Naderu Weekly · Episode 14
Priced is not available
Google published a price this week for a model you cannot call. Gemini 4 Carbon has a rate and no API, Kimi K3.1 Pro has an API and no weights yet, Kimi's K2 line closes to new accounts on 31 October, and DeepSeek starts billing searches your tools already make. Four launches, four gates — and each gate has a date.
Watch on YouTube — 5:45. Every figure below is linked to where we got it. If you would rather have it arrive than remember to check, subscribe to the weekly report — one email a week, one click to leave.
A correction to ourselves first. Last week we promised to measure token efficiency in open weights this week. We have not run that evaluation, and we will not publish numbers we did not measure, so it moves to a later report.
Every gate this week has a date on it
| Date | What changes |
|---|---|
| 26 Oct | GPT-6.1 Terra released at $1 / $5 — callable today |
| 28 Oct | Gemini 4 Carbon announced — cyber defenders first |
| 31 Oct | Kimi K2 models closed to newly registered accounts |
| 1 Nov | DeepSeek Search and Fetch start billing, 00:00 Beijing time |
| by 4 Nov | Kimi K3.1 Pro weights promised |
| 1 Dec | Gemini 4 Argon's introductory rate ends — $4 / $20 |
None of these dates is a model getting better. Each one changes who can use a model, or what it costs.
Carbon: a price for a model you cannot call
Google's own post says Carbon launches at Argon's introductory price — $2 in / $10 out per million tokens, cached input 95% off — then $4 / $20 when the introductory period ends, with no end date given. Access starts with a subset of cyber defenders and phases to trusted partners through the Fairwind Program; developers, businesses and consumers get it "as soon as possible". Every benchmark in the post is Google's own reporting.
So Carbon is not a row on the board. A row needs a model id, a context window and an API page, and Carbon has none of the three yet. It is in the board's coverage ledger and lands the day the Gemini API pricing page lists it.
One run, six prices
Ten million input tokens and two million out, at list rates, uncached, with every request under any long-context step — priced off the board data file:
| Same workload | Job cost |
|---|---|
| Kimi K3.1 Pro Highspeed ($4 / $20) | $80.00 |
| Gemini 4 Argon from 1 December ($4 / $20) | $80.00 |
| Kimi K3 ($3 / $15) | $60.00 |
| Gemini 4 Argon until 30 November ($2 / $10) | $40.00 |
| Kimi K3.1 Pro ($2 / $10) | $40.00 |
| GPT-6.1 Terra ($1 / $5) | $20.00 |
Kimi: a cheaper flagship, and a door closing behind it
Moonshot's pricing page lists Kimi K3.1 Pro at $2.00 in / $0.20 cached / $10.00 out, flat across a 1,048,576-token window, with cache writes at $2.00 for five minutes or $4.00 for an hour. The Highspeed variant is $4.00 / $0.20 / $20.00 for two to three times the speed. The K3.1 Pro guide says the full weights will be released by 4 November. Until they are on Hugging Face under a named licence, the board carries K3.1 Pro as closed, and Kimi K3 stays the open-weight pick on the cards that chose it for its weights. Moonshot claims coding and agentic gains; we have not evaluated them.
The same page carries the bigger change: from 31 October, K2 series models are no longer available to newly registered users. Existing accounts are unaffected, and Moonshot plans to retire every K2 model by the end of December 2026. So Kimi K3.1 Pro replaces Kimi K2.7 Code on the board's agentic-coding card — Moonshot's own K2 notice points migrating users to the K3.1 Pro guide. Per token it costs about twice as much, and until its weights ship the card has no open-weight pick.
1,000 web searches, by provider
| Search tool, fee on top of tokens | Per 1,000 |
|---|---|
| Gemini 3 and newer, Google Search grounding — after 5,000 free a month | $14.00 |
| OpenAI web search | $10.00 |
| Claude web search | $10.00 |
| Grok web_search | $5.00 |
| DeepSeek Search, from 1 November — Fetch is another $0.75 | $0.75 |
DeepSeek is the cheapest by far, and the fee is not the problem. DeepSeek's guide says that when you use its API through DeepSeek Harness, Claude Code or Codex, server-side tool calls already hit the Search and Fetch APIs, free until 00:00 Beijing time on 1 November 2026 — 16:00 UTC on 31 October. After that, every successful call bills: an HTTP 200 with a non-empty result for Search, page content for Fetch. Failures, empty results and timeouts do not. A coding agent can start running up a search bill for calls nobody on your team wrote. The rates sit beside the token prices on DeepSeek's pricing page.
Three moves for Monday
1. Meter the calls you did not write
If an agent runs on DeepSeek through a harness, count its server-side Search and Fetch calls this week, while they are free. From 1 November each successful one bills.
2. Register before the door shuts
If you might need Kimi K2 next quarter, open the account before 31 October — then plan the move to K3.1 Pro anyway, because the K2 line retires by the end of December.
3. Budget on the price you can pay
Plan on what you can call today. Carbon and K3.1 Pro's weights are dated options, not inputs.
Also on the board this week
Grok 4.7, 4.6 and 4.5 now carry xAI's 20% Batch discount; the board had said Batch was not supported on 4.7 and 4.6, and that is corrected. The web-search card had priced Grok 4.7 at $2.00 / $10.00 — its output is $6.00, and that is corrected too. Five rows left the board because their endpoints no longer exist: Gemini 3.7 Flash and 3.5 Flash (routed to 3.8 Flash and 3.6 Flash), Devstral 2, Sonar Pro and Sonar Reasoning Pro. Llama 3.3 70B lost its price — no host we track publishes one any more. And from 26 October, new paid Gemini API projects store requests by default. Every change is in the board's changelog.
Sources
- Naderu model board and its data file — 74 rows, re-verified 29 October 2026.