The 2026 AI API Cost Index: What 22 Models Really Cost Per Workload

Almost every AI API pricing page answers the wrong question. It tells you the cost of a million tokens, when what you actually need to know is the cost of your workload — one support reply, one code review, one document summary — multiplied by how many of them you run.

This index closes that gap. It takes the published list rates of 22 models across four providers, verified against each vendor’s official pricing page in September 2026, and prices six realistic production workloads against all of them. Every number is reproducible from the methodology at the bottom of this page, and the raw dataset is available to download.

Cost per workload

Each workload below states its token assumptions explicitly, because the assumptions matter more than the rates. Change the input-to-output ratio and the ranking can move: output tokens are priced 3–6× higher than input almost everywhere, so a workload that emits a lot is far more sensitive to the output rate than to the input rate.

All figures are standard-tier list prices. Column meanings: Monthly is the no-cache, standard-tier bill; Per 1k req is the same figure divided by thousands of requests; Cached 70% assumes a 70% prompt-cache hit rate; Batch applies the 50% batch discount. The cheapest row in each table is highlighted.

Classification / routing

Ticket triage, intent labels, spam checks. Short in, very short out.

1,500 input + 30 output tokens per request × 500,000 requests/month = 750.0M input and 15.0M output tokens per month. Cheapest: Gemini 2.5 Flash-Lite at $81.0. Most expensive: GPT-5.6 Cyber at $10,500.

Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the downloadable dataset.

Customer support reply

System prompt + retrieved policy + short answer.

3,000 input + 400 output tokens per request × 50,000 requests/month = 150.0M input and 20.0M output tokens per month. Cheapest: Gemini 2.5 Flash-Lite at $23.0. Most expensive: GPT-5.6 Cyber at $3,375.

Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the downloadable dataset.

RAG answer with context

Retrieved passages plus instructions.

12,000 input + 600 output tokens per request × 30,000 requests/month = 360.0M input and 18.0M output tokens per month. Cheapest: Gemini 2.5 Flash-Lite at $43.2. Most expensive: GPT-5.6 Cyber at $5,850.

Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the downloadable dataset.

Coding agent turn

Repo context, tool results, patch output. Heavy output.

60,000 input + 3,000 output tokens per request × 20,000 requests/month = 1200.0M input and 60.0M output tokens per month. Cheapest: Gemini 2.5 Flash-Lite at $144. Most expensive: GPT-5.6 Cyber at $19,500.

Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the downloadable dataset.

Bulk document summarisation

Nightly pipeline over a document backlog.

8,000 input + 800 output tokens per request × 200,000 requests/month = 1600.0M input and 160.0M output tokens per month. Cheapest: Gemini 2.5 Flash-Lite at $224. Most expensive: GPT-5.6 Cyber at $32,000.

Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the downloadable dataset.

Long-document analysis (past the cliff)

Whole-report analysis. Past 200K/272K every affected model charges long-context rates for the ENTIRE request.

300,000 input + 2,000 output tokens per request × 2,000 requests/month = 600.0M input and 4.0M output tokens per month. Cheapest: Gemini 2.5 Flash-Lite at $61.6. Most expensive: GPT-6 Astra at $12,300.

Showing the 10 cheapest of 22 models. The full 22-model matrix for this workload is in the downloadable dataset.

The context cliff: why one extra token can cost thousands

Six of the models in this index change their unit price once a request crosses a size threshold. The mechanism is the same everywhere: above the threshold, the input rate doubles and the output rate rises by half — and critically, that multiplier applies to the entire request, not just the portion above the threshold.

That produces a genuine discontinuity. Take GPT-6 Astra running 2,000 calls a month with a 2,000-token answer:

272,000 input tokens per call	$5,640 / month
272,001 input tokens per call	$11,180 / month

One additional token in the prompt — a slightly longer retrieved document, one more sentence of instruction — costs $5,540 a month. This is not a rounding error and it does not appear on any pricing page as a warning.

The workaround: chunk before you cross

Because the multiplier applies per request, splitting a long prompt into pieces that each stay under the threshold avoids it entirely. Two 150K-token calls cost exactly half of one 300K-token call:

The trade-off is real and worth stating plainly: chunking removes the model’s ability to reason across the whole document at once. It is the right move for extraction, summarisation and per-section analysis. It is the wrong move when the task depends on relationships between distant parts of the input. If you need both, run the cheap pass first and escalate only the cases that need the full context.

