A 438-billion-parameter reasoning model now sits behind an OpenAI-compatible endpoint priced at $0.60 per million input tokens and $1.80 per million output tokens. Its weights are closed, so that large parameter count does not give developers a model they can inspect, tune or run on their own hardware. Multiverse Computing launched Quasar 438B on September 2 as its first large model, with access through the company's CompactifAI service. The launch offers a hosted model rather than a downloadable artifact.
That distinction matters because Quasar's measured performance is substantial. Artificial Analysis lists it 13th for intelligence and 18th for speed among 178 models in its comparison class. It scored 43 on the firm's Intelligence Index and generated output at 176.2 tokens per second. The same page labels Quasar proprietary and reports a one-million-token context window. Developers can test an unusually large, fast model through a familiar interface, while Multiverse retains control of the weights and serving stack.
The speed claim survives a closer look
Multiverse says Quasar produces a 500-token response in 15.3 seconds, including its thinking time. Artificial Analysis separately measures 176.2 output tokens per second. Those figures describe different slices of a request: one covers the whole wait for a fixed answer, while the other measures generation throughput once output begins. Read together, they suggest that reasoning and initial latency take most of the 15.3 seconds, with token generation itself accounting for only a few seconds. The launch comparison is unusually useful because it includes the thinking delay instead of presenting throughput alone.
Three models in Multiverse's chart complete the same 500-token job faster. Nemotron 3.5 Lightning takes 9.4 seconds but scores 24 on the index; Gemini 3.5 Flash-Lite takes 10.8 seconds and scores 37; Gemini 3.7 Flash takes 11.5 seconds and scores 56. Quasar is therefore neither the quickest nor the highest-scoring model in the chart. Its result is the combination: a score of 43 at 15.3 seconds. Artificial Analysis's model page places that combination near the front of a much larger field, while still describing Quasar as expensive against models in a similar price band.
The intelligence score also needs its denominator. Version 4.1.1 of the Artificial Analysis index combines nine evaluations, including Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond and a long-context test called AA-LCR. Multiverse reports a score of 43 for Quasar, compared with 30 for Mistral Medium 3.5, 38 for Nemotron 3 Ultra and 42 for Inkling. Claude Opus 5 leads the launch chart at 63. The company's description of Quasar as Europe's leading model is based on that selected regional comparison, rather than a claim that it leads every model worldwide.
A million-token window with a verbosity bill
A one-million-token context window is the specification most likely to interest teams building research tools or coding agents that must carry large repositories and long histories. Artificial Analysis reports that limit, while Multiverse gives Quasar a 75.0 score on AA-LCR. That long-context score matches Grok 4.6 high in the company's chart and sits close to Claude Opus 5 at 75.7. A context ceiling says how much a request may contain; it does not establish how reliably the model uses every part of a large prompt. AA-LCR supplies a more relevant test, although one benchmark cannot settle that question for a production workload.
Quasar's output volume may be as important as its token price. Artificial Analysis says the model generated 350 million tokens while completing the Intelligence Index, compared with a 67 million median, and cost $1,047.71 to evaluate. Its analysis calls Quasar very verbose. That creates a practical wrinkle for agent loops: a low listed rate can still produce a larger bill when the model reasons at length or sends more text back on each tool step. Teams comparing costs should record completed tasks and total generated tokens, not rely on the price card alone.
At the public rates, a request with 100,000 input tokens and 10,000 output tokens costs $0.078 before retries or later agent turns. The input contributes six cents and the output contributes 1.8 cents. CompactifAI's model catalog confirms the $0.60 and $1.80 rates. Quasar's measured verbosity makes the output estimate the uncertain part. A long-context agent can also resend accumulated history, turning the cheaper input side into a recurring charge.
What developers can use today
The public Quasar documentation identifies quasar-438b as the API model name. It supports chat completions, tool calling and structured output through an OpenAI-compatible payload. Reasoning is always enabled; callers may set reasoning_effort to high or max, which is the default, but cannot turn reasoning off. That last constraint affects latency tests because even a routine request passes through the reasoning path.
A minimal request uses the standard chat-completions shape documented by CompactifAI, with the Quasar model ID substituted:
curl https://api.compactif.ai/v1/chat/completions \
-H "Authorization: Bearer $COMPACTIFAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"quasar-438b","messages":[{"role":"user","content":"Review this patch"}]}'
CompactifAI's quickstart documents that endpoint and bearer-token scheme. Existing OpenAI-style clients should need fewer changes than an integration with a distinct request format. Compatibility at the HTTP layer does not make the underlying model portable: deployment location, batching behavior and updates remain decisions made by the provider.
Multiverse says Quasar works in English and Spanish and positions it for software agents, technical copilots and document research. Its Terminal-Bench v2.1 score is 69.3, 18.7 points above Mistral Medium 3.5 in the published comparison and 19.8 points below Claude Opus 5. That is credible evidence for testing terminal work, though it is still a benchmark result supplied around the launch. Teams evaluating Quasar should use repository-level tasks with pass-or-fail checks, then compare elapsed time and total billed tokens against the model they already run.
The public model description is thin
The parameter count is precise, but the release leaves several technical questions unanswered. CompactifAI's Quasar specification names the publisher, API ID and 438B parameter count. It does not identify the base architecture, say whether all parameters are active for each token, or document a quantization format. The broader catalog leaves Quasar's original architecture blank. Without those details, 438B cannot be translated into a self-hosting requirement or used for a direct architectural comparison.
The launch article also does not provide a training-data description, data cutoff, model card, safety evaluation or weight license. Artificial Analysis's proprietary label closes the weight-access question, but the other omissions still matter to buyers assessing domain coverage and risk. The published evidence is concentrated on speed, price and selected capability tests. Those figures justify an evaluation. They do not explain how Quasar was trained or how it behaves outside the tested tasks.
Quasar's published results pair a 43 intelligence score with a measured 15.3-second response on the company's 500-token test and a price below two dollars per million output tokens. The next useful evidence will come from independent workloads that report success rate, complete request latency and token use together. Developers should also watch whether Multiverse publishes an architecture description and model card, and whether API performance stays consistent once traffic grows. Public changelogs will matter here because a closed service can change without giving customers a new weight file to pin. Until then, Quasar's benchmark position is clear; its behavior under ordinary production load is still an open test.