ONE · WHAT SHIPPED Google put Gemini 3.7 Flash into general use on the thirteenth of August and called it its most intelligent workhorse model for coding and agents. Workhorse is the honest word. This is not the top of the range and it is not sold as such. It is the one you point at the volume work, the long chains of tool calls that make an agent an agent, where the number that matters is what a million tokens costs you rather than what the model scores. The number is seventy five cents per million tokens in and three dollars seventy five out. It carries a one million token context and takes images and audio as well as text. Google's own figures, and they are Google's, put it at 30.4 percent on AutomationBench and 1588 Elo on Code Arena, both of which it puts at the front of its class. The same sheet has it at 65.3 percent on DeepSWE against 69.6 for GPT-5.6 Terra, and 38.1 percent on OSWorld against 50.2 for the same rival. So it loses two of the four it published, which is a more useful thing to know than the two it won. Reach for it where the work is wide and repetitive. Keep something else for the tasks where a single misstep costs an afternoon. There is also a tier you can use without paying anything, with rate limits attached, which is the sane way to find out whether any of the above survives contact with your actual work.  GO TO THE SOURCE The announcement is here, and the rate sheet itself, which is the page that actually matters and the one nobody reads, is here. Read the second one before the first one. The blog post is written to make you want the model. The rate sheet is written to tell you what it does to your monthly bill, and it is the only document of the two that changes on the first of January. TWO · WHAT THE FAST ONE DOES NOT TELL YOU On the same day, OpenAI and Cerebras previewed Ultrafast, a service tier that runs the existing GPT-5.6 Sol at 750 output tokens a second, which they put at up to fourteen times the standard tier. Watching it work is genuinely startling. An agent that used to think in front of you now simply answers. Three things it does not come with. No published rate. No date for general use. No model name you can put in a request. It is a limited preview of a speed, not a thing you can build on this month. This is worth naming plainly because the two announcements landed the same day and read as though they belong in the same decision. They do not. One is a number you can put in a spreadsheet today. The other is a demonstration. The habit worth keeping is to sort every launch into those two piles before you let it change any of your work, and to notice how often the loudest number of the week sits in the second pile. THREE · THE DECISION THAT IS DUE The introductory number holds until the thirty-first of December. On the first of January it becomes a dollar fifty in and seven dollars fifty out, which is exactly double. Nothing about the model changes that day. Only the bill does. Which means the next four months are a measuring window, and most people will not use them as one. Put the model on the volume work now, at the low number, and keep a record of what a normal week actually consumes. Then you will be deciding in December with a real figure in front of you instead of a feeling. If the doubled rate still works at your volume, you carry on and nothing happens. If it does not, you already know, and you have had four months to find the thing you move to. For comparison while you measure: after the thirtieth of July reduction, GPT-5.6 sits at twenty cents in for Luna, two dollars for Terra and five dollars for Sol, per million. And there is a floor under all of this that no vendor sets, which is a model running on your own machine, where the rate is nothing and the expiry date does not exist. Google's own note on running Gemma 4 12B on a laptop is the version of that worth reading, and the wider state of open models this summer is the context around it. PROMPT OF THE DAY The one that reads a rate sheet for you. Reach for this when a model you use publishes new numbers and you want to know what it does to your month rather than what the announcement says about it. Paste the rate page in after the instruction. Here is a rate page for a model. Do four things with it. First, list every number that is temporary, with the exact date it changes and what it becomes. Second, list every number that is permanent. Third, work out the monthly bill at two volumes I will give you, once at today's numbers and once at the numbers after every temporary one has expired. Fourth, tell me which single line in this page would hurt me most if I did not notice it, and say why. Quote the exact wording from the page for anything you assert, and if the page does not say something, say that it does not say it rather than filling it in. The load-bearing line is the last one. Without it the model will quietly complete the parts the page leaves out, and a rate sheet is exactly the kind of document where the missing part is the part that matters. The second calculation is what turns this from reading into a decision. THE NEXT WORKING NOTE A model you chose on quality alone is a model you will re-choose every time somebody publishes a chart. A model you chose on quality and a number you measured yourself is one you can defend in December when the number moves. The second kind of decision survives the news cycle. The first kind is the news cycle. I am putting together a five model rate card: what each of the current working models charges in and out per million, which of those numbers carry an expiry date and when, and the monthly figure at three realistic volumes, each with a worked example. Built so you can answer the December question in a minute rather than an afternoon. Want it when it ships? Reply with send me the rate card and I will get it to you. A QUESTION FOR YOU Do you actually know what your models cost you in a normal month, or is it a number you would have to go and look up. No judgement either way. I am asking because I suspect most of us are in the second group, and I would like to know before I build the card. | 