GPT-5.6 Sol Hits 750 Tokens a Second. Agent Latency Just Became a Buying Decision

OpenAI's new GPT-5.6 Sol runs at up to 750 tokens a second on Cerebras hardware, roughly 5x the ~150 tokens a second most production models deliver today. For a chatbot that reads as "instant either way." For an enterprise agent that chains 30 or 40 model calls to finish one back-office task, that 5x compounds into the difference between a workflow that finishes in seconds and one that finishes in minutes. Speed stopped being a nice-to-have and became a line item on the buying checklist.

Most coverage of the July 9 launch is about price tiers and benchmark scores. The more useful story for anyone actually running agents is throughput, because throughput is the tax you pay on every single step, and agents take a lot of steps.

Why a single speed number matters more for agents than for chat

A person typing to a chatbot reads at maybe 5 to 10 tokens a second. Any model faster than that feels instant, so the last few years of speed gains were mostly invisible to end users. Agents broke that.

An agent doing something real, screening a candidate, matching an invoice to a purchase order, reconciling a payment, doesn't make one model call. It makes dozens: read the document, extract the fields, check them against a system, decide the next action, draft the reply, verify it. Each hop waits on the one before it. So latency isn't a fixed cost you pay once. It multiplies by the number of steps in the workflow.

That's why GPT-5.6 Sol at 750 tokens a second is a bigger deal than it looks. (OpenAI shared the full lineup in its GPT-5.6 launch announcement.) A 40-step workflow that spent 8 seconds per step now spends closer to 2. Across thousands of runs a day, that changes what you can put an agent in front of. Live customer conversations, same-day claims intake, anything with a human waiting, all become viable when the round trip drops under a threshold people will actually tolerate.

The pricing tells you OpenAI is thinking about volume

GPT-5.6 shipped as three models, not one, and the shape of the lineup is the interesting part:

A tiered family priced from $1 to $30 per million output tokens is a company designing for agents that run millions of calls, not for one person in a chat window. Luna at $1 in / $6 out exists so you can run the cheap, high-frequency steps somewhere cheap. Sol exists for the steps where speed or reasoning actually decides the outcome.

Which points at the real lesson, and it isn't "switch everything to Sol."

No single model wins every step of a workflow

The temptation with any new flagship is to route everything to it. That's the wrong read. A candidate-screening workflow might use a cheap fast model to parse 500 resumes, a stronger model to rank the shortlist, and a careful one to draft the outreach. Paying flagship output rates to reformat a phone number is just lighting money on fire.

We've argued before that one model can't fit every business, and GPT-5.6's own three-tier lineup is OpenAI conceding the same point inside a single release. The job isn't to pick the best model. It's to route each step to the model that clears the bar for that step at the lowest cost and latency, and to be able to swap models the week a faster one ships without rebuilding the workflow.

That routing is exactly what a model-agnostic agent platform is for. When Sol is fastest this quarter and something else is fastest next quarter, you want to change one setting, not re-plumb your operation.

What to actually do about it

• Measure your workflows by end-to-end time, not model benchmark scores. A model that's 3 points higher on a benchmark but half the speed can be the wrong choice for a live workflow.

• Find the steps where a human is waiting. Those are where throughput like Sol's pays for itself. Batch steps that run overnight don't need it.

• Don't standardize on one model. Route by step. Keep the expensive, fast model for the moments that need it.

• Check availability before you plan around it. Sol's fastest configuration is rolling out to limited partners first and expanding as capacity grows, so treat the 750 number as the ceiling, not day-one reality for everyone.

The headline is a faster OpenAI model. The story underneath is that agents changed what "fast enough" means, and speed is now something you buy for, not something you assume.

Want the enterprise-agent read on releases like this? We send a short weekly write-up on what actually matters for teams running agents in production. Get it here.