After the GPT-5.6 Price Reduction, What's the Actual Price Difference Between the Latest Large Model APIs at Home and Abroad

Fri, 31 Jul 2026 10:23:33 +0800

The problem with the previous version was not in the calculations, but in sample selection: it picked “representative SKUs,” while users wanted “the latest models from each manufacturer.” Putting the old ERNIE 5.0, GLM-5, Kimi K2.6, and MiniMax M2.7 into a single table labeled “latest prices” made the percentages that followed incomparable.

Only one rule is adopted this time: for each company, only the latest proprietary text model that is publicly available through API is included. OpenAI is an exception that needs to be expanded rather than compressed—GPT-5.6 has three tiers in the same generation: Sol, Terra, and Luna, all of which are available through the API, and Terra alone cannot represent OpenAI as a whole.

First, let’s clarify exactly what was reduced in this “ChatGPT price cut.” On July 30, OpenAI adjusted the standard API price of GPT-5.6 Terra from $2.50 / $15 to $2 / $12, and Luna from $1 / $6 to $0.20 / $1.20; Sol remained unchanged. This is a change in API usage-based pricing, not a discount on ChatGPT/Codex subscription monthly fees. The announcement also clearly states that subscription prices and package quota budgets remain unchanged; only the quota deductions for Terra and Luna in their respective products have been reduced.

A Price List for the Latest Models Only

Synthetic price = 1M uncached input token + 1M output token. Terra’s $2 + $12 = $14, noted as 100%. The RMB quotation uses a fixed exchange rate of $1 = ¥6.7939, used only for sorting; it does not include caching, Batch/Flex/Fast, tool calls, taxes, or enterprise agreement pricing.

Rank	Vendor	Latest Publicly Available API Model	Standard Price (Input / Output, per 1M tokens)	Synthesized Price (USD Estimate)	Relative to Terra
1	Anthropic	Claude Fable 5（2026-06-09）	$10 / $50	$60.00	428.6%
2	OpenAI	GPT-5.6 Sol（2026-07-09）	$5 / $30	$35.00	250.0%
3	Moonshot AI	Kimi K3	$3 / $15	$18.00	128.6%
4	OpenAI	GPT-5.6 Terra（baseline）	$2 / $12	$14.00	100.0%
5	Google	Gemini 3.6 Flash（2026-07-21 GA）	$1.5 / $7.5	$9.00	64.3%
6	xAI	Grok 4.5（2026-07-16）	$2 / $6	$8.00	57.1%
7	Alibaba Cloud	Qwen3.7-Max（2026-06-08 stable snapshot）	¥12 / ¥36	$7.07	50.5%
8	Volcengine	Doubao-Seed-Evolving（rolling SKU updates）	¥6 / ¥30	$5.30	37.9%
9	Baidu	ERNIE 5.1（input ≤32K）	¥4 / ¥18	$3.24	23.1%
10	MiniMax	MiniMax-M3（input ≤512K）	¥2.1 / ¥8.4	$1.55	11.0%
11	OpenAI	GPT-5.6 Luna	$0.20 / $1.20	$1.40	10.0%
12	DeepSeek	DeepSeek-V4-Pro（current public preview flagship）	$0.435 / $0.87	$1.31	9.3%
13	Tencent	Hy3（Guangzhou region）	¥1 / ¥4	$0.74	5.3%

“Latest” in the table comes in two not-so-neat forms: Doubao-Seed-Evolving is a rolling update under the same model ID, without a fixed release date like a regular version number; DeepSeek-V4-Pro is a preview series for the current public API flagship. They still align better with the table’s standard than historical SKUs that have been replaced by later models.

Models that have an API but are not included in the price ranking

Vendor	Latest Public API Model	Why Not Calculate a Percentage
Zhipu (Zhipu AI)	GLM-5.2	Officially confirmed as the latest flagship with API support, but the publicly crawlable official page does not list per-token pricing; the old pricing of GLM-5.1 cannot be used as a substitute.
Cohere	Command A+	Trial and production keys come with limited free usage, while production deployment goes through Model Vault; Cohere has not published per-token pricing that can be compared horizontally with the main table.
Mistral	Leanstral 1.5	A free API endpoint is offered, but it is dedicated to Lean 4 formal verification; it is not a $0 pricing for a general-purpose production API and cannot be included in this general task cost table.

