What is the OpenAI agent message board?

A shared communication layer where autonomous agents can post messages, read responses, and coordinate workflows without human intervention. Discovered in September 2026.

How much does agent-to-agent communication cost?

At GPT-6 Astra rates: $10/M input, $1/M cached, $50/M output. A 10-turn, 3-agent coordination costs ~$15. On GPT-6 Sol: ~$3 for the same exchange.

Can agents run up unbounded bills on the message board?

Yes. Without message budgets, depth limits, and idle timeouts, negotiation loops and broadcast storms can consume millions of tokens. FinOps teams must enforce hard caps per workflow.

Should I use Astra or Sol for agent coordination?

Use Sol for coordination (cheap tokens, sufficient for handoffs). Reserve Astra for the final reasoning/synthesis step where its reasoning pays off.

OpenAI agent message board cost model

Updated September 27, 2026 · first published September 27, 2026

In September 2026, researchers discovered an OpenAI agent message board — a communication layer where autonomous agents can post messages, read responses, and coordinate multi-step workflows without human intervention. For FinOps teams, this introduces a new cost vector: agent-to-agent token consumption that scales with conversation length, not task complexity.

What the message board enables

The message board acts as a shared context bus. An orchestrator agent posts a task decomposition; specialist agents claim subtasks, post findings, and negotiate handoffs. Each message is a full prompt-completion cycle billed at the model's standard rates.

Cost model

Component	Rate (GPT-6 Astra)	Rate (GPT-6 Sol)
Agent post (input)	$10.00 / 1M	$2.00 / 1M
Agent read (cached input)	$1.00 / 1M	$0.20 / 1M
Agent reply (output)	$50.00 / 1M	$10.00 / 1M

A 10-turn coordination between three agents at Astra pricing: ~300K tokens = $15. Same exchange on Sol: $3. The difference is whether the task requires Astra's reasoning or if Sol's cheaper tokens suffice.

Runaway scenarios

Negotiation loops: Two agents disagree on a handoff and iterate 50+ times before converging.
Broadcast storms: One agent posts to all subscribers; each replies, triggering more posts.
Orphaned conversations: Agents wait for a response that never comes, polling the board every 30 seconds.

FinOps controls

Message budget per workflow: Hard cap on total board tokens per task ID.
Depth limits: Max 3 agent-to-agent hops before human escalation.
Model routing: Route coordination to Sol; reserve Astra for the final synthesis step only.
Idle timeout: Auto-close threads after 5 minutes of inactivity.

Bottom line

The message board makes multi-agent systems composable — but composability without cost guards is a budget leak. Treat agent-to-agent tokens as a distinct cost center with its own alerts and budgets.

Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →

Back to research