Meta's New Model Priced Below DeepSeek, Which Announces Price Hike on the Same Day
0xBroomberg
Meta released its coding model Muse Spark 1.2 on August 6, pricing it below DeepSeek's comparable service via a data-for-discount scheme; the same day DeepSeek announced a significant price increase — the divergence signals the AI model race is shifting from price wars to an Agent-capability showdown.
How cheap is Meta's "Contributor" plan, exactly?
Meta offers two API tiers: Standard and "Contributor." The Contributor plan requires users to let Meta train on their data in exchange for steep discounts.
Contributor input: $0.10 per million tokens — 12.5× cheaper than Standard. Cached input: $0.002/M tokens — 75× cheaper. Output: $0.20/M tokens — over 21× cheaper.
This means → Contributor-tier input and output costs both undercut DeepSeek-V4-Flash (uncached input ¥1/M tokens, output ¥2/M tokens).
In plain terms = Meta isn't trying to profit from APIs — it wants your usage data. One commenter put it bluntly: "They don't care about margin; they just want the data."
Where does Muse Spark 1.2 rank in coding benchmarks?
On Terminal-Bench 2.1, DeepSWE 1.1, and other code benchmarks, Muse Spark 1.2 places second only to Claude Opus 5, ahead of GPT-5.6, Grok 4.5, and Gemini 3.6.
It tops the leaderboard on MCP Atlas — an Agent tool-use benchmark — and trails only Opus 5 on the complex-reasoning GDPVal-AA V2 test.
In a GPU-kernel iterative-optimization test spanning 1,000+ tool calls (up to 24 hours of runtime), it beat Gemini 3.6 Flash and GPT-5.6 Terra but fell short of GPT-5.6 Sol and Opus 5.
It scored 54 on the AI Analysis Index (AII), tying Meta with SpaceX AI for third place among U.S. labs.
What can the Muse Code agent actually do?
Muse Code is a terminal-based coding agent — an AI tool that autonomously handles programming tasks from the command line — capable of end-to-end software engineering in large codebases: planning changes, writing code, verifying results.
Its architecture pairs a main loop with asynchronous background agents that run continuously to cut latency.
This means → less human intervention on complex multi-step tasks. Zuckerberg says users can install it with "a single line of code."
Why is DeepSeek raising prices while its rival cuts them?
DeepSeek-V4-Flash launched its API on July 31 and quickly built momentum; benchmarks showed it far surpassing the earlier V4-Pro-Preview, scoring near GPT-5.6 Sol xhigh on Almanbench.
Platzi CEO (co-founder of Latin America's leading online-education platform) posted: "I genuinely cannot understand how DeepSeek-V4-Flash achieves this speed and quality."
The same day Meta cut prices, DeepSeek announced it "plans to raise API prices across the board in the near term, with a significant increase expected" — widely read as a deliberate exit from the pure price war toward quality-driven competition.
This reflects a strategic pivot: DeepSeek is aggressively recruiting Agent Harness talent and absorbing top open-source projects; its Agent product (possibly named DeepSeek Code) is in active development.
Where will this race be won or lost?
Whether Meta's data-for-discount play triggers a mass developer migration depends on two things: users' tolerance for data-privacy trade-offs, and whether Muse Spark 1.2's real-world engineering performance matches its benchmark scores.
Other leading Chinese labs are also loosening pricing: Zhipu's GLM models will see a notable price increase once current promotions expire.
This means → the decisive dimension of competition is shifting from price to Agent capability — whichever coding agent proves more reliable on real tasks wins the developer.
Content is for reference only, not financial advice.