# Introductory AI model prices: how to budget for the step-up

URL: https://infro.io/blog/ai-model-launch-pricing-budgets
Author: INFRO Editorial (https://infro.io/blog/editorial)
Publisher: INFRO
Published: 2026-10-27
Last modified: 2026-10-27
Last reviewed: 2026-10-27 (next review due 2026-11-26)
Topic: Cost and FinOps
Format: Guide

> Budget a model with an introductory or time-limited rate on the rate that will apply when the work runs, not the launch rate. Keep a price schedule with effective dates per model, forecast month by month from it, and decide before the step-up whether to stay, move work to batch or caching, or switch. As of 27 October 2026, Gemini 4 Argon moves from $2 input and $10 output per million tokens to $4 and $20 on 1 December, and Gemini 3.8 and 3.6 Flash from $0.75 and $3.75 to $1.50 and $7.50 on 1 January 2027.

Key takeaways:
- Gemini 4 Argon is paid-tier only at $2 input and $10 output per million tokens through 30 November 2026, then $4 and $20; output includes thinking tokens.
- Gemini 3.8 and 3.6 Flash move from $0.75 and $3.75 to $1.50 and $7.50 on 1 January 2027, and context caching and cache storage double the same day.
- Not every time-limited price has an end date: OpenAI says GPT-5.6 Sol's promotional pricing is available at least through 21 November 2026, which is a floor, not an end.
- Scheduled changes move both ways: Anthropic's planned rise for Claude Sonnet 5 on 1 September 2026 did not happen, and the introductory rate became standard.
- Forecast month by month from a price schedule with effective dates; a single rate per model hides the step.
- Ask before you commit whether the introductory rate applies on cloud platforms and aggregators, and whether a negotiated rate covers a model launched after the contract.

Budget an AI model with an introductory or time-limited price on the rate that will apply when the work actually runs. As of 27 October 2026, Google lists Gemini 4 Argon at $2 input and $10 output per million tokens through 30 November, then $4 and $20 from 1 December[1], and Gemini 3.8 and 3.6 Flash at $0.75 and $3.75 through 31 December, then $1.50 and $7.50 from 1 January 2027[1]. A plan built on the launch rate under-forecasts every month after the step. The fix is mechanical: a price schedule with effective dates, a month-by-month forecast built from it, and a decision made before the step rather than after the first invoice.

This guide is for both the engineers who pick models and the people who sign off the spend. It covers the time-limited prices on the table now, the shapes they come in, a forecasting method with code, the contract questions worth asking and the alerts to set. The yearly planning method around it is in [AI budget planning for 2027](https://infro.io/blog/ai-budget-planning-2027).

## The time-limited prices on the table now

| Vendor and model | What the vendor says | Rate now | Rate after |
| --- | --- | --- | --- |
| Google, Gemini 4 Argon | Paid tier only; prices "through November 30, 2026" and "starting December 1, 2026" | $2 in, $10 out | $4 in, $20 out |
| Google, Gemini 3.8 Flash and 3.6 Flash | "through December 31, 2026" and "starting January 1, 2027" | $0.75 in, $3.75 out | $1.50 in, $7.50 out |
| OpenAI, GPT-5.6 Sol | Promotional pricing "available at least through November 21, 2026" | Promotional | Not dated |
| Mistral, Mistral Large 4 | Launch pricing "50% off for 2 weeks" from the 6 October preview | $1.36 in, $4.18 out listed today | — |
| DeepSeek, server-side Search and Fetch | Free until 31 October, 16:00 UTC | Free | $0.75 per 1,000 calls per endpoint |
| Z.ai, GLM-5.3 family | Cached input storage "Limited-time Free" | Free | Not dated |

Table: Time-limited AI API prices, standard tier, per million tokens unless stated, from each vendor's page read on 2026-10-27.

