On the evening of August 5 Eastern Time (early morning of August 6 Taiwan Time), Zuckerberg personally posted an announcement unveiling Meta’s latest model, Muse Spark 1.2. The pricing of this new model pushes the price war to a new high: under the Contributor plan, which agrees to data sharing, input costs just $0.10 per million tokens and output $0.20, making it quite a bit cheaper than DeepSeek Flash. The timing is spot on: that very same day, DeepSeek announced on its developer platform, “We plan to comprehensively raise API service pricing in the near future, with a substantial expected increase.Chinese AI vendors are raising prices, Meta is cutting prices, and the price war in the AI circle has kicked off. The low price comes with just one condition: agreeing to let Meta use your data to train models.
Zuckerberg posts announcement: models and tools launch simultaneously.
While announcing Muse Spark 1.2 on X, Zuckerberg also unveiled his own coding agent. Muse Codeas the showcase medium for this model. Muse Code is a terminal programming tool that can take on complete software engineering tasks for large repositories, including planning changes, writing code, and verifying results. The two are trained together to deliver optimal performance when used in combination.
Releasing Muse Code in beta today. It’s a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results. Powered by Muse Spark 1.2, a coding-focused model update. pic.twitter.com/xqavk41w6v
— Mark Zuckerberg (@finkd) August 5, 2026
According to Meta’s official description, Muse Spark 1.2 significantly expands training compute and training environment diversity for programming tasks on top of 1.1. The improvements focus on code generation, complex debugging, codebase understanding, and end-to-end development workflows, with long-horizon tasks such as large repository generation, large projects, and automated research as the primary training directions. The model supports a 1 million token context window, enough to fit an entire large project into a single session.
Introducing Muse Code (beta), a terminal coding agent built for long-horizon software engineering, powered by our new Muse Spark 1.2 model.
Muse Code plans, implements, and validates complex, multi-file changes across large repositories with persistent sub-agents that solve… pic.twitter.com/uEMb1XL9Y0
— AI at Meta (@AIatMeta) August 5, 2026
Meta launches Muse Spark 1.2 model with superior capabilities and prices slashed even lower than DeepSeek.
The model isn’t tied to Muse Code—just connect it to other agents and it’s ready to use.
The most noteworthy thing about Muse Spark 1.2 is that it isn’t locked into Meta’s own tool ecosystem. The official page lists three ways to use it: Muse Code, the Meta Model API, and OpenRouter. The Meta Model API provides an interface compatible with the OpenAI SDK, and OpenRouter already lists muse-spark-1.2. In other words, developers can integrate it into any agent framework they’re already using, without having to migrate to Muse Code.
This makes Muse Spark 1.2’s low price meaningful in a completely different way. In the past, price cuts by model vendors only had an effect when tied to their own tools. Muse Spark 1.2 takes an open API approach: Claude Code users can replace the underlying model with Muse Spark 1.2, and Codex users can do the same, so the discount directly shows up on the API bills of existing workflows. For developers, that is more meaningful than yet another new tool.
Performance: closing in on the Claude lineup, but still trails Opus 5.
Based on the benchmark data Zuckerberg released, Muse Spark 1.2’s programming capability lands in the top tier. It scored 82.9% on Terminal-Bench 2.1, second only to Opus 5 (86.7%) from the Claude family, and ahead of GPT 5.1 Codex (81.8%) and Grok 4.5 Build (81.6%). However, on DeepSWE 1.1, which is closer to real-world tasks, Muse Spark 1.2’s 59.3% trails Opus 5 (65.0%) and GPT 5.1 Codex (64.8%), still leaving a gap from the top tier. Compared with the previous generation, Muse Spark 1.1, scores have improved noticeably across the board—for example, DeepSWE rose from 53.0% to 59.3%.
Data from the AI evaluation organization Artificial Analysis also shows a similar conclusion: Muse Spark 1.2 scored 54.1 on the Intelligence Index, slightly ahead of DeepSeek Flash’s 49.9 and GPT-5.6 Luna’s 51.2, while Terminal-Bench 2.1 stands at 80.1%.
