DeepSeek V4.1 Flash repriced twice in five days as GPT-6.1 Terra appears on OpenRouter

A weekly price-tracking post on dev.to documented a volatile stretch for LLM API costs: DeepSeek's V4.1 Flash tier was repriced twice within five days, while four previously unseen OpenAI model IDs carrying a "Terra" label appeared on OpenRouter. The post, published on 1 November, draws on a tracker that checks prices daily across OpenAI, Anthropic, Google, DeepSeek, and xAI.

DeepSeek V4.1 Flash moved twice in five days

According to the tracker, DeepSeek V4.1 Flash was listed on 26 October at $0.02 per million input tokens and $1.00 per million output tokens. By 31 October the same model was listed at $0.045 input and $0.50 output. Input pricing more than doubled; output pricing was cut in half.

Which side of that trade a given team lands on depends entirely on workload shape, the post notes. Input-heavy pipelines — long-context retrieval, document processing, large system prompts — ended the week more expensive. Workloads that generate a lot of tokens relative to what they consume came out ahead.

The tracker describes two changes in five days on the same model as unusual, observing that most providers adjust prices slowly and with advance notice. This one reportedly shifted mid-week without any announcement.

Kimi K3.1 Pro surfaced at premium rates

On 27 October the tracker logged its first pricing record for Kimi K3.1 Pro: $4.00 per million input tokens and $20.00 per million output tokens. That places it in the same general band as GPT-4-class models — a premium position rather than a budget one. The author adds a caveat: with no earlier snapshot of the model, there is no way to tell whether this represents a price change or simply the first time the tracker caught it.

Four GPT-6.1 Terra models appeared on OpenRouter

The same day, four new OpenAI model IDs showed up in the tracker's OpenRouter data:

• openai/gpt-6.1-terra-pro
• openai/gpt-6.1-terra-pro:batch
• openai/gpt-6.1-terra
• openai/gpt-6.1-terra:batch

The ":batch" suffixes follow OpenAI's established pattern for asynchronous, discounted batch processing. The "Terra" label is the novelty — the post says it is not a suffix OpenAI has used on any model the tracker monitors, and it explicitly declines to guess whether it marks a capability tier, a regional variant, or something else entirely. No pricing came through with the listings, so there are no numbers to evaluate yet.

Why it matters

Anyone running DeepSeek Flash in production should redo their cost arithmetic with the 31 October figures. Input-heavy workloads now pay roughly 2.25x more per input token, while output-heavy ones effectively received a discount.

The episode is also a reminder that static pricing references decay quickly. Budgets, contracts, and long-running batch jobs priced off a spreadsheet from last month can drift from reality without anyone noticing — a live pricing source beats a cached one, particularly before committing to a contract or a lengthy batch job.

For teams evaluating GPT-6.1 options, the Terra batch variants are the ones to watch once pricing appears: batch versions traditionally come at a significant discount to synchronous endpoints, which matters for any workload that can tolerate async turnaround.

One caveat: all of this comes from a single tracker's observations, and the source itself flags open questions — Kimi's pricing history and the meaning of "Terra" are both unresolved. Treat the Terra models as unconfirmed until OpenAI publishes details.