The cheapest rate on the market is no longer paid in money: Meta bills its coding agent twelve times cheaper in exchange for the right to train on your code

On August 5, 2026, Meta released in beta Muse Code, a terminal coding agent built on its Muse Spark 1.2 model, and so took its place against Claude Code, Codex and Gemini CLI. The event is not the agent: it is its pricing grid. Meta offers the same model at two prices. The tier called Standard is billed $1.25 on input and $4.25 on output per million tokens, and Meta uses neither the requests nor the responses. The tier called Contributor is billed $0.10 and $0.20 (about twelve times cheaper on input and twenty-one times cheaper on output), and Meta reserves the right to use your requests and the responses produced to train its future models. In other words, the cheapest line on the market in the coding-agent segment is not a price in money: it is a barter. And during the same week, two other prices moved in the opposite direction from the one everybody is commenting on: ByteDance’s generative video rose by 53%, and the H200 GPU-hour settled 25% above its August 2025 level.

What Meta put on the market on August 5

Muse Code is a command-line coding agent, available in early beta on macOS and Linux, whose announced differentiating feature is the management of persistent asynchronous background agents: tasks that outlive the session that launched them. The model that powers it, Muse Spark 1.2, succeeds Muse Spark 1.1, Meta’s first paid API, which we recorded at $1.25/$4.25 in our August 2 grid. The Standard tier of version 1.2 reproduces exactly those amounts, with cached input at $0.15.

The novelty is the second tier. The muse-spark-1.2-contributor endpoint is billed $0.10 on input, $0.002 for cached input and $0.20 on output. The quid pro quo is explicit in the published terms: requests and completions may be used to improve Meta’s models. The discount is not marginal: it places Muse Spark 1.2 below DeepSeek V4 Flash ($0.14/$0.28) and level on input with Ministral 3 3B, that is, in the zone of the cheapest models on the market, for a model that plays in another capability category.

Three restrictions come with this rate, and they matter as much as the price. The permitted throughput is 60 requests per minute against 3’000 on the Standard tier, a factor of fifty. Quotas are assessed over a five-hour rolling window. And access is limited to a list of selected countries, not detailed in the launch documents. Meta also says it is starting to accept zero data retention requests, presented as a feature for enterprises; enterprise pricing terms, volume discounts and service commitments do not appear in the launch documents and are a matter for commercial negotiation.

One point deserves to be flagged for what it is: reported information, not confirmed by the documentation we were able to read directly. Several third-party guides published around August 5 claim that Muse Code installs by default on the Contributor tier, and that an explicit action is therefore needed to switch to Standard. We did not find this mention in the launch materials themselves, and we do not take it up as an established fact. But the practical consequence is independent of its verification: any organisation deploying this tool must check the active tier before opening a client code repository, because the question does not arise the other way round. A setting that decides whether a client’s code goes into a training pipeline or not is not a comfort parameter.

The benchmark has lost its unit along with the price

The announced performance illustrates the same problem of commensurability. Meta claims 82.9% on Terminal-Bench 2.1 for Muse Code on Muse Spark 1.2, measured on its own evaluation harness. As of August 6, Muse Spark 1.2 did not appear on the verified Terminal-Bench leaderboard. Vals’s independent evaluations place it 5th on the Vals index, at a cost of $0.69 per test, but 14th on Terminal-Bench measured on a common harness. A third-party reading sums up the gap with a useful formula: of the 6.7 points Muse Spark 1.2 gained on Terminal-Bench, two survive a constant harness.

The comparison with the competition should be taken with the same caution, because the published rankings do not reconcile. One reading gives GPT-5.6 Sol 89.5% and Claude Opus 5 89.1%; another, on verified data, gives Sol 85.77%, Opus 5 84.64% and Kimi K3 80.90%. We do not settle it: the gap of nearly four points on the same models indicates that these figures circulate without a common harness, and none should be cited as an absolute measure. The only robust statement is this one, and it comes from reading Meta’s materials: Claude Opus 5 is ahead on every table Meta chose to show, including the one Meta built and controls. A supplier that publishes a comparison in which it loses is selling something other than performance. Here, it is selling price. And the price, on the tier that makes the difference, is paid in data.

At the other end, prices are rising

The market’s dominant narrative is one of continuous price collapse. It is accurate for the text token, and wrong everywhere else. Two counter-examples date from this week.

