6Re-check prices monthly. This market repriced four times in 2026 alone.

FAQ

Sources

1. 01Cheapest AI APIs: live 2026 prices across 60+ hosts - TokenMonopoly (opens in a new tab)
2. 02Introducing DeepSeek-V4.1-Flash - DeepSeek (opens in a new tab)
3. 03GLM-5.3 - Z.ai (opens in a new tab)
4. 04Ultramarine 1 Flash - Paper Instruments (opens in a new tab)
5. 05Introducing LongCat-2.5 - Meituan (opens in a new tab)

Elsewhere on this site

The selected work catalogue lists the SaaS products and dashboards behind notes like this one.