Prompt caching: break-even on the second request

Caching has an entry cost. Writing a prefix to the cache is billed at 1.25× the normal input rate on both OpenAI and Anthropic, so caching is only worth it if you read the prefix back often enough to recover that premium.

The arithmetic is simple. If the write costs W and each read costs R (both relative to uncached input), caching beats not caching after W / (1 − R) reads:

Every provider that publishes both rates lands between 1.28 and 1.39 reads. In practice that means caching is profitable from the second request that shares a prefix — a very low bar for any workload with a fixed system prompt, a stable tool definition set, or a document that gets queried repeatedly.

The reasoning-token tax

Reasoning tokens are billed as output tokens. That matters because two of the four Anthropic models here have thinking always on — you cannot switch it off — while OpenAI’s GPT-6 family lets you set reasoning effort to none.

For a high-volume, low-difficulty task such as classification or routing, that difference shows up directly in the bill. A model with mandatory reasoning will emit thinking tokens for a task that needed none, and every one of them is charged at the output rate. Before committing to a per-token price, check whether you can actually turn reasoning off on that model — the column below is the one most comparisons omit.

Full price list

List rates per 1M tokens, standard tier, verified September 2026. Google’s Gemini 3 Flash family is on promotional pricing through 31 December 2026 and doubles on 1 January 2027, so treat those rows as temporary.

Methodology

So that the numbers can be checked rather than trusted:

1. Rates. Taken from each provider’s official pricing page in September 2026: platform.openai.com/docs/pricing, anthropic.com/pricing, ai.google.dev/gemini-api/docs/pricing, api-docs.deepseek.com/quick_start/pricing. Standard service tier only. Batch, flex, priority and fast tiers are excluded except where the tables say otherwise.
2. Cost formula. cost = input_M × (1 − hit) × input_price + input_M × hit × cached_price + output_M × output_price. Cache rates apply to input tokens only.
3. Long-context handling. Where a provider documents a threshold, the multiplier is applied to the whole request — input and output both — not just to the tokens above the threshold. That matches the vendors’ published wording.
4. Cache break-even. reads = cache_write_multiplier / (1 − cache_read_ratio), using the 1.25× write multiplier that OpenAI and Anthropic both publish. Providers that do not publish a write rate are omitted from that table rather than estimated.
5. Workload volumes. The six workloads are illustrative, not benchmarks. Their token counts are stated so you can substitute your own. Nothing here is a measurement of speed, quality or reliability — this index is about price only.

Two caveats worth repeating. List prices change without notice, and Google’s promotional Flash pricing expires on 1 January 2027. And a cheap model that fails the task is more expensive than an expensive model that succeeds — the point of this index is to show where the price differences are large enough to be worth testing, not to declare a winner.

Frequently asked questions

Why do the same token counts cost different amounts on different models?

Because providers set their own rates, and the input and output rates move independently. Output is where the divergence is widest: at the time of writing, output rates across these 22 models span from $0.40 to $75 per million tokens — a 187× range — while input rates span 125×.

Should I just pick the cheapest model?

No. The right conclusion is that the spread is large enough to justify routing different traffic to different models. Classification and extraction rarely need a frontier model; multi-step reasoning and code generation often do. The saving comes from matching, not from minimising.

How much does caching actually save?

On the workloads above, a 70% cache hit rate cuts 40–60% off the input portion of the bill. Since input is usually the larger token count but the cheaper rate, the total saving typically lands between 20% and 40%. The tables show the exact figure for each workload.

Is chunking a long document always safe?

No. It is safe when each chunk can be understood independently — extraction, summarisation, per-section classification. It breaks tasks that require cross-document reasoning, comparison, or a global view. Test on your own evaluation set before shipping it.

How often should I re-run this analysis?

Quarterly at minimum, and immediately after any provider announces a price change. Two of the models here carry promotional pricing with a known expiry date, and providers routinely retire models on short notice.

Related guides: How to Reduce AI API Costs · Prompt Caching Explained · Batch AI APIs · How AI API Pricing Works · AI API Cost Calculator