The first is generative video. ByteDance’s Seedance 2.5, whose API opened on August 7, 2026, is billed 52.9% more per token than Seedance 2.0 without video input, and 48.8% more with it. On matched five-second examples, the real gap comes out at about +45% at 480p and +52% at 720p. The rate is $10.97 per million output tokens, or, on the Volcano Engine grid, 70 yuan per million billable tokens without video input and 42 yuan with. Billing covers only the video produced: text requests, the first-frame image and reference images are free. And the published range of resolutions has narrowed: Seedance 2.0 lists 480p, 720p, 1080p and 4K, while the public grid for 2.5 lists only 480p and 720p. More expensive, for fewer resolutions available: the opposite trajectory to text.

The second counter-example is compute. The median on-demand price of an H200, recorded across a panel of suppliers, has risen about 25% since August 2025, from $3.50 to $4.39 an hour per GPU. Dispersion remains extreme: from $2.30 to $13.78 an hour depending on the supplier, for a cohort median around $4.11, with one specialised supplier showing $3.99 on demand and $1.99 spot. For the H100 80 GB, readings range from $1.49 to $6.98 an hour, with another panel giving $2.19 to $11.06. A warning accompanies these figures and is worth keeping: on the H200, supply is volatile, and “available” frequently means a capacity request or a reservation of several weeks. An hourly price displayed on a machine you cannot get for three weeks is not a price.

This is the context in which to read Anthropic’s announcement of August 5: the confirmation of the formation of an in-house chip design team, with the ambition of co-designing the silicon and the Claude models, and a stated target of reducing per-token inference cost by about 50%, specialised inference chips being credited with gains of 50 to 67% per token at production scale. The AWS, Google, Nvidia and AMD partnerships remain central; the company is recruiting semiconductor engineers at up to $485’000 a year. Two reservations, explicit in the coverage: no timetable is given, and it is not specified whether Anthropic intends to control manufacturing. Industry sources put the development of an advanced AI chip at close to half a billion dollars. The message sent to the market is clear (the cost of the token is now a silicon problem, not a commercial-margin one), but it will have no effect on an invoice for several years.

“Open weights” is not a price either

The third register in which the unit of account slips away is that of open-weight. We noted on the morning of August 10 that the weights of Qwen3.8-Max (2.4 trillion parameters, 95 billion active, a 1-million-token context, $2/$6 on the API) and of Qwen3.8-27B were announced for the week of August 10 on Hugging Face and ModelScope. The state of that promise at the time of writing deserves precise note: Alibaba has made Qwen3.8-Max widely accessible to international users ahead of the weights’ publication, it would be the first Max-class model Alibaba opens, and the licence has still not been announced.

This last point is not an administrative detail. Weights published under an Apache 2.0 licence, an MIT licence, or an in-house licence with restrictions on commercial use or active-user thresholds do not define the same asset, nor the same legal risk, nor the same value for an organisation considering sovereign hosting. As long as the licence is not published, “open weights” describes an intention, not a purchase option. The reading rule is the same as for Meta’s Contributor rate: what determines the real cost is not the displayed number, it is the clause that comes with it.

Comparison table: API prices as of August 10, 2026

Two lines enter the grid since our reading this morning: the two tiers of Muse Spark 1.2, which replace the now obsolete Muse Spark 1.1 line. The prices below are the suppliers’ standard rates, excluding cache, excluding Batch and excluding marketplaces.