This is not an omission—it’s exactly what you encounter when “looking only at the latest”: some new products only offer trial credits first, some are sold exclusively through enterprise channels, and some only provide model descriptions in public materials without listing per-token pricing. Filling the table with previous-generation prices might make it look complete, but it would actually turn the issue from “temporarily not comparable” into an incorrect number.

What This Price Adjustment Means Within OpenAI’s Own Three Tiers

Sol’s composite price is 2.5 times that of Terra (150% higher); Terra is in the middle; Luna is 10% of Terra. Therefore, this price reduction of Terra and Luna does not mean that “flagship models have become cheaper across the board.” Instead, OpenAI has further widened the gap between the mid-tier and low-cost tiers while keeping Sol positioned at the premium price point.

OpenAI’s stated reason is improved inference efficiency: the model, inference system, and the agent layer connecting tools and context are all being optimized. The announcement states that production kernel optimization reduced end-to-end service costs by 20%, and experiments improved token generation efficiency by more than 15%. This is a vendor disclosure that can explain why there is room for price reductions, but it is not a complete cost table available for external audit. As for whether the low price is mainly aimed at competing for high-volume API workloads, it can only be an analysis of the pricing structure, not an official motive.

This Table Is Not a Substitute for Model Selection

Price sorted from expensive to cheap does not mean tasks sorted from good to bad. Long context triggers tiered pricing, and cache hits, reasoning tokens, tool calls, retry counts, region, and enterprise contracts all change the actual bill. The “1M input + 1M output” in the table is only a unified filter before procurement.

When making a real choice, it’s more reliable to test with ten to twenty actual tasks: first look at success rate, latency, and human rework, then record input, output, reasoning tokens, cache hits, and total cost. What you should ultimately compare is the “total cost per successful task,” not an isolated per-unit price on this table.

References

OpenAI: GPT-5.6 Release and Three-Tier API Pricing, published on 2026-07-09, accessed: 2026-07-31.
OpenAI: GPT-5.6 Price-Performance Announcement, published on 2026-07-30, accessed: 2026-07-31.
OpenAI API: GPT-5.6 Terra Model Page, accessed: 2026-07-31.
Anthropic: Claude Fable 5, accessed: 2026-07-31.
Google: Gemini API Changelog and Model Pricing, accessed: 2026-07-31.
xAI: Grok 4.5 and Pricing Information, accessed: 2026-07-31.
Alibaba Cloud Bailian: Model Pricing, accessed: 2026-07-31.
DeepSeek: V4 Release and Pricing Information, accessed: 2026-07-31.
Zhipu: Latest Model Releases and Model Overview, accessed: 2026-07-31.
Volcano Engine: Doubao Large Model and Pricing, accessed: 2026-07-31.
Tencent Cloud: Hy3 Model List and Pricing, accessed: 2026-07-31.
Baidu: ERNIE 5.1 Release and Qianfan Model Service Pricing, accessed: 2026-07-31.
Moonshot AI: Kimi Open Platform, accessed: 2026-07-31.
MiniMax: Pay-As-You-Go Pricing, accessed: 2026-07-31.
Cohere: Command A+, accessed: 2026-07-31.
Mistral: Leanstral 1.5, accessed: 2026-07-31.
Bank of China: Foreign Exchange Quotation, using the USD spot exchange selling rate on 2026-07-24, accessed: 2026-07-31.

写作附记

Writing Instructions

This article is a snapshot of public API list prices as of 2026-07-31 and does not constitute advice on procurement, investment, or product selection. It only includes verified vendors and does not represent an exhaustive list of all API providers worldwide; model capabilities, tokenizers, context, caching, inference tokens, tool charges, regions, and enterprise contracts vary, and the percentages in the table should not be treated as equivalent performance or actual task cost rankings.

Original Prompt

ChatGPT’s series of models have officially announced price reductions. Let’s organize the relevant information: why are they cutting prices? Compile the API prices of major model providers, both domestic and international, into a table. Using ChatGPT 5.6 terra as the baseline, calculate the price comparison percentages of each provider in the table, sorted from most expensive to cheapest.

The compiled table information is incomplete—other ChatGPT models aren’t listed, and the models from major providers aren’t the latest released ones. I only need the latest model prices from each provider.