The sources: Google's pricing page for both Gemini rows, including the paid-tier restriction on Argon and the note that its output price includes thinking tokens[1]; OpenAI's pricing page for GPT-5.6 Sol[2]; Mistral's changelog entry of 6 October for the launch pricing[3] and its model page, which today lists $1.36 input, $0.14 cached input and $4.18 output with no launch price shown[4]; DeepSeek's V4.1-Flash-1017 notice for the tool calls[5]; and Z.ai's pricing page, which gives no end date for its free cache storage[6].

The step-ups reach beyond the headline rates. On Argon, context caching moves from $0.10 to $0.20 and the Batch and Flex rates from $1 and $5 to $2 and $10[1]. On the Flash models, context caching moves from $0.075 to $0.15 and cache storage from $0.50 to $1.00 per million tokens per hour on 1 January[1]. Argon became generally available on 13 October[7], so any team that adopted it this month adopted it at the first rate.

## Five shapes of a time-limited price

Read the wording, because it decides what you can plan:

1. **Fixed end, next rate published.** Gemini 4 Argon and the Gemini Flash models state both dates and both rates[1]. This is the easy case: the step goes straight into the schedule.
2. **A floor, not an end.** "Available at least through November 21, 2026" for GPT-5.6 Sol[2] says when the promotion cannot end before, not when it ends. Plan as if it ends on that date, and treat anything after as upside.
3. **A duration counted from launch.** "50% off for 2 weeks" for Mistral Large 4[3] puts the end around 20 October without a date, and the regular rate lives on another page[4]. Work out the date yourself and confirm it.
4. **Free until.** DeepSeek's tool calls cost nothing until a stated hour, then $0.75 per 1,000 calls per endpoint[5]; Z.ai's cache storage is free for a "limited time" with no date[6]. A free line has no history in your bills, so it is the easiest to miss.
5. **A scheduled change that does not happen.** Anthropic launched Claude Sonnet 5 at $2 and $10 as introductory pricing through 31 August 2026 with a rise to $3 and $15 planned for 1 September; the rise was cancelled and the introductory rate became standard[8]. Keep the step in the forecast until the pricing page changes, and expect to remove it sometimes.

Rates also differ by where you buy. Perplexity's changelog lists Gemini 4 Argon on its Agent API "with introductory pricing through November 30, 2026"[9]; a cloud platform or aggregator decides its own dates and rates, so the price list of the platform that bills you is the one that counts.

## Build the step into the forecast

A forecast that holds one rate per model cannot show a step. Keep a price schedule instead, one row per model, rate and period, with the page and date you read it from:

```python
from datetime import date

# One row per rate period. Rates in USD per 1M tokens. Source and read date kept with the row.
SCHEDULE = [
    {"model": "gemini-4-argon", "from": date(2026, 10, 13), "to": date(2026, 11, 30),
     "input": 2.00, "output": 10.00, "source": "ai.google.dev/gemini-api/docs/pricing", "read": "2026-10-27"},
    {"model": "gemini-4-argon", "from": date(2026, 12, 1), "to": None,
     "input": 4.00, "output": 20.00, "source": "ai.google.dev/gemini-api/docs/pricing", "read": "2026-10-27"},
    {"model": "gemini-3.8-flash", "from": date(2026, 9, 2), "to": date(2026, 12, 31),
     "input": 0.75, "output": 3.75, "source": "ai.google.dev/gemini-api/docs/pricing", "read": "2026-10-27"},
    {"model": "gemini-3.8-flash", "from": date(2027, 1, 1), "to": None,
     "input": 1.50, "output": 7.50, "source": "ai.google.dev/gemini-api/docs/pricing", "read": "2026-10-27"},
]

def rate_on(model: str, day: date) -> dict:
    for row in SCHEDULE:
        if row["model"] == model and row["from"] <= day and (row["to"] is None or day <= row["to"]):
            return row
    raise LookupError(f"no rate for {model} on {day}: treat as unpriced, not as zero")

def month_cost(model: str, first_day: date, input_mtok: float, output_mtok: float) -> float:
    r = rate_on(model, first_day)
    return input_mtok * r["input"] + output_mtok * r["output"]
```

The function prices a month at the rate in force on its first day, which is right for steps that fall on the first of a month, as both Google steps do. For a step mid-month, split the month. Two rules matter more than the code:

- **Count thinking tokens as output.** Google's Argon and Flash output prices include thinking tokens[1]. A forecast built from visible answer length under-counts output on any reasoning model.
- **A missing rate is unpriced, not zero.** A model with no row should stop the forecast, not add nothing to it.

