Agora-2 vs MiniMax M3: One GPU Per Participant, or Thirteen Cents Per Million Tokens

Two ways to run a crowd of agents, priced in two different currencies. MiniMax M3 costs $0.30 per million input tokens and $1.20 per million output — a rate that makes a thousand-agent experiment a line item rather than a funding round. Agora-2, the multi-agent world model Odyssey previewed on September 21, 2026, costs one NVIDIA RTX PRO 4500 per participant view: a four-player session runs on four GPUs, with one rendering model per card generating that player's 640×480 perspective, and one of those four cards also carrying the shared simulation and the sixteen autonomous agent policies. The MiniMax M3 number is per token and scales with how much your agents think. The Agora-2 number is per eyeball and scales with how many simultaneous participants you put in the world. Comparing them is comparing a reasoning budget with a rendering budget, and for anyone planning a multi-agent study in 2026, that turns out to be the useful thing to do.

The token side of the ledger

MiniMax M3 is MiniMax's open-weight flagship, released May 31, 2026: a 428-billion-parameter sparse mixture-of-experts with roughly 23 billion parameters active per token, a 1,048,576-token context window of which MiniMax guarantees 512K usable, text, image and video input, and a score of 29 on the Artificial Analysis Intelligence Index v4.3.2. It reports 83.5 on BrowseComp and 76.1 on BankerToolBench in vendor figures — MiniMax's own numbers, unreproduced here — and Artificial Analysis measures it at 145 output tokens per second, the tenth-fastest model on that board.

The relevant number for agent fleets, though, is the cache-read rate: $0.06 per million cached tokens, one fifth of the input price. Agent loops are prefix-dominated — the same system prompt, the same tool schemas, the same accumulating history, re-sent every turn — so the effective cost of a loop is far closer to $0.06 than to $0.30 for the bulk of its tokens. That is the arithmetic that makes a 1,200-agent run, of the kind METR documented inside OpenAI's July 2026 ExploitGym evaluation, something a research group can actually reproduce rather than merely read about.

The GPU side of the ledger

Agora-2's cost structure is disclosed in its own implementation appendix, and it is refreshingly concrete. One RTX PRO 4500 Blackwell Server Edition per view. Four for a four-player session. Those cards generate each participant's view at a target of fifteen latent updates and thirty displayed frames per second at 640×480, and the simulation advances on a 30 Hz tick. The report also gives the training scale for the surrounding work: an effective global batch of 128 across 32 B200 GPUs.

Translated into the same unit as MiniMax M3, that is one data-centre GPU per concurrent participant, plus the shared simulation riding along on one of them. It is a fixed cost that does not care how long your agents deliberate and does not drop when they go quiet. Two consequences follow, and they point in opposite directions. At low participant counts with heavy reasoning per agent, the token model is obviously cheaper — you are paying cents for thinking and nothing for the second agent in the room. At high participant counts with dense interaction, the arithmetic reverses: twenty concurrent participants in a shared world is twenty GPUs, a number that does not grow with the number of tokens each one spends.

• Unit of cost — Agora-2 one RTX PRO 4500 per participant view vs MiniMax M3 $0.30/$1.20 per 1M tokens

• Cache reads — Agora-2 not applicable, no token billing vs MiniMax M3 $0.06 per 1M cached

• Independent score — Agora-2 none published vs MiniMax M3 AA Intelligence Index 29 (v4.3.2)

• Context — Agora-2 shared state plus bounded per-view history vs MiniMax M3 1,048,576 tokens, 512K guaranteed

• Output — Agora-2 640×480 video at a 30 fps target vs MiniMax M3 text

• Access — Agora-2 browser preview, no API, no weights vs MiniMax M3 open weights, community licence, API

Why the two costs are not substitutes

It is tempting to read this as an either/or, and it is not one. MiniMax M3 is a policy brain: it reads a partially observed situation, reasons, calls tools and emits an action. Agora-2 is an environment in which actions have consequences visible to other participants — a simulation model that predicts how combined actions change shared state, a world server that applies them, and per-view renderers so that each participant sees the same world from their own perspective. Odyssey's sixteen autonomous entities are not driven by any language model; they are recurrent scalar and spatial policies trained with reinforcement learning, consuming a sixteen-observation history and acting every four simulation ticks. That design choice is the tell. At thirty ticks a second, deciding what to do is a control problem, and control problems are solved by training a small policy rather than by calling an API.

So the honest framing for a team planning multi-agent work is that these live at different layers and you need a budget for each. The reasoning layer is cheap and elastic and you can shop for it. The environment layer, if you want something other than a game engine, is currently a research preview with a GPU-shaped footprint and no published API. Both facts are new enough — M3 since May, Agora-2 since September — that almost nobody has a cost model for the combination yet.

What is verified and what is a target

MiniMax M3's independent score, context window, modalities and price are published facts, checkable today. Its BrowseComp and BankerToolBench figures are vendor-reported and should be labelled as such whenever you quote them.

Agora-2's numbers come entirely from Team Odyssey's technical report and have been reproduced by nobody outside the company. Its rendering result — 30.11 dB PSNR for the structured-prefix renderer against 20.67 dB for a VAE baseline on thirty monitoring clips — is a comparison of complete systems with unmatched training budgets, as the report itself notes. Its 45.8% denoiser-time reduction versus cross-attention comes from randomly initialised denoisers on synthetic inputs and measures execution cost rather than image quality. And its limitations section states plainly that the fifteen-updates-per-second and thirty-frames-per-second rates are targets; that sustained frame rate and input-to-display latency during live multi-agent interaction "have not been quantitatively evaluated"; that long-horizon visual quality, cross-view agreement and end-to-end action adherence remain unevaluated; that the environment needs an instrumented engine with explicit geometry and a fixed appearance vocabulary; and that transfer beyond the training domain is untested. Anyone building a cost model on one GPU per view should read that section first, because the throughput per GPU is exactly what has not been measured.

The call

If you are building agent fleets — many models, long loops, tool calls, no rendering — MiniMax M3 is the cheapest credible way to do it at scale, and the cache-read rate is the number that decides your bill. It is routed on OrcaRouter at MiniMax's own list price with zero markup, so the $0.06 cached rate is what you pay, with failover across providers if an endpoint degrades mid-run. For fleets specifically, the pass-through matters more than usual: a reseller margin on cached tokens is a margin multiplied by every turn of every agent.

Agora-2 is not something you can route to — there is no API, no price and no weights, only Odyssey's browser preview — and we do not host it. What it is, is the first shared-world simulator built specifically so that many participants, human and synthetic, can be watched interacting under controlled conditions, and the report underneath it is unusually forthright about how far it currently is from a product. If your problem is reasoning at volume, MiniMax M3 is available now and cheap. If your problem is what happens when a thousand of those reasoners share a world, Odyssey has built the arena and left the hard measurements for someone else to run.