LLM API Price War Intensifies: 66 Models Priced Up to 95% Below Official Rates

If you thought LLM API pricing had already hit rock bottom, this real-time catalog from the aggregation platform UplinkSwitch might change your mind. The platform pulled 66 chat models from 11 developers with official list pricing from its own API gateway's real-time catalog and placed "official list price" side by side with "actual platform selling price." The result is striking: all 66 models are priced below official list rates on both input and output sides, with the largest discount reaching 95%.

It's worth clarifying that this data reflects the discount a reseller/aggregation platform offers relative to official list prices — not the absolute level of each developer's official pricing itself. But even with that lens, it remains a useful sample for observing the reality of LLM inference costs.

Value Snapshot: Each Developer's Lowest-Priced Model

The original article sorts by "output price first, input price second," filters for each developer's lowest-priced model (skipping models that fail real-device validation), and lists the comparison against official list prices:

Developer	Lowest-Priced Model	Input Price	Output Price	Official List Price
Anthropic	Claude Haiku 5.5	$0.06	$0.15	$0.10 / $0.50
DeepSeekAd	DeepSeek V4 Flash 0423	$0.0279	$0.0558	$0.30 / $1.20
Google	Gemini 3.1 Flash Lite Preview	$0.15	$0.90	$0.25 / $1.50
MiniMax	MiniMax M2.5	$0.162	$0.648	$0.27 / $1.08
Moonshot AI	Kimi K2.5	$0.27	$1.35	$0.45 / $2.25
OpenAI	GPT-6 Luna	$0.0625	$0.15	$0.10 / $0.50
Qwen	Qwen3.7 Plus	$0.192	$0.768	$0.32 / $1.28
Tencent	Hy3	$0.0417	$0.1668	$0.139 / $0.556
xAI	Grok 4.20 Multi-Agent (0309)	$0.75	$1.50	$1.25 / $2.50
Xiaomi	MiMo V2.6 Flash	$0.084	$0.09	$0.14 / $0.28
Z.ai	GLM 5.3 Flash	$0.084	$0.15	$0.15 / $0.50

A few signals worth noting:

• The low-price segment is already dominated by Chinese vendors and lightweight models. DeepSeek V4 Flash, Tencent Hy3, OpenAI GPT-6 Luna, and Anthropic Claude Haiku 5.5 all have input prices down to a few cents per million tokens.
• The gap between official and channel prices is widespread. Most models in the table see discounts of 40%–70%, with DeepSeek V4 Flash 0423 reaching 90% off on input and 95% off on output.
• xAI is the most expensive in this sample. Its lowest-priced model costs $0.75 input and $1.50 output — clearly higher than other vendors' entry-level offerings.

Breaking Down Each Developer's Pricing Structure

Anthropic: 12 Models, Clean Tiering

Anthropic's lineup runs from Claude Haiku 5.5 ($0.06/$0.15) all the way up to Claude Fable 5 ($3.00/$15.00) — a roughly 50x spread. One detail: the four models from Opus 4.6 to Opus 5 are priced identically ($1.50/$7.50), while Opus 5.5 actually drops slightly to $2.10/$6.00 with 47%/70% discounts. Sonnet 4.6's cache pricing is listed as "at input rate," meaning this model doesn't offer a separate cache discount tier.

DeepSeek: The Largest Discounts in the Entire Dataset

Among DeepSeek's 6 models, V4 Flash 0423's 90%/95% discount is the most extreme in the whole dataset. Its sibling V4.1 Flash also carries 80%/85% discounts. The more premium V4 Pro series narrows to 40%–70%, suggesting that the more flagship the model, the less room there is for channel negotiation.

Google: 11 Models, Dense Flash-Series Iteration

Gemini's lineup is almost entirely filled with Flash variants: 3.5 Flash, 3.6, 3.7, and 3.8 Flash are all priced at $0.45/$2.25 with a uniform 40% discount. The true flagship, Gemini 4 Argon, is priced at $0.60/$3.00 — a 70% discount off the official $2.00/$10.00. Gemini 3.1 Flash Lite Preview is the lowest across the entire line ($0.15/$0.90).

OpenAI: 13 Models, the Richest Discount Gradient

OpenAI has the most models of any developer. GPT-6 Luna ($0.0625/$0.15) and Claude Haiku 5.5 are priced almost identically, suggesting entry-level market pricing is converging. GPT-6 Astra and Claude Fable 5 share the same official list price of $10.00/$50.00, placing them in the flagship tier. One detail: GPT-6 Luna's input discount is only 37%, one of the few figures in the table below 40%.

Other Developers

• Moonshot AI: 4 Kimi models, with prices rising from K2.5 to K3 ($0.27/$1.35 up to $0.90/$4.50) — a major version jump accompanied by roughly a 3x price increase.
• MiniMax, Tencent, Qwen, Xiaomi, Z.ai: 2–4 models each, with discount rates holding steady at 40%–70% — standard channel discounts.
• xAI: 5 Grok models, with 4.3 and later versions uniformly discounted at 70%, though absolute prices remain the highest in the entire table.