> **Example: What the Argon step does to one workload**
> Suppose a coding-assistant feature sends 400 million input tokens and receives 80 million output tokens (thinking included) a month on Gemini 4 Argon. At the rate through 30 November: 400 × $2 + 80 × $10 = $1,600 for November. From 1 December: 400 × $4 + 80 × $20 = $3,200 a month. A November-to-January forecast at the launch rate reads $4,800; with the step it is $1,600 + $3,200 + $3,200 = $8,000. The same volumes on Gemini 3.8 Flash cost 400 × $0.75 + 80 × $3.75 = $600 in December and $1,200 from January. Volumes are hypothetical, at Google's standard-tier list rates read on 2026-10-27; Batch and Flex halve the Argon figures.

## Decide before the step, not after the invoice

The month before a step is the cheapest time to change course. Four options, with the cases where each loses:

| Option | When it works | When it loses |
| --- | --- | --- |
| Stay and budget the new rate | The model is clearly better on your evaluation at the post-step rate | When nobody re-ran the evaluation at the new rate |
| Move eligible work to batch or flex | Work that can wait; Argon's Batch and Flex rates after 1 December equal its standard rates before it | Interactive traffic; anything with a latency target |
| Cache more | Long shared prefixes; note Argon supports implicit caching only, not explicit | Prompts that differ from the first token |
| Switch model | A model at a lower list rate passes your evaluation | When the cost of re-qualifying prompts, tools and tests exceeds the difference |

Table: Choices before a scheduled price rise. Argon batch and caching terms from Google's pricing page, read 2026-10-27.

Whatever you choose, compare options on cost per completed task, not on the per-token rate: a model at half the rate that needs twice the retries or twice the reasoning tokens costs the same.

## Contract and procurement questions

For a finance or procurement lead, the questions to put to the vendor or account team before committing:

1. **What is the rate after the introductory period, in writing, and from what date?** A pricing page can change; a contract term cannot without notice.
2. **Does our negotiated rate cover a model launched after we signed?** If it does not, a new model runs at list until someone amends the contract.
3. **Does the introductory rate apply where we actually buy?** On Vertex AI, Bedrock, a cloud listing or an aggregator, the rate and its dates are set separately.
4. **How much notice do we get of a price change?** Compare the notice period with your own budget cycle.
5. **Does usage on an introductory rate count toward our commitment at the introductory or the standard rate?** It changes how fast a commitment burns down.
6. **What happens to batch, caching and priority rates at the step?** On Argon all of them double together[1]; that is not true of every vendor.

## Alerts and controls that catch the step

- **A calendar entry per step**, 30, 14 and 7 days ahead, owned by the person who owns the model choice.
- **Budget thresholds sized on the post-step rate**, so the first alert after the step is expected rather than a surprise.
- **A check of the vendor's pricing page in the week before the step.** Steps get cancelled and extended, as Sonnet 5's was[8].
- **A reconciliation of the first invoice after the step** against your estimate, per model, so a rate you mis-recorded is found in one month rather than three.

For a [negotiated rate](https://infro.io/blog/negotiated-rates-list-price-drift), the same step can apply to your contract or not, depending on how the contract is written. That is a question to answer before the step, not from the invoice.

