AI API Pricing in August 2026: A Reproducible 11-Model Baseline

AI API price tables look comparable because they all use dollars per million tokens. They are not comparable until you align the model status, billing tier, context tier, cache semantics, time conditions, and token mix. A copied “input price” can be a standard rate, a discounted batch rate, a promotional rate, or an off-peak rate. Treating those as one column produces a clean-looking table and a misleading estimate.

This baseline solves a narrower problem: it records one explicit calculator default for 11 directly priced API models and shows how to reproduce a single workload calculation. It is a dated reference, not a real-time feed or a recommendation to buy a particular model. It is not a capability ranking.

The source pages were checked on August 22, 2026. The 11 rows were manually approved by the TokenCostAI site owner on August 22, 2026. Prices can change after that date, so confirm the linked provider page before making a purchasing decision.

Companion data files:

• pricing-baseline.json contains the normalized rates, scope notes, source links, and worked-scenario results.
• pricing-baseline.csv contains the same 11 rows in a spreadsheet-friendly form.

The comparison contract

Every primary price in this baseline uses USD per one million tokens. The three calculator columns mean:

• Fresh input: input that is billed at the model’s selected default input rate.
• Cached input: an eligible cache hit or cached-read rate, not a cache-write charge.
• Output: model-generated tokens, including reasoning or thinking tokens when the provider includes them in output billing.

The default is the provider’s paid Standard direct-API rate where that concept exists. For OpenAI, this baseline uses the Standard short-context tier. For Anthropic, the cached-input column is the cache-hit and refresh rate; five-minute and one-hour cache-write prices remain separate metadata. For Google, the primary row is the paid Standard rate rather than Batch, Flex, or Priority. For DeepSeek, the primary row uses the higher peak rate so a default estimate does not quietly assume that every request lands in an off-peak window.

That scope is deliberate. It creates a reproducible starting point, but it does not erase provider-specific rules. If your workload uses long context, cache writes, batch processing, data-residency endpoints, audio, images, tools, grounding, or storage, the three-column result is incomplete.

The August 22 baseline

Provider	Model	Fresh input	Cached input	Output	Default rate basis
OpenAI	GPT-5.6 Sol	$5.00	$0.50	$30.00	Standard, short context
OpenAI	GPT-5.6 Terra	$2.50	$0.25	$15.00	Standard, short context
OpenAI	GPT-5.6 Luna	$1.00	$0.10	$6.00	Standard, short context
Anthropic	Claude Opus 5	$5.00	$0.50	$25.00	Base input, cache hit, output
Anthropic	Claude Sonnet 5	$2.00	$0.20	$10.00	Base input, cache hit, output
Anthropic	Claude Haiku 4.5	$1.00	$0.10	$5.00	Base input, cache hit, output
Google	Gemini 3.7 Flash	$0.75	$0.075	$3.75	Paid Standard promotional rate through 2026-12-31
Google	Gemini 3.5 Flash	$1.50	$0.15	$9.00	Paid Standard
Google	Gemini 3.5 Flash-Lite	$0.30	$0.03	$2.50	Paid Standard
DeepSeek	DeepSeek V4 Flash	$0.44	$0.014	$1.32	Peak rate used as conservative default
DeepSeek	DeepSeek V4 Pro	$1.32	$0.044	$3.96	Peak rate used as conservative default

These rows answer “what rate did this calculator use?” They do not answer “which model is best?” Capability, latency, task completion, retries, context limits, regional availability, support, and non-token fees all affect cost per completed task.

A worked calculation with one token mix

To make the table testable, use the same simulated monthly workload for every row:

• 1,000 requests;
• 2,000 fresh input tokens per request;
• 1,000 eligible cached-input tokens per request;
• 500 output tokens per request.

The monthly totals are 2,000,000 fresh input tokens, 1,000,000 cached-input tokens, and 500,000 output tokens. Because all rates are per million tokens, the formula simplifies to:

Scenario cost = (2 × fresh-input rate) + (1 × cached-input rate) + (0.5 × output rate)

Model	Fresh-input cost	Cached-input cost	Output cost	Scenario total
GPT-5.6 Sol	$10.000	$0.500	$15.000	$25.500
GPT-5.6 Terra	$5.000	$0.250	$7.500	$12.750
GPT-5.6 Luna	$2.000	$0.100	$3.000	$5.100
Claude Opus 5	$10.000	$0.500	$12.500	$23.000
Claude Sonnet 5	$4.000	$0.200	$5.000	$9.200
Claude Haiku 4.5	$2.000	$0.100	$2.500	$4.600
Gemini 3.7 Flash	$1.500	$0.075	$1.875	$3.450
Gemini 3.5 Flash	$3.000	$0.150	$4.500	$7.650
Gemini 3.5 Flash-Lite	$0.600	$0.030	$1.250	$1.880
DeepSeek V4 Flash	$0.880	$0.014	$0.660	$1.554
DeepSeek V4 Pro	$2.640	$0.044	$1.980	$4.664

The totals are arithmetic outputs, not measured application bills. The scenario assumes that all stated cached tokens qualify and hit the cache. It excludes cache writes, storage, Batch/Flex/Priority differences, long-context uplifts, tools, search grounding, image and audio tokens, retries, taxes, and exchange-rate effects.

