DeepSeek V4 Pricing Revealed: Flash Variant Ultra-Low Cost, New Gemini Model Surfaces

The price war in the large model API market is far from over. This week's LLM Price Watch tracking data surfaced several noteworthy new entries: pricing for two DeepSeek V4 variants has emerged, Gemini 3.8 Flash's output price came in higher than expected, and four new model IDs were recorded for the first time.

DeepSeek V4: The Low-Price Strategy Continues

Pricing for two DeepSeek^Ad V4 variants was recorded on October 12:

• deepseek-v4-1-flash: input $0.30/M, output $1.20/M
• deepseek-v4-pro: input $1.32/M, output $3.96/M

The Flash variant's pricing essentially continues DeepSeek's consistently aggressive approach. An input price of $0.30 means cost-sensitive tasks like classification, routing, and structured extraction have a new candidate. Of course, cheap doesn't mean it fits your use case — actual performance still needs to be benchmarked against real tasks.

The Pro variant's positioning is more interesting. Its $1.32 input price is lower than GPT-4o and roughly in the same range as Claude Haiku 3.5's input price, while the output-side comparison depends on the specific input/output token ratio. If V4 Pro's quality holds up under scrutiny, it has a chance to carve into the middle ground of "stronger than Flash, but doesn't require flagship-level budgets" — precisely the most fiercely contested price segment right now.

Gemini 3.8 Flash: Output Price Higher Than Expected

The tracker recorded gemini-3-8-flash pricing at $0.825/M input and $4.125/M output (verified October 18).

The output side is worth watching out for. Google's Flash series has long been known for being cheap — Gemini 2.0 Flash's output price was only a fraction of this figure. If you're running high-output workloads — summarization, long-form content creation, agent loops that generate large numbers of tokens — it's best to run the token math before treating this model as a drop-in replacement for your current one. The "Flash" in the name makes it easy to assume it's cheap, but the data doesn't support that assumption.

Four New Model IDs Surface in the Tracker

Four new records entered the tracking system this week:

deepseek/deepseek-v4.1-flash-1017 (first seen October 18) — a date-suffixed checkpoint or an updated version of the V4 Flash line. DeepSeek has a habit of releasing timestamped variants, so users who pin model versions in production should take note.

qwen/qwen3.7-plus:batch (first seen October 17) — a batch endpoint variant of Alibaba's Qwen 3.7 Plus. Batch interfaces typically trade slower response times for substantial price discounts; if you have asynchronous workloads, it's worth waiting for full pricing to come out and running the numbers.

google/gemini-4-argon and google/gemini-4-argon:batch (both first seen October 14) — the most noteworthy new faces this week. The Argon codename suggests it may be a specific tier or configuration within the Gemini 4 family rather than the flagship itself. The simultaneous appearance of both a standard and a batch version usually means Google is preparing for an official launch. There's currently no pricing data in the system, so its cost positioning can't be determined.

Batch endpoints are a billing model that cloud providers offer for asynchronous tasks: users package a large number of requests into a single job and submit it, and the system processes it over hours or even longer. In exchange, the unit price is typically 30%–50% lower than real-time interfaces. It suits scenarios that don't require immediate responses, such as offline data cleaning, bulk content generation, and evaluation runs. Note that batch interfaces often have limits on job size, completion deadlines, and result retention periods, and once submitted they usually can't be canceled — so they suit teams that already have a task queue system, not online businesses that need immediate responses. Qwen 3.7 Plus's :batch variant is exactly this type of interface; before pricing is announced, the discount magnitude should be the key criterion for deciding whether to migrate.

Date-suffixed model IDs (like -1017) reflect the model version pinning mechanism. Providers continually update the weights on the same model line, causing the output behavior of identically named models to drift over time. If a production system depends on specific output formats or stability, it should explicitly specify date-suffixed snapshot versions, and after a new version is released, compare it against a regression test set before deciding whether to upgrade.

Practical Recommendations

• High concurrency, cost-sensitive: DeepSeek V4 Flash's $0.30/$1.20 is worth including in your benchmarks, but be sure to validate quality against real tasks first.
• Mid-tier reasoning tasks: V4 Pro enters a crowded battlefield, but its pricing is competitive enough to put on your candidate list.
• Gemini users: 3.8 Flash's output price is higher than the "Flash" name suggests — run a token cost estimate before migrating.
• Keep an eye on Gemini 4 Argon: no pricing yet, but the simultaneous appearance of two variants is a signal.

Price data changes quickly; the above reflects what the tracker detected this week, and some figures may have changed by the time you read this.