Super cheap: two modes, with a price difference of 12 to 21 times.
The most notable feature of Muse Spark 1.2 is its pricing. Meta split it into two different model IDs, with prices differing by more than an order of magnitude:
- muse-spark-1.2（Standard）Input $1.25 (approximately NT$41) / million tokens, output $4.25 (approximately NT$140), cached input $0.15, and your data will not be used for training.
- muse-spark-1.2-contributor（Contributor）: Input $0.10 (about NT$3.3), output $0.20 (about NT$6.6), cached input $0.002, provided you agree to Meta using your data to “improve products.”
When compared with other models, the position is clear at a glance (all prices below are in USD per million tokens):
- Muse Spark 1.2 Contributor$0.10 input / $0.20 output
- DeepSeek Flash$0.14 input / $0.28 output (cached input $0.0028)
- GPT-5.6 Luna$0.20 input / $1.20 output
- Muse Spark 1.2 Standard$1.25 input / $4.25 output
The Contributor plan, which agrees to data sharing, is priced a notch below DeepSeek Flash, with cache hit input at only $0.002—almost free. In contrast, the Standard plan, which doesn’t share data, is priced around Gemini 3.6 Flash.Vals AI It also lists the cost for Spark 1.2 to complete tasks at only $0.69, which is 3 times cheaper than Kimi and over 10 times cheaper than Fable/Opus/5.6 Sol. The price-performance ratio is indeed very high (the table below shows the cost without data sharing; sharing would make it much lower):
The price of cheap: your code will be used for training
Official documentsIt’s stated very plainly: Contributor mode is “licensing Meta to use your prompts and completions to train future models in exchange for heavily discounted pricing.” In practical terms, the input price gap is 12.5 times, while the output price gap is as high as 21 times. Estimating at US$1 ≈ NT$33, the cheapest Contributor plan costs only NT$6.6 per million output tokens.
In other words, to get this ultra-low price, you have to accept Meta analyzing your code and conversations. For individual developers or non-sensitive projects, this trade-off might be worthwhile, but for enterprise users, feeding repository content into Meta’s training pipeline involves code confidentiality and data compliance issues, making it a decision that requires careful evaluation. As a result, the Contributor program has sparked discussion in the community, and the media has also responded with headlines like…Use data for a 21x discount.” from that angle to report on this release (editor’s note: but at least they were honest about it — some companies might take your money and still use your data for training).
The timing is telling: DeepSeek raised prices, Meta cut theirs.
The most subtle timing of this release is that it came almost simultaneously with DeepSeek’s price increase announcement. On August 6, DeepSeek posted a pop-up notice on its developer open platform stating that “we plan to raise overall pricing for DeepSeek API services in the near future, with a relatively significant expected increase,” though specific rates and the effective date have yet to be announced. The price adjustment applies only to API services and does not affect regular conversations on the web or app. Prior to this, DeepSeek had already adjusted API prices multiple times this year.
Meta unveiled a lower-priced coding model at the same time DeepSeek is gearing up for a price hike — the intent is fairly obvious. The roster of players in the AI price war keeps growing. Meta is taking on DeepSeek’s open-weights approach with a closed-source model, trading Contributor discounts for data and competing for developers with low prices. The next thing to watch is whether other vendors will follow suit once DeepSeek officially raises its prices. For developers, the direct upside of this competition is more options to choose from. But before handing your code over to any model provider, reading the fine print matters more than rushing to grab the cheapest deal.
Conclusion
The release of Muse Spark 1.2 has pushed the model price war to a new stage. In terms of performance, it has firmly secured a place in the top tier; in pricing, it pioneered a new “data-for-discount” model; its open API approach allows the low prices to be directly applied to other agent frameworks; and the timing was precisely set on the same day DeepSeek announced a price increase. For developers, this is good news—more options and lower prices. However, the data terms of the Contributor model also remind us that there’s no such thing as a free lunch; behind ultra-low prices, there is always a corresponding cost.
Source: KOCPC Chinese