LLM API pricing in 2026: what inference really costs

$0.14 per million input tokens. That is DeepSeek V4 Flash, the cheapest general-purpose API on the market in August 2026, and its price has not moved in months. The ceiling on the same list: $30 per million input tokens for OpenAI's GPT-5.5-pro. Two orders of magnitude between floor and ceiling, with the whole stack shifting underneath.

This is a snapshot of first-party list prices normalized to US dollars per million tokens at the standard, non-batch tier, collected from provider pricing pages in early August 2026. Providers below are named; the numbers are theirs.

The price floor keeps sinking

DeepSeek built the V4 generation around two price points. V4 Pro, the flagship, sits at $0.435 input and $0.87 output per million tokens, below most Western mid-tier models. V4 Flash, the fast model, holds the floor at $0.14 input and $0.28 output, cheaper than almost anything else with a 1M-token context window. Automatic context caching makes repeated input cheaper still: a V4 Flash cache hit costs $0.0028 per million tokens.

On July 31 it updated V4 Flash to a new checkpoint with 280 billion total parameters and 13 billion active, and kept the price untouched. The model is more capable at the same $0.14/$0.28. The point of the release was not a price cut; it was shifting the comparison from unit price alone to how much work the same dollar completes.

None of this is a law of physics. DeepSeek has signaled a broader API price increase, and as of mid-August no replacement rates or an effective date have been published. The floor it set is a decision, and decisions can be revised. As we wrote in August, the ASIC-vs-GPU cost-per-token war in custom inference silicon is the structural force keeping this floor low, but the signaling from DeepSeek is the first sign that the cheapest lane in the market may stop getting cheaper.

The real price war moved to cache reads

Base input prices have been roughly stable through 2026. The competition went elsewhere: into the price of a cache hit. Anthropic, OpenAI and Google all price a cache read at 10% of base input, a 90% discount. DeepSeek prices its cache-hit lane at 98% off. xAI is the clearest illustration of the direction: it cut Grok 4.5's cache read from $0.50 to $0.30 between July snapshots, deepening that model's discount from 75% to 85% while leaving the $2.00 base input price untouched.

For agentic workloads this column predicts the invoice better than the headline input price. Every call re-sends the same system prompt, the same tool definitions, the same context prefix. Repeated input routinely dominates total token volume in production. A provider with a higher base rate and a cheaper effective caching lane can win on the real bill. That is why procurement math now runs the cache column before anything else.

There is a second divergence hiding in the same tables: flat versus tiered context pricing. Anthropic bills its 1M-token context window at standard rates, a 900K-token request costs the same per token as a 9K one. Google and xAI took the opposite path, roughly doubling per-token rates above 200K tokens. For long-context applications, the pricing structure matters more than the sticker rate.

Per-token is not per-task

Two models with identical $/1M stickers can produce meaningfully different invoices for the same document. Anthropic's newest tokenizer emits roughly 30% more tokens per text than its earlier models, heavier on code, structured data and non-English text, lighter on plain prose. Other providers publish no tokenizer efficiency data at all. Cross-provider per-token comparison is therefore an approximation, not an equivalence.

This is the argument for switching the unit of accounting. Cost-per-task captures what an invoice actually shows: the tokens burned per completed answer, including reasoning turns, retries, tool calls and context overhead. The models that look cheap per token are frequently the wrong choice for long-context or agentic workloads once the whole workflow is counted. Output tokens compound the effect: they are priced two to six times higher than input tokens across the board, so the models that generate more text per answer carry a multiplier that list-price tables hide.

Only 22% of organizations track AI spend by transaction. In a market where the gap between the cheapest and most expensive model is two orders of magnitude and the tokenizer can silently add a third of your bill, the visibility gap is the most expensive hidden cost in large language model (LLM) API economics.

The September calendar is the tell

Anthropic launched Claude Sonnet 5 in late June at an introductory $2 per million input tokens and $10 per million output, with an expiry date of August 31. On September 1 it was scheduled to step up to standard Sonnet rates of $3/$15. On August 11, Anthropic cancelled the step: the introductory pricing is now permanent.

The move reads as a response to a market that is not cooperating with price increases. OpenAI cut rates on GPT-5.6 Luna and Terra in July. DeepSeek's floor sits at $0.14. The most-quoted reply to the Sonnet 5 announcement put a competing model at $0.14 per million input tokens, the exact figure DeepSeek quotes for V4 Flash. In a market where list prices keep falling, a scheduled increase is an announcement of the direction you think the market is heading, and cancelling one is the reverse.

There is a structural reason Anthropic is defending this price point rather than walking it up. The company is reportedly in talks to acquire Decart for roughly $6 billion, an AI infrastructure startup squarely aimed at the inference-cost problem. Sonnet 5's pricing exists to displace Opus-class spend on long-running agent work, and the September step-up would have broken that displacement math. The deal, if it closes, is the clearest signal yet that the competitive front in this market is cost per unit of work, not model quality alone.

The buy decision is three levers, not one price

The most common mistake in LLM procurement is treating the input-price column as the decision. In 2026 the same workload can land at a 10x different invoice depending on three levers, none of which shows up on a pricing page.

The first is caching strategy. For agentic traffic, the system prompt and tool definitions are re-sent on every call, so cached input routinely dominates total token volume. A provider with a 90% cache discount and a deeper minimum-cacheable prefix can beat a cheaper headline rate on the real bill. The second is batch and off-peak processing. Providers discount non-urgent work meaningfully, so workloads that tolerate queueing can run at a fraction of interactive rates. The third is model routing: simple queries to a cheap fast model, complex reasoning to a frontier one. Hybrid routing cuts average cost by 40-60% in production deployments while holding output quality, because most traffic does not need the most expensive tier.

Each lever shifts the effective price of a finished task more than any base-rate difference between providers. That is why cost-per-task, not cost-per-token, is becoming the standard unit in serious procurement documents, and why the teams that optimize all three levers routinely run the same product at a quarter of the cost of teams that optimize none.

What the same dollar buys

List prices, per million tokens, standard tier, early August 2026:

Read the table as a floor-and-ceiling map, not a verdict. The floor keeps the whole market honest; the ceiling is what pays for the frontier. Between them sits the decision most teams actually face: which lane of this stack your workload belongs in, and whether your bill is driven by the price column or the cache column.