LLM API pricing

How llm-provider access is priced — per-token, split input vs output, with output 2–6× more expensive than input across providers llm-api-pricing-comparison. The 2026 market spans a ~600× cost spread.

The September repricing (2026-09-22)

One day, two labs, and every price moved down (opus-and-sol-and-luna-willison, checked against openai-api-pricing-docs and anthropic-model-docs):

OpenAI halved its line-up in a new generation. GPT-6 Sol at $2/$10 is half of GPT-5.6 Sol's promotional $4/$20; GPT-6 Luna at $0.10/$0.50 is half of GPT-5.6 Luna and sits at the market's floor beside Mistral Small 3.2. GPT-6 Astra ($10/$50) is the new premium tier.
Anthropic cut its flagship 20%. Opus 5.5 at $4/$20 against Opus 5's $5/$25, with cache hits at 0.05× input instead of 0.1×.

The result is a set of exact price matches across labs. GPT-6 Sol equals Sonnet 5 at $2/$10. Opus 5.5 equals GPT-5.6 Sol's promotional rate at $4/$20. The two premium tiers (Fable 5.1, Astra) share $10/$50. Whatever the labs compete on above the floor, it is not list price, since the lists now agree to the dollar at three points.

Three tiers (mid-2026 snapshot — volatile, cite the source/date)

Budget — Mistral Small 3.2 $0.10/$0.30 (in/out per 1M); GPT-4.1 Nano $0.10/$0.40; deepseek V3.2 $0.27/$1.10.
Mid / production — GPT-5.4 $2.50/$15; Claude Sonnet 4.6 $3/$15; Gemini 3.1 Pro $2/$12; Claude Sonnet 5 $2/$10 (2026-08-30, anthropic-model-docs).

A successor priced below the model it replaces (2026-08-30). Sonnet 5 launched at $2/$10 billed as introductory through 2026-08-31, with $3/$15 — Sonnet 4.6's rate — scheduled for 1 September. Anthropic's pricing docs now state the increase will not occur and $2/$10 is the standard price. So the mid tier's own floor moved: the current Sonnet undercuts the previous Sonnet by a third at the same position in the line-up. Every "the frontier premium survives / doesn't" reading on this page rests on sticker prices that a vendor can quietly decline to raise, and this is the first instance the spoke has caught of that happening rather than the reverse.

Frontier / reasoning — GPT-5.4 Pro $30/$180; o3 $15/$60; Claude Opus 4.7 $5/$25.

Frontier snapshot (mid-2026, per 1M in/out — grok-4-5-price-vs-benchmarks): Grok 4.5 $2/$6, Opus 4.8 $5/$25, GPT-5.5 $5/$30, Fable 5 $10/$50. Grok 4.5 undercuts the whole group; Fable 5 sits at the top — pricier than Opus, corroborating the Fable-is-expensive read.

Budget tier, frontier comparison set (2026-08-26): Qwen3.8-Flash lists at $0.15 / $0.47 per 1M with 1M context — inside the budget band above, on a model its maker benchmarks against Claude Opus 4.6. Whatever the benchmarks are worth, the rate card is the clearest statement yet of what the collapsing floor is doing to the middle.

A third axis: priced by who you are

The rate card above and the subscription bucket below both bill usage. research-access-programs · claude-message-batches-api starts somewhere else: a provider gives a class of user access on terms nobody else can buy. Anthropic's scientist program (2026-08-27) opens 10,000 seats for one year — standard free, premium $15/month for 5× the usage limits — to any principal investigator at an academic or nonprofit institution, with up to $50,000 in credits per project available on top through its AI for Science program. The base those limits multiply is unpublished, so the discount cannot be converted to a per-token figure and does not belong in the tables above. What makes it a pricing fact anyway is the second gate: the same post caps biology and chemistry researchers at Opus-class models, so a subsidized buyer is handed a lower price and a lower ceiling at once.

A different axis: speed-priced, subscription buckets

Most of the market above competes on a per-token rate card. Cerebras doesn't: as a wafer-scale inference provider (cerebras-inference) reselling open models, it sells throughput, leads on "20× faster," and publishes no per-million-token price at all (corrected 2026-09-22: the pricing page now lists GPT OSS 120B at $0.35/$0.75 and Qwen 3.8 27B at $0.99/$1.49 per million input/output tokens cerebras-pricing, so Cerebras now runs both shapes). Its paid access is structured as daily-token subscription buckets — Cerebras Code Pro $50/mo → 24M tokens/day, Max $200/mo → 120M tokens/day — plus a $5-credit free trial and a self-serve Developer tier from $10 cerebras-pricing. Both coding tiers were listed sold out (mid-2026, and still on 2026-09-22), a capacity signal: fast inference is rationed, not just priced. So the market now has two pricing shapes — the per-token rate card, and the speed-provider's flat-subscription-plus-cap — and cost-per-task comparisons must account for both.

Cost levers (stack to ~25% of list)

Prompt caching — 50–90% off repeated input prefixes (aggressive at anthropic/OpenAI). At Anthropic it's also a throughput lever, not just a discount: under cache-aware ITPM, cached input tokens don't count toward the rate limit either claude-api-rate-limits.
Batch processing — flat 50% off for non-real-time work, applied to input, output and special tokens alike. At anthropic the ceiling is 24 hours (most batches finish inside an hour) and unprocessed requests expire unbilled; batch traffic also has its own rate limits and doesn't spend the Messages-API budget, so like caching it buys throughput as well as price claude-message-batches-api. It stacks with caching only weakly — cache hits inside a batch are best-effort, because the requests run in any order.
Model routing — send simple tasks to budget models; cuts 70–90% of qualifying calls. Bought as a product this is the ai-gateway layer: omniroute ships 19 routing strategies with cost-optimized and headroom variants that pick the cheapest backend still holding quota.
Prompt compression — squeeze the prompt before it's billed. omniroute stacks 12 engines and claims 15–95% savings (89% "average", self-reported). Unlike caching and batching this lever is client-side: it needs no cooperation from the provider, and it attacks tokens-per-task rather than price-per-token.
Free-tier aggregation — treat many providers' free pools as one budget. omniroute totals ~1.53B tokens/month across 43 pools, with 15 of them flagged as ToS-questionable. The lever works; the hazard comes attached.

Strategic reads

deepseek disrupted the bottom (OpenAI-compatible + ultra-low cost), pulling the floor down; google undercuts on sticker price; OpenAI offers the widest range; Anthropic competes on consistency + caching. A US frontier lab has now joined the price war: Grok 4.5 prices below the field and argues cost-per-task (per-token price × tokens-per-task) is the metric that matters — xAI claims 4.2× fewer tokens than Opus 4.8, so a lower benchmark still wins on total cost (grok-4-5-price-vs-benchmarks). That reframes the sticker-price comparisons above: tokens-per-task is a hidden multiplier on the headline rate. Because output dominates cost, context/output optimization is the highest-leverage spend control. Pricing is the most volatile fact in this wiki — always date it.

llm-api-pricing-comparison · deepseek · anthropic · llm-provider · llm-benchmarks · cerebras-pricing · cerebras-inference · ai-gateway · omniroute · research-access-programs · claude-message-batches-api