Perplexity, Google, Anthropic and Mistral reset token prices

Perplexity lowered the listed rates for perplexity/low on 2026-09-16: input pricing moved from $3 to $0.2 per million tokens, and output pricing moved from $15 to $1.2 per million tokens.

That price move sits alongside lower listed rates from Google AI and Anthropic, as well as mixed changes from Mistral AI. The repricing record for the period from 2026-09-14 through 2026-09-21 covers models used for different purposes, but each entry separates the price of material sent to a model from the price of material it generates.

That distinction matters when reviewing an AI bill. Input pricing matters to anyone placing long documents, retrieved material or conversation history into a prompt. Output pricing matters to anyone generating substantial responses. A rate card alone does not establish a deployment's total spend, because the record does not state how many input or output tokens a particular workload uses.

Perplexity/low now carries lower input and output rates

The perplexity/low change is direct on each side of the request. Its input rate was $3 per million tokens and is now $0.2. Its output rate was $15 per million tokens and is now $1.2. The stated effective date is 2026-09-16.

For teams that put long source material into prompts, the input line is the relevant change: $0.2 replaces $3 for each million input tokens. For applications that generate extended answers or other text, the output line changes from $15 to $1.2 for each million output tokens. The two changes should be considered separately in any cost review.

The record gives price movement, not a usage forecast. A system that sends little prompt material and produces substantial output has a different cost mix from one that repeatedly submits large document collections for analysis. Neither pattern can be inferred from the listed model price alone.

Google AI lowered rates for its named preview models

Google AI changed prices for gemini-2.5-computer-use-preview-10-2025 effective 2026-09-14. Its input rate moved from $1.25 to $1 per million tokens. Its output rate moved from $10 to $5 per million tokens.

The same date applies to gemini-robotics-er-2-preview. Its input price moved from $2 to $1 per million tokens, while its output price moved from $10 to $5 per million tokens. The current listed rates therefore match across the input and output fields, even though the earlier input prices differ.

For a user of either named Google AI model, the output rate is now $5 per million tokens. The input rate is now $1 per million tokens. That is useful for budgeting prompts and completions independently rather than treating a model as having one blended token price.

The model identifiers also matter. gemini-2.5-computer-use-preview-10-2025 and gemini-robotics-er-2-preview have separate price histories in the record. A family name is not enough to establish which rate applies to a specific integration.

Anthropic lowered Claude Sonnet 5 while Mistral's changes split

Anthropic lowered the listed price of claude-sonnet-5 effective 2026-09-14. Input pricing moved from $3 to $2 per million tokens. Output pricing moved from $15 to $10 per million tokens.

Mistral AI's changes do not point in the same direction across its named models. ministral-3b-latest moved from $0.04 to $0.1 per million tokens for input, and from $0.04 to $0.1 for output. ministral-8b-latest moved from $0.1 to $0.15 for input and from $0.1 to $0.15 for output.

mistral-small-latest moved differently. Its input rate changed from $0.2 to $0.15 per million tokens on 2026-09-14, while its output rate remained $0.6 per million tokens. That makes this an input-side change rather than a full input-and-output reduction.

The Mistral entries are a reminder not to generalise from a vendor-level label. A lower rate on mistral-small-latest input does not describe the listed movement for ministral-3b-latest or ministral-8b-latest. The exact model identifier and the input or output field determine the relevant price.

Input and output prices affect different parts of a request

A prompt can contain instructions, user text, retrieved documents and prior conversation material. Those are input tokens for billing purposes. A generated answer is output, so a lower input rate helps document-heavy prompting while a lower output rate helps work that produces more generated material.

The price records show several ways those fields can move. perplexity/low, gemini-2.5-computer-use-preview-10-2025, gemini-robotics-er-2-preview and claude-sonnet-5 have lower listed input and output rates. mistral-small-latest has a lower input rate while keeping its $0.6 output rate. ministral-3b-latest and ministral-8b-latest have higher listed rates on each field.

This is why a single headline price can conceal the operational question. A team reviewing a model change needs the relevant input price, the relevant output price and its own token mix. The provided rates do not establish model quality, latency, reliability or any other service characteristic.

Replacement models make pricing an operational migration issue

The price change for claude-sonnet-5 lands alongside a recorded model transition. claude-3-5-sonnet-latest was retired on 2026-09-14, with claude-sonnet-5 listed as its replacement. The replacement's listed rates changed on the same date, from $3 to $2 for input and from $15 to $10 for output per million tokens.

Mistral AI also recorded a retirement on 2026-09-14: pixtral-large-latest was retired with mistral-small-latest listed as its replacement. mistral-small-latest has an input rate of $0.15 after moving from $0.2, and an unchanged output rate of $0.6. These named transitions are the immediate watchpoints for teams maintaining model inventories: migration decisions should check the replacement identifier and each current token rate, rather than assuming a vendor family has one stable price.