## Forecasting and enforcing a price step with INFRO

INFRO is a control plane in front of the provider accounts you already hold; each provider bills you directly, and INFRO's charge is a separate subscription. Several parts of it apply to a price step:

- **Budgets and forecast alerts.** Budgets set a cap per organization, project, member and key with alerts at your thresholds, and a budget forecast to be exceeded becomes a finding routed to email, Slack and webhooks. The forecast is drawn from your own spend history; it has no way to know a vendor's future rate, so size the budget on the post-step rate yourself.
- **Negotiated rates.** You can enter the rate your contract carries for a catalog model, optionally per provider, and estimates use it before the list rate.
- **Reconciliation.** For OpenAI and Anthropic, a daily comparison of INFRO's estimate with the cost reported by the provider's own cost API flags a persistent gap; when a vendor changes a published list rate, INFRO records the change with its source and effective date.
- **Commitments.** A committed-spend agreement can be recorded and tracked against reported cost, with a forecast to the end of the term.

What INFRO does not do: it estimates each request at the rate in force in its catalog when the request is made, so it never prices this month's traffic at next quarter's rate; the forward schedule above stays your job. Gemini 4 Argon is not in INFRO's model catalog as of 27 October, so INFRO has no rate to estimate it with yet, and under a blocking budget a request that could reach an unpriced model is refused unless you allow it. Reconciliation reads OpenAI's and Anthropic's cost APIs only, so a Google rate step is not checked against Google's reported cost.

## Questions

### How much does Gemini 4 Argon cost after 1 December 2026?

Google's pricing page lists $4 per million input tokens and $20 per million output tokens, including thinking tokens, from 1 December 2026, up from $2 and $10 through 30 November. Context caching moves from $0.10 to $0.20 and Batch and Flex from $1 and $5 to $2 and $10. The model is not available on the free tier.

### When does Gemini 3.8 Flash get more expensive?

On 1 January 2027. Google lists $0.75 input and $3.75 output per million tokens through 31 December 2026 and $1.50 and $7.50 from 1 January 2027 for both Gemini 3.8 Flash and 3.6 Flash. Context caching moves from $0.075 to $0.15 and cache storage from $0.50 to $1.00 per million tokens per hour.

### Should I commit spend to a model that is on an introductory rate?

Only on the post-introductory rate. Size any commitment from the rate that will apply for most of the term, get that rate and the date it starts in writing, and ask whether your negotiated terms apply to models launched after the contract was signed.

### Does a launch price on the vendor's API apply on Vertex AI, Bedrock or an aggregator?

Not automatically. Cloud platforms and aggregators publish their own price lists. Perplexity's changelog, for example, lists Gemini 4 Argon on its Agent API with introductory pricing through 30 November 2026. Check the price list of the platform that actually bills you.

### What if the vendor cancels a scheduled price rise?

It happens. Anthropic announced Claude Sonnet 5 at introductory pricing through 31 August 2026 with a rise planned for 1 September, then kept the introductory rate as the standard price. Keep the step in the forecast until the vendor's pricing page changes, and record the source and date you read it.

## Sources

1. Gemini Developer API pricing — Google. https://ai.google.dev/gemini-api/docs/pricing (accessed 2026-10-27)
2. Pricing — OpenAI. https://developers.openai.com/api/docs/pricing (accessed 2026-10-27)
3. Changelog — Mistral AI. https://docs.mistral.ai/resources/changelogs (accessed 2026-10-27)
4. Mistral Large 4 — Mistral AI. https://docs.mistral.ai/models/mistral-large-4-0 (accessed 2026-10-27)
5. DeepSeek-V4.1-Flash-1017 — DeepSeek. https://api-docs.deepseek.com/news/news261017/ (accessed 2026-10-27)
6. Pricing — Z.ai. https://docs.z.ai/guides/overview/pricing (accessed 2026-10-27)
7. Gemini API release notes — Google. https://ai.google.dev/gemini-api/docs/changelog (accessed 2026-10-27)
8. Pricing — Anthropic. https://platform.claude.com/docs/en/about-claude/pricing (accessed 2026-10-27)
9. Changelog — Perplexity. https://docs.perplexity.ai/changelog (accessed 2026-10-27)