Model	Input ($/M tokens)	Output ($/M tokens)	Note
Claude Fable 5 (Anthropic)	10.00	50.00	frontier; available on Google Cloud
GPT-5.6 Sol (OpenAI)	5.00	30.00	Batch/Flex 2.50/15.00; Fast mode 10.00/60.00
Claude Opus 5 (Anthropic)	5.00	25.00	ahead of Muse Spark 1.2 on every table published by Meta
Kimi K3 (Moonshot, open-weight)	3.00	15.00	1M context at no extra cost; cache hit 0.30
GPT-5.6 Terra (OpenAI)	2.00	12.00	Batch/Flex 1.00/6.00; Fast mode 4.00/24.00
Gemini 3.1 Pro (Google)	2.00	12.00	4.00/18.00 beyond 200k context
Claude Sonnet 5 (Anthropic)	2.00	10.00	launch rate until August 31, then 3.00/15.00
Qwen3.8-Max (Alibaba)	2.00	6.00	2.4T params / 95B active; weights announced this week, licence unpublished
Grok 4.5 (xAI)	2.00	6.00	coding model; EU access still restricted
Gemini 3.6 Flash (Google)	1.50	7.50	geared to multi-step automation
Qwen3.7 Max (Alibaba)	1.475	4.425	marketplace reading; no context tiers
Muse Spark 1.2 · Standard (Meta)	1.25	4.25	launched August 5; cache 0.15; Meta does not train on your data
Grok 4.3 / 4.20 (xAI)	1.25	2.50	1M context
Inkling-Small (Thinking Machines)	0.58	1.20	open weights; output at 1.44 at third parties, not corroborated
Mistral Large 3	0.50	1.50	December 2025 rate; Batch at −50%
DeepSeek V4 Pro	0.435	0.87	supplier rate, identical on marketplaces
Qwen3.7 Plus (Alibaba)	0.32	1.28	0.96/3.84 beyond 256k context
Gemini 3.5 Flash-Lite (Google)	0.30	2.50	very high throughput
GPT-5.6 Luna (OpenAI)	0.20	1.20	−80% on July 30; default on Free and Go since August 6
DeepSeek V4 Flash 0731	0.14	0.28	82.7 on Terminal-Bench 2.1; ≈ −37% on marketplaces
Ministral 3 3B (Mistral)	0.10	0.10	flat rate
Muse Spark 1.2 · Contributor (Meta)	0.10	0.20	cache 0.002; Meta may train on your requests and responses; 60 RPM; selected countries
Qwen3.7 Flash (Alibaba)	0.03	0.13	under 32k prompt only; 0.10/0.40 up to 256k; 0.20/0.80 beyond

Muse Spark 1.2 rates (both tiers) verified on August 10, 2026; the other lines are carried over from our reading on the morning of August 10. Cache, promotional and launch rates expire without notice. On the market floor, two readings do not reconcile: one tracker gives, as of August 8, Llama 3.1 8B Instruct at $0.02 input as the cheapest production model, while another retains Qwen3.7 Flash at $0.03; the scopes (“tracked production model” versus supplier grid) differ and these two figures are not comparable.

The table above cannot accommodate media models, and that is precisely the subject of this article: they are not billed in the same unit.

Model	Billing unit	Price	Note
Seedance 2.5 (ByteDance)	per second of video	$0.09/s at 480p · $0.21/s at 720p	API opened August 7; +52.9% per token vs 2.0; 1080p and 4K absent from the public grid
Seedance 2.5 (ByteDance)	per million output tokens	$10.97	free inputs: prompt, first frame, reference images
Qwen Image 3 Pro (Alibaba)	per image	≈ $0.034 to $0.04	depending on supplier and resolution tier; free for a limited time on Bailian

An identical five-second 720p clip is billed $1.244 with a short reference and $4.838 with a thirty-second reference: a factor of 3.9 for an identical deliverable. No per-million-token grid makes it possible to anticipate that gap.

What this means for a Swiss organisation

Three readings, in the order in which they commit a decision.

First, a discount paid in data is not a discount, it is a transfer of rights. For a fiduciary, a law firm, a private bank or a developer under client mandate, the code and documents processed do not belong to whoever processes them. Running them through a pricing tier that authorises their use to train a third party’s model is not a budget trade-off: it is a question of professional secrecy and contractual confidentiality obligation, and it is settled before looking at the price, not after. The gap of twelve to twenty-one times is real and attractive; it is also the measure of what your data is worth to the party asking for it. Two practical rules follow: check the active tier of any agent tool before bringing it near a client repository, and demand the applicable regime in writing. Meta is starting to accept zero-retention requests, which presupposes making them.

Next, falling prices are not a law of nature, they are a commercial policy by segment. This week sums up the situation: the text token falls, generative video rises 53%, the H200 GPU-hour exceeds its level of a year ago by 25%, and the only player announcing a halving of inference cost does so by building its own chips, with no timetable. An organisation that built its budget plan on the assumption of continuous deflation has extrapolated one segment to the whole market. The hedge is not to forecast the price, it is to measure its real cost per completed task and to be able to move that task when the grid shifts.

Finally, the displayed figure has ceased to be a unit of comparison, in all three registers at once. Price, because the cheapest line is paid for in rights of use over your data. Performance, because the same model loses nine places depending on the harness that measures it, and because two published rankings differ by almost four points on Sol and Opus 5. And availability, because “open weights without a licence” and “a GPU available within three weeks” describe intentions, not purchase options. What remains measurable is what happens at your end: which task was run, by which model, under which data regime, at what observed cost. An organisation that instruments this compares real quantities. An organisation that reads grids compares units that no longer have a common measure.