Why the OpenAI rows specify short context

The OpenAI API pricing page separates Standard pricing from other processing modes and distinguishes short- and long-context rates for relevant models. A single “GPT-5.6 price” without that scope is incomplete. This baseline records the Standard short-context table because it is a clear direct-API default for the site calculator.

That choice does not mean a long-context request is billed at the same rate. It also does not mean Batch, Flex, or Priority can be substituted without changing the calculation. A later calculator revision can expose those modes as separate options; until then, the visible rate-basis label is part of the number.

OpenAI also lists a cache-write price for the GPT-5.6 rows. The current three-field calculator does not charge cache writes, so those values are retained in the JSON metadata instead of being folded into cached input. Treating a write as a read would understate the first cached request.

Why Claude cache writes remain separate

The Anthropic pricing documentation publishes base input, five-minute cache-write, one-hour cache-write, cache-hit/refresh, and output columns. Only the cache-hit/refresh column maps to this baseline’s cached-input field.

For example, Claude Sonnet 5 is recorded at $2 per million base-input tokens, $0.20 per million cache-hit tokens, and $10 per million output tokens. Its five-minute and one-hour cache-write rates are higher than the base-input rate and remain separate fields in the JSON asset. A workload that creates many cache entries but rarely reuses them can cost more than a hit-only estimate suggests.

Anthropic’s page also states that the $2/$10 Sonnet 5 input/output rates are standard rather than ending on August 31, 2026. That dated clarification is why the baseline does not attach an August expiration to the Sonnet row.

Why Gemini 3.7 Flash needs an expiration date

The Gemini Developer API pricing page lists multiple processing modes and, for Gemini 3.7 Flash, a Standard promotional price through December 31, 2026. The baseline uses $0.75 input, $0.075 cached input, and $3.75 output per million tokens for that dated period.

The same source lists higher Standard rates beginning January 1, 2027: $1.50 input, $0.15 cached input, and $7.50 output. Both schedules are preserved in the JSON asset. A maintenance check must warn before the promotion expires; silently leaving the promotional number in place would turn a verified snapshot into stale data.

Google also offers Batch pricing at lower rates. Those values are metadata, not the default row, because a workload must satisfy Batch behavior and latency constraints before claiming the discount. The visible Standard result keeps an interactive calculator from implying that an asynchronous processing mode applies to every request.

Why DeepSeek uses peak pricing by default

The DeepSeek pricing page publishes separate cache-hit, cache-miss, and output rates for peak and off-peak periods. The primary rows use peak prices: $0.44/$0.014/$1.32 for V4 Flash and $1.32/$0.044/$3.96 for V4 Pro.

Off-peak prices are stored separately: $0.22/$0.007/$0.66 for V4 Flash and $0.66/$0.022/$1.98 for V4 Pro. They are useful for a scheduled workload, but they are not a safe universal default. Peak windows and weekend rules are time-dependent provider conditions. A budget calculator should not assume a discount unless the user deliberately selects it and understands the schedule.

Why Llama is not in this table

The previous TokenCostAI catalog included illustrative Llama hosting prices under Meta. That was not a sound direct-provider comparison. Meta publishes open-weight models, but there is no single universal Meta API rate that represents every cloud host, inference platform, region, quantization, or self-hosted GPU setup.

Llama can return in a future comparison only when each row names the actual hosting provider and links to that provider’s price page. Self-hosting requires a different cost model based on hardware time, utilization, storage, networking, and operations. Inventing one per-token “Meta price” would hide those choices.

How to reuse the data responsibly

Start with the JSON or CSV asset rather than copying the rendered table. Keep the currency and per-million-token unit attached to every value. Preserve the source URL, review date, model status, and rate-basis note. If you change the simulated token mix, recompute the totals instead of scaling a rounded display value.

Before using the data for procurement:

• open the provider source and confirm the model still exists;
• confirm the processing tier, context tier, region, modality, and cache behavior;
• use provider-reported token counts rather than character estimates where possible;
• add cache writes, storage, tools, grounding, retries, and other applicable charges;
• compare task success, latency, and retry rate alongside token cost;
• record a new review date when any value changes.

This baseline is intentionally conservative about what it claims. It provides a transparent calculator input and a reproducible arithmetic example. It does not convert a price snapshot into a promise about model quality, future pricing, or the cheapest completed task.

Method and limitations

The source catalog follows a manual publication gate: collect official pages, create a candidate, document differences, obtain per-row human approval, package the approved file, and only then prepare a website update. The approved candidate and generated package remain versioned local files. No database or paid data feed is required.

The verification script checks that all 11 rows match the approved candidate, that IDs are unique, that source URLs and rate-basis notes exist, and that each displayed scenario total can be recomputed from the stated formula. This catches transcription and arithmetic errors. It cannot detect a provider changing its website tomorrow, which is why the visible review date and official source links remain necessary.

Editorial review: the price rows and this article were separately approved by the site owner on August 22, 2026. The visible review date is not a claim that prices update in real time.