The Price of Intelligence Just Changed

OpenAI’s steep API cuts and DeepSeek’s latest model upgrade point to a new phase of AI competition—one measured not only by capability, but by the cost and speed of getting useful work done.

A customer-support agent may call an AI model thousands of times each day: classify the request, find an order, extract key details, draft a response and decide whether a human should intervene.

At that scale, a modest benchmark gain matters less than the bill.

OpenAI has cut the API price of GPT-5.6 Luna by 80% and Terra by 20%. It also renamed Priority Processing as Fast mode, advertising up to 2.5 times faster inference for Sol—at twice the standard price.

That last detail is important. Faster does not mean cheaper.

DeepSeek, meanwhile, has put an updated V4 Flash API into public beta. The architecture is unchanged, but the company says new post-training improves agentic work: tasks in which a model plans, uses tools and completes multiple steps.

The announcement appeared on DeepSeek’s official social channel; its API changelog had not yet indexed the July 31 update when this issue was reviewed.

DeepSeek has published favourable comparisons with GLM 5.2. Those results are vendor-reported, however, and testing conditions may differ. Until independent evaluations run the models through the same tasks with the same tools, budgets and scoring rules, “outperforms” is too strong.

Our read: these announcements resemble the opening of a price contest—but not necessarily a coordinated one. There is no evidence that OpenAI made its cuts in direct response to DeepSeek. What the releases do show is that model providers are increasingly selling intelligence as a menu of trade-offs rather than one premium product.

For builders, the practical response is routing. Use a lower-cost model for extraction, classification and routine replies. Escalate ambiguous cases to a stronger model. Pay for faster processing only where waiting has a measurable cost.

A useful comparison now looks like this:

The cheapest token can still produce an expensive failure. The strongest model can be wasteful on easy work. The winner will often be the system that assigns each task to the right level of intelligence.

AI models used to arrive as fairly simple propositions: this one is smarter, this one is cheaper, this one is open.

The boundaries are blurring. Anthropic’s Claude Opus 5 offers five effort levels. OpenAI separates standard and faster processing. Sparse open-weight systems such as Kimi K3 activate only part of their enormous networks for each token.

These are different technical approaches, but they point in the same direction: users are gaining more control over how much computation a task receives.

For example, contract names and dates may need little computation; an unusual liability clause may justify much more. A leaderboard score without latency, effort setting and total task cost therefore tells only part of the story.

What to watch next: model routers that make these choices automatically—and whether they are transparent enough for customers to understand why a task became slower or more expensive.

Anthropic released Claude Opus 5 with a one-million-token context window, adaptive thinking and five effort levels. Standard API pricing is $5 per million input tokens and $25 per million output tokens.

Anthropic and independent evaluators report strong results on several agentic and knowledge-work tests, but there is no universal “best model.” Rankings shift with the task, effort level, latency allowance, scoring method and budget. High-effort runs can also take many minutes.

Moonshot AI released the full weights and technical report for Kimi K3, a multimodal mixture-of-experts model with 2.8 trillion total parameters, 104 billion active for each token and a one-million-token context window.

The distinction between total and active parameters matters. K3 selects 16 of 896 routed experts as it processes a token, aiming to gain capacity without using the entire network every time.

The architectural significance lies in how K3 allocates computation, not merely in its total parameter count. It is also not a laptop model. Storing and operating the full system still requires datacenter-class infrastructure. Its custom licence means “open-weight” should not be confused with unrestricted open source, and headline benchmark comparisons remain partly developer-reported.

In a peer-reviewed Nature study, researchers introduced ContactSeek, a method that uses AlphaFold 3 contact probabilities to identify residues affecting the specificity of gene-editing systems.

In plain language, the researchers used the model’s signals to decide where a molecular tool might be redesigned, then tested those changes experimentally across Cas9-TadA and Cas12a-based editors.

That moves structure prediction from something scientists inspect toward something that helps guide engineering choices. It could help laboratories prioritize promising mutations before running costly experiments.

The boundary is equally clear: improved precision in laboratory assays is not proof of clinical safety or successful treatment in people.

According to Axios, Google announced Gemini Robotics 2, describing systems that combine dexterous movement with higher-level planning and coordination between multiple robots.

The compelling possibility is not simply a robot completing one rehearsed action. In a potential deployment, one machine might hold an object steady while another manipulates it, or a group might reorganize a workspace without every motion being scripted in advance.

For now, the evidence comes mainly from Google’s demonstrations and reported claims. A polished demo does not establish reliable performance across unfamiliar objects, lighting conditions, hardware or busy workplaces. Independent testing and fuller technical documentation will matter.

NVIDIA’s Cosmos 3 Edge provides a useful parallel. Its four-billion-parameter world model is designed to process text, images, video, sound and action on supported NVIDIA hardware near the robot itself. That could reduce cloud delays and keep sensitive camera data local. NVIDIA reports real-time robot-control results, but those figures remain company-reported and broad deployments have not yet been established.

Taken together, the direction is clearer than the maturity: robot intelligence is moving toward coordinated planning and faster decisions made closer to the machine.

MiniMax H3 can use text, images, video and audio as references while generating clips up to 2K resolution and 15 seconds long.

That combination makes it more useful than a one-prompt video generator. A creator could provide a product photograph for appearance, a short clip for camera movement, audio for timing and a written instruction for the final scene.

A sensible first test is small: generate the same ten-second product shot three times, changing only one reference. Check whether the object remains recognisable, the motion follows the example and the audio stays aligned.

The model currently supports clips from four to 15 seconds. MiniMax has said open weights will come later, but they should not be treated as available until the files and licence are actually published. As with any generative-video system, users should also confirm that they have permission to use the faces, voices and reference media they provide.

MiniMax documents H3 through its developer platform, but this review did not independently establish consistent pricing or regional availability.

Cursor’s new India-only Start plan costs ₹649 per month including tax and accepts UPI as well as cards.

The plan includes more agent requests than the free tier, Cursor’s own models, cloud agents, mobile access and extensions. A solo developer could ask a cloud agent to draft a bug fix, monitor its progress from a phone and return later to review the proposed change.

“More” does not mean unlimited. Advanced third-party models, Bugbot, Automations and software-development-kit access remain Pro features. Still, local pricing and UPI support solve two concrete problems that global AI launches often overlook: purchasing power and payment friction.

The clearest pattern this week was not a single model taking first place, but the appearance of more controls over cost, speed and effort.

Builders can choose cheaper inference, faster inference or deeper reasoning. Researchers can inspect open weights, provided they have the infrastructure. Creators can combine several reference types, while robots may split work across machines and make more decisions locally. More choice is useful, but it shifts responsibility toward the buyer.

It is: “What level of intelligence does this task deserve—and what will it cost when the system has to do it a thousand times?”

IntellicaHub explains consequential developments in AI, robotics and emerging technology without the hype.