AI Digest: Manus, DeepSeek V4, Gemini 3.7, and OpenAI Ultrafast — August 14

Article spoiler:On August 14, four AI developments arrived before San Francisco had finished its morning coffee. Each one shifts somethi…We care about our clients, so we made a short takeaway from this article. Press to quickly get the point.

On August 14, four AI developments arrived before San Francisco had finished its morning coffee. Each one shifts something for businesses evaluating or running AI workflows: Manus loses its Meta deal and offers free access through August 25, DeepSeek ships a strong model and immediately reprices its API, Gemini 3.7 Flash enters at competitive rates that expire at year-end, and OpenAI previews a speed mode that will either solve a specific bottleneck or simply accelerate budget consumption, depending on what you are building.

The AI landscape in August 2026 has not settled into a predictable cadence. Four developments arrived on a single day, each carrying a different kind of practical implication for businesses that use or are evaluating these tools. Here is what happened and what we think it means.

Manus goes free after the Meta deal falls through

The planned acquisition by Meta was blocked. Manus returns to independent Chinese ownership, and the company is offering Manus 1.6 and Manus 1.6 Lite at no cost through August 25 as part of the transition. Existing users with active projects have a data migration deadline of August 23.

What GLC thinks. The free access window is a retention mechanism while Manus reorients following a deal that did not close. For teams that were mid-evaluation of Manus as an agent platform, the capability of version 1.6 and 1.6 Lite remains what it was before the acquisition fell through, so the evaluation itself is still worth completing. The ownership shift reintroduces a question that applies to any AI tool operated by a Chinese company and used in European business contexts: where data resides, under which governance framework, and what the implications are for GDPR compliance. That question belongs in any enterprise evaluation of Manus regardless of the ownership structure, and the current moment makes it timely to put it explicitly on the table.

What to do. If your team has active projects in Manus, August 23 is the data migration deadline. If you were evaluating Manus for agent workflows, the free window through August 25 gives time to complete a structured assessment with a data governance checklist alongside the capability review.

DeepSeek V4 Pro-0813: strong model, significant price change

DeepSeek released V4 Pro-0813, and early benchmarks show performance that tests competitively with much higher-priced models in three specific areas: coding, review of large technical projects, and long multi-step reasoning chains. The same release came paired with a pricing restructure effective August 16: DeepSeek's API will move to peak and off-peak pricing windows, and certain specific operations will increase in cost by up to twelve times under the new rate card.

What GLC thinks. The model's capability profile looks genuinely useful for technical workloads, and for SMBs running development projects or working with large technical documents, V4 Pro-0813 deserves a proper evaluation. The pricing announcement is the more immediately operational story: any workflow currently calling DeepSeek's API should be cost-modelled against the new rate structure before today's deadline. The logic of running everyday and lighter tasks on V4 Flash while reserving V4 Pro for genuinely complex reasoning mirrors the same routing principle we covered in the GPT-5.6 Sol/Terra/Luna release — a pattern that is appearing across multiple providers.

What to do. Audit your current DeepSeek API usage against the new rate card before the August 16 cutover. If your workflows include batch processing jobs, the peak/off-peak structure may make it worth shifting those to off-peak windows. For new technical workloads, V4 Pro-0813 is worth a direct benchmark against your own representative tasks rather than relying on published scores.

Gemini 3.7 Flash: Google's efficient tier, on time as usual

Gemini 3.7 Flash arrived, and Google's benchmarks place it close to Claude Sonnet 5 and GPT-5.6 Terra in coding tasks and agent workflows. Google has maintained its consistent pattern of shipping the efficient-tier model before its flagship, and Gemini 3.7 Flash follows that cadence. Pricing is set at $0.75 per million input tokens and $3.75 per million output tokens through December 31, after which the standard rate is scheduled to double to $1.50/$7.50.

What GLC thinks. At the current price point, Gemini 3.7 Flash is worth including in any comparative evaluation where Terra or Sonnet 5 serve as the existing baseline for agent or coding tasks. Google's self-reported benchmark figures warrant the standard level of scrutiny: they are a useful starting signal and a reasonable basis for deciding which tasks to test, but your own representative workloads are the relevant performance test for any production decision. The year-end pricing schedule is worth tracking: the Q1 2027 economics at $1.50/$7.50 look different from the current introductory rate, and any workflow adoption decision should be modelled at both price points.

What to do. If you run agent or coding workflows at any meaningful volume, the current Gemini 3.7 Flash rate makes a short evaluation worthwhile before December 31. Build the Q1 2027 pricing into the business case now rather than revisiting it at renewal.

OpenAI Ultrafast: 750 tokens per second, price unknown

OpenAI previewed Ultrafast mode for GPT-5.6 Sol, targeting up to 750 tokens per second, approximately fourteen times faster than the standard operating speed. Access is currently limited to selected API clients, and OpenAI has not announced pricing.

What GLC thinks. 750 tokens per second is a speed that opens up specific use cases that the standard rate makes impractical: live voice interfaces, real-time customer-facing agents where response latency is part of the user experience, and interactive workflows where the user is waiting for each output before taking the next action. For the large majority of SMB analytical, content, and operational workflows, the speed differential between Ultrafast and standard Sol does not produce a meaningfully different outcome. A report, a strategy document, or a batch of classified support tickets does not become more valuable because it arrived fourteen times faster. The relevant variable will be the per-token cost, which OpenAI has not published. When pricing is announced, the question is simple: does a specific bottleneck in a live user-interaction workflow justify the premium. For most SMB workflows today, the answer will be no, but the exceptions are worth identifying now.

What to do. Identify the two or three workflows in your current or planned AI setup where response latency is actually the binding constraint on the user experience. Those are the candidates for Ultrafast when pricing and general access arrive. Everything else runs on standard Sol, or on Terra and Luna where the economics are better.

---

Every week, staying current with the AI market well enough to make useful implementation decisions requires time that most SMBs have reasonably allocated to running their actual business. Part of what GLC covers in its consulting work is precisely that: following, evaluating, and translating what is happening in the market into specific decisions for client workflows and budgets. If you want a read on how any of this week's releases applies to what you are building or planning, Get in touch.