The Frontier Fills Up, and Its Price Converges

MENLO PARK, Calif., Oct. 17 — MetaMeta on Monday released Muse Wonder, its first model built to compete at the frontier of artificial intelligence, and priced it for developers at exactly what Anthropic charges for Claude Opus 5.5.

That price is $4 for every million tokens a customer sends the model and $20 for every million it generates. From Dec. 1, Google’s Gemini 4 Argon will cost the same on its standard paid tier. Google opened its newest frontier model to developers and businesses on Tuesday at an introductory $2 and $10 on that tier, good through Nov. 30, and to paying subscribers in its Gemini app.

The other moves came later in the week. On Thursday Meta released the weights of a second model, Muse Spark Open, for download under its own license, and Microsoft’s in-house laboratory said it had finished pre-training a large model of its own. On Saturday DeepSeek, of Hangzhou, upgraded its Flash model and left its prices, less than a tenth of the $4-and-$20 rate, unchanged.The TokenThe smallest unit of text a model recognizes: a word, a number or a punctuation mark, in DeepSeek’s definition. Prices are quoted per million.

The frontier is filling up, and its list price is settling. That matters to the laboratories, which must earn back what it cost them to get there; to the companies building products on their models, which buy the tokens; and to the cloud providers, which sell several laboratories’ models side by side. One of them, Microsoft, now trains its own.

The Meter

The list price, though, is not the bill. Artificial Analysis, an independent firm that runs its own tests, scores models on a composite index and records the weighted average cost of each task in it.The IndexArtificial Analysis’s Intelligence Index v4.3.2 combines 10 evaluations, among them Terminal-Bench 4.0 and Humanity’s Last Exam. Settings tested: “Max, Default Fallback” for Opus 5.5, “high” for Argon, “xhigh” for Muse Spark Open and “max” for the rest. Figures as its site showed them on Oct. 17. Opus 5.5 scored 58, Argon 53 and Muse Wonder 52; OpenAI’s GPT-6 Astra and GPT-6.1 Sol scored 53 and 52.

At the identical list price, Opus 5.5 cost a weighted average of $5.98 per task on Artificial Analysis’s independent index, and Muse Wonder $4.52. Across the index, Opus generated 260 million output tokens and Muse Wonder 130 million. Astra, whose Standard short-context input and output rates are two and a half times as high, generated 60 million and cost a weighted average of $3.26 per task. Anthropic also cautions against reading much into the scores: “at these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences.”

The Margin

At Anthropic, $4 and $20 is the second shelf: its Claude Fable 5.1 lists at $10 and $50, as does OpenAI’s Astra at its Standard short-context rate. Of the cheaper model, Anthropic says, “Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that.” By its account, the price fell because the cost did.

Meta is holding something back, too. Muse Wonder Full, the stronger version, which scored 56 on the independent index, goes only to subscribers of Meta’s Muse Code and is not sold through its developer service. Meta’s open-weight model, Muse Spark Open, comes with conditions: its license requires notice to Meta and a separate written agreement before the model or its derivatives are offered as a model-access service, if the licensee and its affiliates together exceed $20 million in gross revenue in any consecutive 12 months. Internal use and features in products that are not model-access services are exempt. On the independent index, the 816-billion-parameter model scored 39, behind China’s GLM 5.3 at 45 and Kimi K3 at 44.The HardwareMeta says the model needs at least 700 gigabytes of GPU memory in its compressed NVFP4 version, and that it is optimized to run on two Mac Studios with M5 Ultra chips.

Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that.

Microsoft’s in-house laboratory, Microsoft AI, said its MAI-Base-2, with 2.8 trillion parameters, 80 billion of them active, was trained on 16,000 GB200 chips over 38 days, and that it plans to license the weights to trusted partners through Foundry, its service for businesses. The model is still in post-training. Microsoft estimates that its training recipe could match DeepSeek-V4-Pro-Base, which it calls the strongest open base model, with roughly a fourth to a twelfth of the computing DeepSeek used. Mustafa Suleyman, who runs Microsoft AI, told employees at a town hall on Thursday that “we now have a stronger pre-trained base model than the model that currently powers GPT 5.6 Sol,” trained on a quarter of the compute OpenAI used, Business Insider reported. Mr. Suleyman said the comparison rested on private tests of how well models understand and predict patterns, drawing on technical information Microsoft holds about OpenAI’s models. Microsoft declined to comment to Business Insider, and OpenAI did not respond; Microsoft’s published account does not mention OpenAI. Microsoft’s finance chief, Amy Hood, called model diversification “a margin improvement opportunity” on the company’s July 29 earnings call, according to its transcript.

For customers, the pressure runs the other way. DeepSeek charges 30 cents per million uncached input tokens and $1.20 per million output at peak hours, and half that otherwise.Off-PeakPeak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. Its service accepts requests in Anthropic’s format. Its new search API costs 75 cents per 1,000 successful requests, but those search requests are currently free when using the DeepSeek API through Claude Code or Codex; billing starts at midnight Beijing time on Nov. 1. When DeepSeek cut prices in September, shares of MiniMax, Z.ai and Alibaba fell sharply in Hong Kong, Bloomberg reported. Christian Kroll, chief executive of the German search engine Ecosia, which is dropping Mistral, told Politico that partnering with Melious, a platform hosting a range of open models, among them China’s Qwen, GLM and Kimi, had “roughly cut our costs in half while improving quality and performance.” Mr. Kroll told Politico he sees open-weight models as a huge opportunity for Europe.

None of these model announcements came from OpenAI, which paused training of its most advanced systems in late September; a report elsewhere in this edition takes up why.

Any lead may be brief. An analysis by the Organization for Economic Cooperation and Development, cited by Fortune, found that on average a model holds the best performance for its price for only a few months. DeepSeek begins billing those searches at midnight Beijing time on Nov. 1 (16:00 UTC on Oct. 31). Google’s introductory rate for Argon runs through Nov. 30, and its standard price doubles on Dec. 1. Microsoft reports its quarter on Oct. 28; it has given no date or price for MAI-Base-2.

Daily Cortex is written with AI models, including Anthropic’s Claude, and fact-checked with OpenAI’s GPT models.