Naderu by BytesBrains All reports  ·  Model board →

Naderu Weekly · Episode 15

Downloadable is not unrestricted

5 November 2026 Week of 2 November 5 min episode · 6 min read

Kimi K3.1 Pro's weights shipped on the deadline and Mistral Large 4's halved its API price. Both downloads clear a terabyte, and both carry a hosted-competition clause in the licence beside them — and the clauses do not match. Meanwhile DeepSeek's search meter is running, Kimi's K2 line is closed, and the Decisions API is generally available with caching it did not have in beta.

Watch or listen

Watch on YouTube — 5:13. 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. OpenAI's Decisions API went generally available on 22 October, with prompt caching the beta did not have. It happened between our passes and last week's report never said so — corrected here and on the board.

Every gate this week has a date on it

DateWhat changes
29 OctMistral Large 4 weights out, API halved to $0.70 / $2.10
31 OctKimi K2 models closed to newly registered accounts
1 NovDeepSeek Search and Fetch start billing, 00:00 Beijing time
4 NovKimi K3.1 Pro weights, on the deadline
21 NovGPT-5.6 Sol promo rate ends
1 DecGemini 4 Argon's introductory rate ends — $4 / $20

None of these dates is a model getting better. Two are prices going up, one is a door closing, and two are downloads getting bigger.

Kimi kept the date — read the file beside it

moonshotai/Kimi-K3.1-Pro landed 4 November: ninety-six shards, 1.56 terabytes — we counted the file listing. It is the day Moonshot's own guide promised, so last week's closing question is answered and this week's board row is open for the first time. The API is unchanged at $2.00 in / $0.20 cached / $10.00 out, flat across the million-token window.

Beside it, the Kimi K3.1-Pro License. Commercial use is granted — until you run inference for third parties above $20M aggregate revenue in any twelve months. Then you need a separate agreement with Moonshot before any commercial use. Above 100M monthly users or $20M monthly revenue, the name goes on the product. Internal use and Moonshot's official products and certified partners are exempt. It is the same line K3's licence drew.

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 workloadJob cost
Kimi K3.1 Pro Highspeed ($4 / $20)$80.00
Gemini 4 Argon from 1 December ($4 / $20)$80.00
Kimi K3.1 Pro ($2 / $10)$40.00
Gemini 4 Argon until 30 November ($2 / $10)$40.00
GPT-6.1 Terra ($1 / $5)$20.00
Mistral Large 4 ($0.70 / $2.10)$11.20
At seventy cents in and two-ten out, Mistral's open-weight flagship is the cheapest run on this chart — and it is also downloadable. The API price and the download are two different products with the same name, and this chart prices the first.

The cached footnote: with a full cache hit, that run is $4.90 on Mistral Large 4 — seventy cents of input plus full-price output — against $22.00 on K3.1 Pro. Same ninety percent off the input at both labs; the gap is output, where Mistral charges a fifth of Kimi. OpenAI's pricing page and Google's carry the Terra and Argon rates this table uses.

Three licences, one question

Line the three licences up by what they ask of you. DeepSeek's V4.1 Flash is MIT — nothing to sign, self-host or fine-tune freely. Kimi draws its line at $20M in hosting revenue, with attribution above that and partners carved out. Mistral draws no line at all: any model-as-a-service business needs a commercial licence first, however small; only internal use and Mistral's official products are exempt.

The cheaper the API, the more the licence matters — because the moment the API is cheap enough to build on, you are the host the clause was written for. Read it before you pull eleven hundred gigabytes, not after. The weights sit at mistralai/Mistral-Large-4-1T-A52B, about 1.06 terabytes on the main branch.

DeepSeek's search meter is running

DeepSeek's guide bills Search and Fetch at $0.75 per 1,000 successful requests each since 00:00 Beijing time on 1 November 2026 — confirmed live this pass. When you use its API through DeepSeek Harness, Claude Code or Codex, server-side tool calls hit the same APIs. Only an HTTP 200 with a result bills — a non-empty results array for Search, page content for Fetch. Failures, empty results and timeouts do not. The meter that was a warning last week is a bill this week.

And the Decisions API is generally available since 22 October, per OpenAI's changelog. Shared input now caches at the standard gpt-6-luna rates ($0.01 read, $0.125 write, 1,024-token minimum), questions carry images — inline or hosted URLs — and each request holds up to 400 questions. The Decisions guide carries the details.

Three moves for Monday

Licence

1. Find the hosting clause

If you serve model output outside your company, read the model-as-a-service clause before the weekend. Kimi gives you to $20M in hosting revenue; Mistral gives you nothing.

Revision

2. Pin the revision

Weights and prices moved in the same week. Benchmarks attach to a revision, not a name — pin the weights and the API rate before you deploy.

Budget

3. Budget the dated prices

Two cheap rates expire this month: the GPT-5.6 Sol promo ends 21 November and Argon doubles 1 December. Both doublings are already dated, so plan them now.

Also on the board this week

All 74 rows re-fetched against first-party pages and re-dated 5 November, except Mistral Medium 3.5 and Large 3, whose prices render client-side only and are carried under named deferred entries. The Decisions row moves from beta to GA with cache rates ($0.01 read, $0.125 write). Kimi K3.1 Pro and K3 join the Batch API at 60% of standard, so all four Kimi rows carry a 0.4 batch discount. Hugging Face's provider note is corrected — free accounts no longer get included credits. Tencent Hy4 (released 4 November, API-only) and OpenAI's GPT-6-Cyber (announced 5 November, no pricing row yet) are recorded in the deferred ledger, not added. Carbon re-checked: still no API page. Every change is in the board's changelog.

Sources