AI Radar for Businesses — Saturday, August 15, 2026
· CompaniesAutomation
Today, business AI is played out in what each model call costs. Google published Gemini 3.7 Flash on August 13 with an introductory price of $0.75 per million input tokens and $3.75 output—half the previous version—but only until December 31: on January 1, it shifts to $1.50 and $7.50. The same day, DeepSeek presented V4-Pro and announced a hike of up to 1,100% starting August 16, with time-of-use billing. IBM and OpenAI sealed an alliance bringing GPT-5.6 and ChatGPT Work to IBM Consulting, and OpenAI launched Ultrafast: latency is now sold as a separate line item.
Yesterday's radar was about who pushes the button; today's is about how much it costs to push it. Within twenty-four hours, Google has halved the price of its agent-oriented model—with an expiration date of December 31—and DeepSeek has announced it will quadruple its own starting Sunday, also debuting time-of-use pricing. In parallel, IBM has become OpenAI's installation arm within large enterprises, and OpenAI has begun selling speed as a separate line item on the bill.
Google slashes price of its agent-powering model by half, but only until year-end
On August 13, Google released Gemini 3.7 Flash, just three weeks after 3.6 Flash, with an introductory price of $0.75 per million input tokens and $3.75 per million output tokens: half that of the previous version. This price is valid until December 31, 2026; on January 1, it rises to $1.50 and $7.50. In coding capabilities, it sees a strong increase—DeepSWE v1.1 up from 49.0% to 65.3% and FrontierCode 1.1 Main from 34.4% to 43.6%—it accepts text, image, audio, and video with a one-million-token window, and is available now via the Gemini API, on Antigravity, and the Gemini Enterprise agent platform. For your business: an introductory price with an expiration date is not a cost basis, it's a promotion. If you are going to build an automation on top of it, calculate the cost per operation using the January rates ($1.50 and $7.50), not the August ones, and ensure the process still makes sense at double the price. And build your model call behind your own abstraction layer so you can switch providers without rewriting the workflow: the discount lasts four months, your automation should last longer. Source
On Sunday, the competitor's cheap model quadruples in price—and charges differently by the hour
On the same August 13, DeepSeek introduced V4-Pro and announced a price hike effective August 16 at 16:00 UTC, reaching up to 1,100% in some cases: V4-Pro output goes from $0.87 to $3.96 per million tokens during peak hours, and V4-Flash from $0.28 to $1.32. What is truly new is the debut of time-slotted billing—peak hours are 01:00-04:00 and 06:00-10:00 UTC, with the rest of the day billed at half price—after having promised in May to keep promotional rates permanent. For your business: look at the clock before the price. That peak window of 06:00 to 10:00 UTC is 08:00-12:00 in Spain: exactly when you launch your morning processes. If you have batch jobs that aren't urgent—transcriptions, email classification, contact enrichment, summaries—moving them to the afternoon costs you half as much without touching a single line of code. And the underlying lesson goes beyond this provider: if your use case was only profitable at the offer rate, it wasn't profitable. Source
IBM becomes the OpenAI installer within large enterprise
IBM and OpenAI announced an alliance on August 13 that puts GPT-5.6, Codex, and ChatGPT Work inside IBM Consulting Advantage, the platform through which IBM provides its consulting services. IBM enters OpenAI's Elite partner tier, creates a dedicated practice with thousands of certified engineers and consultants, and will deploy embedded teams at client sites for three fronts: converting legacy operations into AI-ready workflows, modernizing applications and development, and cybersecurity; the sector focus is banking, government, telecommunications, and retail, and functionally, finance, procurement, customer service, and HR. For your business: the message is that the bottleneck is no longer the model, but integrating it into an existing process—and that is precisely what an SME can do without hiring thousands of consultants. What IBM will charge a premium to do at a bank (connecting the model to legacy systems, defining who approves what, leaving an audit trail) you can do on a small scale this week in a single procurement or service process. And if you sell professional services to large companies, expect that from now on, your proposal will be compared to IBM's. Source
Speed starts being sold separately
On August 13, OpenAI introduced Ultrafast, a new service tier for its GPT-5.6 Sol API that reaches 750 output tokens per second, up to 14 times faster than standard processing, running on Cerebras chips. It is in limited preview for select customers and targets cases where response time is king: incident response, financial analysis, customer service and real-time voice, and commerce. The figures should be read with caution: the measurements are from the manufacturer itself, taken in July, not from an independent third party. For your business: the takeaway isn't the record, but that latency is becoming a product with its own price. Before considering paying for it, measure where the time actually goes in your automation: in a voice agent or live chat, the model wait is noticeable, but in most internal workflows, the delay lies in intermediate steps, system waits, and the legacy application you are querying. Paying for model speed for a process that runs overnight is throwing money away. Source
What to watch tomorrow?
Tomorrow, Sunday at 16:00 UTC, DeepSeek's new rates take effect: if anything of yours depends on that provider, you'll see the first real bill on Monday. And Google's large model, Gemini 3.5 Pro, remains pending with no release date while the budget range is updated every three weeks.