Naderu Weekly · Episode 13
The contributor discount is real — if your repo already is
A frontier coding run that costs forty dollars on Google's newest model costs one dollar and forty cents on Meta's. The models are comparable. The difference is what you agree to share. This week: Meta's ten-cent contributor tier, Google's doubling Argon clock, and the public-repo playbook for agentic work.
The episode cut publishes Thursday — the link lands here on upload. 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.
Every cheap price this week has a date on it
Before the models, the calendar. Google's pricing page puts Gemini 4 Argon at $2 in / $10 out per million tokens through 30 November 2026, then $4 / $20 from 1 December. Gemini 3.8 and 3.6 Flash double on New Year's Day. Anchor budgets on the post-cliff rate and treat everything before it as a discount — a team that priced Argon at two dollars permanently gets a one hundred percent rise delivered as a calendar event.
Argon: Google's flagship with an expiry date
Per the Argon model page, the budget is 1,048,576 tokens shared between input and output, with 65,536 default max output rising to the full million under long decode. Function calling, computer use, code execution and search grounding are supported. Paid tier only, no free tier, implicit caching exclusively.
The benchmarks need a warning label, and this one is second-hand: early coding reviews are mixed, with subscriber default mode reportedly benching below the high mode behind Google's 77.9% DeepSWE figure. Google's numbers, Google's tests — evaluate Argon on your own repo before migrating a lane. Grounding bills apart: 5,000 free searches a month, then $14 per 1,000.
One run, five prices
Ten million input tokens and two million out — a serious overnight agent run — priced off the board data file:
| Same workload, different counterparty | Job cost |
|---|---|
| Argon introductory ($2 / $10) | $40.00 |
| Spark standard ($1.25 / $4.25) | $21.00 |
| DeepSeek V4.1 off-peak ($0.15 / $0.60) | $2.70 |
| GPT-6 Luna ($0.10 / $0.50) | $2.00 |
| Spark contributor ($0.10 / $0.20) | $1.40 |
Twelve times cheaper is a contract, not a coupon
Meta's pricing page runs identical weights at $0.10 in, $0.002 cached, $0.20 out in contributor mode — in exchange for permission to use the prompts, code, files and context you send, plus the responses you get back, to train future Meta models. Limits tighten to 150 RPM against 3,000 standard. The guarantee is the model id: call muse-spark-1.3-contributor, and note Muse Code subscriptions already default to contributor mode.
The same page previews a second closed model, Muse Wonder 1.0 at $4.00 / $0.20 / $20.00 with no contributor tier — see the models overview. And the open-weight Spark release (816B parameters, 42B active) is already served by Together at $1.25: same rate as closed Spark, so the choice is custody and licence, not price.
Elsewhere on the board: MiniMax M3.1 Flash went GA at $0.15 / $0.60 — half of M3, with weights now on Hugging Face — Mistral Large 4 previews at $1.36 / $4.18, and DeepSeek's flash endpoint now serves the 1017 checkpoint at unchanged rates.
Three moves for Monday
1. Sort traffic by custody
Send already-public work — open-source agents, docs bots, eval harnesses — to oss/ contributor lanes, where every member trains and choosing the lane is the consent. Cruise is ours, so we say so: it is an AI API gateway sold by BytesBrains, the company Naderu is a venture of. See the Cruise API reference (ours) and Cruise pricing (ours) (provider list plus 3%).
2. Pin the contributor id
Pin muse-spark-1.3-contributor, not the family name, and meter the 150 RPM ceiling before launch — contributor throughput is throttled far harder than standard.
3. Keep private traffic standard
Customer and regulated code stays on standard tiers, where retention terms hold and nothing trains on your data. No discount survives a compliance incident.
Sources
- Naderu model board and its data file — 77 rows, re-verified 22 October 2026.