What "per million tokens" pricing and model tiering actually mean

LLM APIs are universally priced as "USD per million tokens," split between input (prompt) and output (completion), with output typically 4–5x the input price. This is because autoregressive generation requires serial token-by-token computation, while input can be prefilled in parallel. Within a single developer, models are typically tiered by capability and cost: lightweight/Flash/Haiku-class models target high-concurrency, low-complexity tasks, while flagship models target complex reasoning — with prices that can differ by dozens of times. This is the structural reason behind the "roughly 50x spread from $0.06 to $3.00 input price" mentioned in the article: the price gradient reflects differences in model scale and inference compute investment, not merely market pricing strategy.

How to Interpret "95% Below Official Price"

This is the part of the data most likely to be misread. The original methodology is explicit: list prices are the per-million-token rates developers themselves publish, read from official pricing pages; percentages are rounded down; and 0% means flat or above list price.

In other words, these figures describe "the discount of aggregation platform prices relative to official list prices," not "the models' official prices themselves." For developers, choosing an aggregation platform can lower short-term costs, but it also means entrusting availability, latency, and compliance responsibility to an intermediary. The original article itself notes: when citing this index, cite the index data source rather than this article, because prices change in real time with each request.

Worth noting is the data consistency handling: dated snapshot versions and different effort variants are merged into the base model; models without official list prices and third-party modified versions are excluded. This keeps the comparison baseline relatively clean.

The aggregation platform / API reseller business model

Aggregation platforms (also called API gateways or resellers) don't train models — they buy inference capacity in bulk from model developers and resell it to end users on a pay-as-you-go basis. Their discount room typically comes from three sources: volume-purchase tiered discounts, the platform's own customer-acquisition subsidies, and arbitrage on idle compute. For developers, the value of such platforms lies in a unified billing entry point and convenient multi-model switching, but the cost is introducing an intermediary layer: requests take an extra hop, and latency and availability depend on the platform's own stability; data also flows through a third party, so compliance-sensitive scenarios need extra evaluation of data retention and cross-border transfer terms. So the "95% discount" is fundamentally a concession on the procurement path, not a decline in model capability or inference cost.

The billing logic of prompt caching

LLM API caching mechanisms allow developers to pre-write recurring long prefixes (such as system prompts, long documents, and few-shot examples) into a cache. Subsequent requests that hit the cache are billed at a lower "cache read price," typically 10%–20% of the input price, with some vendors also charging a one-time "cache write price." This significantly lowers costs for long-context, multi-turn conversation, and RAG scenarios. The article notes that Anthropic Sonnet 4.6, the OpenAI GPT-4o series, and the Tencent Hy series are marked as "at input rate" — meaning these models did not offer a separate cache discount tier at the time, and cache hits are billed at the regular input price. When building cost models for repeated long-context calls, these must be accounted for separately rather than assuming other models' cache discount assumptions apply.

Practical Implications for Developer Selection

Several actionable judgments can be drawn from this price table:

1. Entry-level model costs are approaching "utility prices." More than one model has an input price of a few cents per million tokens, making them suitable for high-volume, low-complexity tasks like classification, extraction, and simple summarization.
2. The price gradient across flagship models is far steeper than mid-tier. The 50x spread from $0.06 to $3.00 input price shows that task difficulty — not pure price comparison — remains the first decision in model selection.
3. Cache discount differences are significant. Many models in the table have cache prices at 10%–20% of input price, but Anthropic Sonnet 4.6, the OpenAI GPT-4o series, and the Tencent Hy series are marked "at input rate" with no cache benefit — scenarios involving repeated long-context calls need separate accounting.
4. Don't look only at discount rate. A 40% discount on a $0.10 model versus a 40% discount on a $10 model represents a 100x difference in absolute savings.

The original article also provides raw data in JSON and CSV formats (cached for five minutes, with a generation timestamp header) and pricing pages broken down by model family, covering eight series: Claude, GPT, Gemini, Grok, DeepSeek, Kimi, GLM, and Qwen. For teams building cost models, using the structured data directly is more reliable than reading the article.

Conclusion

The greatest value of this monthly price index isn't any specific number — it's the structure it reveals: entry-level prices in the LLM inference market are rapidly converging, flagship prices remain clearly differentiated, and the gap between official and channel prices has become the norm. For developers, the real cost optimization space lies both in model selection and in the procurement path.

One-line Summary

All 66 LLM APIs on aggregation platforms are priced below official rates, with discounts up to 95%, revealing inference cost structure and procurement path differences.

Paragraph Summary

The article draws on the real-time catalog of aggregation platform UplinkSwitch to compare official list prices against actual platform selling prices for 66 chat models from 11 developers. It finds that all models are priced below official rates on both input and output sides, with the largest discounts reaching 90% on input and 95% on output (DeepSeek V4 Flash 0423). The author breaks down pricing structures by developer: entry-level model prices have converged to a few cents per million tokens, flagship model prices span roughly 50x, and xAI has the highest absolute prices. Cache discounts show structural differences, with some models offering no separate cache benefit. The article emphasizes that these figures reflect channel discounts rather than official pricing itself, and advises developers to consider task difficulty, caching strategy, and absolute savings — not just discount rate — when selecting models.